🧣｜comfy-ui

[halcyon dawn]

[lone cape]

hi 🙂

[thorny lily]

new channel!! neat!

[sly sundial]

Allow me to kick things off with a selection of workflows:
https://civitai.com/models/119257/gtm-comfyui-workflows-including-sdxl-and-sd15

[midnight solstice]

huh

where's comfy

[fickle remnant]

Whats that new color

[lusty pewter]

Nice channel add.

[karmic creek]

Finally! this was a very needed channel 👏

[desert sedge]

Nice 🙂

[karmic creek]

sneak-peek of a useful custom node we've been developing for a project

[silver nebula]

Hurrah! I have been asking about a problem with InstantID - namely "assert 'detection' in self.models"

[karmic creek]

Does anybody else have this warning? [AnimateDiff] - WARNING - xformers is enabled but it has a bug that can cause issue while using with AnimateDiff.

seems to be working fine but i'm cautious

[karmic creek]

didn't know about this in comfy, just have tried faceid and plus face in ip adapters and they work fine, some more than others but you can get consistent good work

[silver nebula]

Very few people seem to know what to do with this error? "assert 'detection' in self.models"

[karmic creek]

dont know if this helps but it's a similar error https://github.com/deepinsight/insightface/issues/1525

[silver nebula]

Thank you. Will give it a go.

... on my way to Venice, Italy!!! Arrividerce!

[karmic creek]

nice!

[silver nebula]

Si, mi molto piace - pizza, pasta, calzone, Valpolicella ... :-d

[river frost]

although it's a 3rdparty, krita plugin with certain inpaint settings - freezes on this step, slowly increasing RAM usage:

[lone cape]

SDXL turbo mode problem solved. In Civitai the "turbo" safetensors are fake. No useful result if set to step 1. But the official turbo safetensor really works with low step. 1 step at the image, "turbo" models downloaded from civitai never get similar results. But really no need special nodes and workflow, just set the cfg around 1..1.2, and the step between 1 to 3.

[fickle remnant]

Yep got the same

xformers is outdated anyways now, Flash Attention V2 is my friend now

Not sure if I can get it running in comfy tho

[stoic rock]

Looks very interesting. I understood your graph up until the presampling node.

[rose storm]

although they are not fully turbo models, they are usually merged with turbo and allow generation with few steps.

I'm using DreamshaperTurbo most of the time. It needs 6 steps (instead of 1 as in turbo), but it allows arbitrary resolutions 1024 and above (not just 512 as in turbo) and it gives me much better quality images as turbo or sdxl base

so if you need one step generation, base turbo is the way to go. If you want speed and good quality, the merged turbo models might be a good compromise

[strong flower]

some SDXL Turbo models allow UltimateSDUpscale with only 1 step per tile too, with LCM

[slim root]

My problem with Turbo was I could only get good quality portrait shots and background was always blurred out of focus. That may have changed with these merged models but so far it was only useful in those situations.

[lone cape]

what else I tested later, and can't see in the first picture I attached, when steps is increased, the turbo workflow picture stays good, but the non-turbo one goes bad. For this effect also need the official turbo model. Basically, my problem with the models I downloaded from civitai before, that they gave exactly the same result as the non-turbo workflow, even though the two do not have the same back-end. Everything seems ok with the turbo model downloaded from Huggingface:

• gives a usable result at 1 step, with both turbo and non-turbo workflows. 1 sec runtime
• if increase the number of steps, the turbo and non-turbo workflow will show more and more differences, the turbo workflow remains only usable. 2 steps are enough, 3 steps are perfectly fine but only using turbo workflow
• At 1024 resolution, the real turbo model gives completely wrong results

[rose storm]

as said,you don't need a turbo workflow. Just set cfg and steps to low value and you can use the normal workflow

and sdxl and sdxl turbo have the same backend. In principal, sdxl turbo is just a fine-tune of sdxl base

[rose storm]

I don't wanna make advertisement for DreamshaperTurbo 😅 but it does not have this problem. It just gives me better results than any other model I tried so far with a lower number of steps. the only downside of it is that it almost ignores negative prompts (due to low cfg). So whenever I need negative prompts I have to go back to a non-turbo model.

[rose widget]

Is it the sampler at fault? I have no idea how "DPM++ 2M SDE Heun" is called in comfy.

[slim root]

You may want to use StableSwarmUI if you are just starting out. It uses comfy but has a sleek UI front end that's easier to navigate. Then when you get more comfortable you can switch to the noodle nodes.

[rose widget]

Thanks, but I don't have any problems with nodes. I am used to them from working with Blender. Just wondering if my bad results are due to my setup, or settings, or something else entirely.

[fickle remnant]

Sorry this image is against our NSFW rules and ill delete this message

[rose storm]

I'm not entirely sure but I think the "Heun" is redundant as both terms "2m" and "Heun" refer to the same thing, a second derivative optimizer. So "dpmpp 2m sde" should be the same as "dpm++ 2m sde heun"

[rose widget]

I see, thank you. For some reason it's called "2m Heun" in SD UI made by AUTOMATIC1111.

[fluid talon]

new channel for comfy, interesting

[eternal idol]

Hi, do any of you know is there preview image node with golden ratio grid? 🙂

[lone cape]

That can't be true. The result of turbo and normal workflow is similar with exactly same setting is only if the step = 1. If more then 1 step, start getting differences. Not just the difference, but the non-turbo wf going wrong while the turbo wf stay good (to step 3). Pic 1 using step 3, the framed cyan wf only the non turbo. The black is the comfy recommended wf, visible the difference, I cant wire the output of comfy turbo nodes to non-turbo sampler. With step 6 (pic 2) is too much for turbo wf, but the framed normal wf much more worst than another two. The full running time with official safetensor is 17 sec, the sampler running 2 sec if step 6, 1 sec is step 3 or lower, the used model on this example is 12.9 GB. Wait a bit for the similar test result if I load 'turbo' safetensor downloaded from Civitai. I will save screenshot with merged turbo checkpoints.

[hollow jewel]

hi

[slim root]

@hollow jewel are you a real person? If so I only add people I know irl as a friend. If your some bot then go away.

[hollow jewel]

yes I am

[slim root]

That's exactly what a bot would say.

[lone cape]

New test with "turbo" safetensor 'realvisxlV30Turbo' from civitai. 6.4GB. The very first loading time was about exactly 1158 seconds (while the first step was 254 sec). Then the running time is 23 sec what is not bad. With step 1, all workflow getting unuseful shit, compare to the 1 step settings in prev. comment using the 'official' turbo model. The pic 2 using step 10. The result maybe useful, but exactly similar in all workflows (getting similar result with all step from 1 to 10), while from step 2 the result getting different if used Stability's safetensor. This is why I dont think that the merged models same as than the official turbo. Not the workflow can be same if using fake turbo checkpoints, if the workflow different, you will get exactly same result (the flower in the desk have small difference in normal mode). For real turbo models different workflow needed.

[strong flower]

it's a merge, on the page it's indicated that it needs 4-10 steps

and SDXL models need 1024x1024 images, 512x512 was for SD1.5

[lone cape]

https://huggingface.co/stabilityai/sdxl-turbo

"Limitations
The generated images are of a fixed resolution (512x512 pix), and the model does not achieve "

[rose storm]

I don't know how your turbo workflow exactly looks like, as you use a lot of custom nodes, but the normal comfyui turbo workflow is running the same code as the non-turbo workflow

also check the readmes. Sone turbo models do only work with specific samplers

DreamshaperTurbo turbo for example is only working with dpm++ sde karras

[strong flower]

sdxl-turbo =/= realvisxl turbo :/ You can see that the example images on Civitai model page are 1 million pixels (1366x768 etc.). I'd check with the parameters they say

[lone cape]

Here is the original comfy turbo workflow, not same as the default non-turbo. I used exactly same wf in my test, the back nodes is the comfy recommended turbo wf. The dark blue nodes are from the "easy nodepack", but that just a test, the result of dark-blue is exactly similar than the comfy turbo. https://comfyanonymous.github.io/ComfyUI_examples/sdturbo/

[lone cape]

A merged checkpoint can overwrite the details of Stability? If can its ok, but I think then this checkpoint is not turbo.

[thorny lily]

SDXL Turbo 1.0 is a model which generates at 512
"Turbo" is a concept that can be applied to any model and any resolution

the model is effectively just a reference example of the turbo training technique ("Adversarial Diffusion Distillation")

LCM is a similar (but distinct) technique. While some do match the training, a lot of community "Turbo" models are made by just merging the Turbo reference and/or LCM reference in. The word tends to be used more at this stage to just indicate the idea that a model is designed to generate with very few steps in the way Turbo (ADD) and LCM are.

(Any model that requires more than 4 steps to generate coherently is not Turbo in any technical application, although models that use in the range of 8-12 are still are least matching the level of speedup that LCM did)

[strong flower]

Thanks a lot Alex for all these infos 🙂

[sly sundial]

I've only tried a normal sdxl model with an LCM lora, but I get poor results from that. Is that not expected to be as good as merging a model with the turbo model?

[merry ermine]

Can't wait for mesh graphormer to be able to detect deformed hands T_T

[steep hawk]

what's your favorite checkpoints to use with AnimDiff?

i found that my stuff are always a bit to woobly, i'd like to have sharp edges for more technicals stuffs

[rose storm]

I mean, that works already, isn't it?

[merry ermine]

it works on mostly featured hands, but not when the hands come out real bad, like when subject is far from 'camera view' for full body portraits.

I guess instead of a hand detector I need a blob at end of arm detector.

[rose storm]

I mean, you probably would have to combine that with an detailer

[merry ermine]

yeah I've been using mesh graphormer with detailer

[royal pond]

Awesome new channel! Now people can finally stop posting workflow debugs in SDXL.

[obtuse latch]

question for images, I want all of my generations to output 1920x1080, could someone show me the nodes or workflow to achieve that? having some issues with my upscaler.

[sly sundial]

Why don't you set it to 1920x1080? It's set to 1344x768 in the screenshot.

[strong flower]

You can generate in 1344x768, upscale 2x (which is in fact 4x) to 2688x1576, then use an image editor to resize it to 1920x1080 (or a Comfy resize node, there probably is one)

[winter forum]

ComfyUI devs/users might want to search for Deforum in their Manager.

[sly sundial]

Why?

[winter forum]

To find out why. We are working on Deforum Nodes and they are available for testing.

[fickle remnant]

And now?

[sly sundial]

Does anyone know if there is a workflow for SVD v1.1?

[slim root]

https://civitai.com/articles/3355/hxsvd-harrlogosxsvd-txt2video-comfyui-workflow-generate-and-animate-text-with-svd-v2-out-now

Same as SVD

[sly sundial]

I don't think this has nodes that support v1.1 though.

[slim root]

I just swapped out the SVD model for 1.1. are there different settings I don't know about?

[sly sundial]

I thought it needed some specific nodes that support it. I was using these, but they fail when I try using v1.1 model...
https://github.com/thecooltechguy/ComfyUI-Stable-Video-Diffusion

[slim root]

As far as I'm aware as long as you are in the latest comfy support is built in and you just have to add the model like in that workflow I sent.

I'm generating videos using my same settings from the original model and I'm not having any issues.

[sly sundial]

Doesn't work for me, but it could be because I'm running torch 2.2

The errors are pointing that way anyway

RuntimeError: CUDA error: invalid configuration argument
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.```

Same error with svd_xt model too 😦

[sly sundial]

Fixed it! Comfy was using xformers, so needed to run it with --use-pytorch-cross-attention

[sly sundial]

An interesting concept for a node...send the prompt to Bing/Dalle, retrieve the output image and use it to assist SD with the prompt understanding/composition. This would allow much more prompt control, as it's basically using the capabilities of Dalle3.

...unless that's what the next model will be capable of already!

https://tenor.com/view/spongebob-squarepants-begging-pretty-please-beg-on-your-knees-pray-for-mercy-gif-26344462

[slim root]

That would be super expensive for something you could achieve with a specific workflow.

[river frost]

It looks like some things remain in VRAM until restart (leak?)? Even though it's supposed to be cleared in normal vram mode...

[sly sundial]

What would be expensive about it?

[slim root]

The API calls to DALL-E

[sly sundial]

Ah, yes. I got used to using it for free and forgot you have to pay for API

[surreal whale]

Look what came to Comfy https://github.com/XmYx/deforum-comfy-nodes

[winter forum]

example workflow in the repo/examples folder, you might want to bypass the hybrid node as you wont have a video loaded in the video loader node, and you'll also notice, that it has a frame by frame logic, so you'd want to enable auto-queue in Extra Settings. Then you can iterate settings while running. The latent in the iterate node is only reset when the incoming latent is None or reset switch is enabled.

[surreal whale]

what type of input is the video loader looking for? ive never used deforum prior, i however did sucessfully use these nodes

[winter forum]

when selecting a file, enter a * in the filename area and press enter, then you can actually select a video file, on each iteration it will emit the next frame.
This is due to Comfy's image uploader being reused.

could be used for controlnet, ipadapter, init, or as plugged in now, hybrid motion input

the video save node will save it's cache when max_frames is reached, Deforum Animation Params

[surreal whale]

Thanks

[olive yoke]

This is already something people do. Usually with controlnets. However the control net checkpoints themselves are not very well trained

[obtuse latch]

Could someone help me with the switch clip and switch, I don't fully know what those are.

[winter forum]

LoRA's are applied to both the Unet (model) and the Clip model. Clip weight is the lora's strength in the clip model. the switches toggle whether a lora is turned on in this node or not. there is not clip switch.

[obtuse latch]

Thanks, just so I understand the CLIP is the text prompt after it's been processed?

And how is that different than the model_weight_x there?

[winter forum]

no, clip is the clip text encoder model that encodes your prompt and negative prompt into conditionings. the clip_weight is the selected LoRA's weight on the clip model, and the model weight is the lora's weight (strength) in the Unet (sampling model)

[obtuse latch]

Thank you, one final question Unet (sampling model) what exactly would that be? What it decides to create from the noise on the model?

Just trying to understand what exactly is happening at these levels.

[winter forum]

that's the main model practically, that was trained against clip embeddings and vae encoded images, and returns the encoded image (tensor) output to be decoded by the vae when sampling

it's always denoising a noisy / noise latent tensor

tensor is just an array of numbers easily held in the gpu, decided to be the intermediary format of machine learning.

[obtuse latch]

got it, I understand that, so quick follow up, when I set a lora weight on the model, is it just inserting itself into the noise?

[winter forum]

no, a LoRA is practically a really small part of the model that gets injected into the model

merged on runtime

its not as deep as a finetune, it only affects X layers

[obtuse latch]

Ah so that's why there's usually steps to start at.

Thanks mate, appreicate your time.

[winter forum]

well, the timesteps are also tensors in order to be processable in our context, so yes

ml ops on high resolution stuff is really expensive (disco diffusion), that's why a VAE encoder/decoder was implemented, that almost lossless compresses the image into 1/8th of its size, but adding a channel, so it's becoming a latent image.

[rose storm]

no, that's another reason

you run the unet multiple times (for each timestep) independently from each other. You can enable and disable loras at every timestep as you want (you could even change models at each timestep, but switching model usually takes more time as they are too large to fit in your vram). Enabling a lora only at timestep X allows you to activate the changes of the lora at the fine details of the image, while keeping composition unchanged

[obtuse latch]

Anyone have any suggestions on why my upscales are losing detail (looks alomst like it's clay)?
Left upscale, right before upscale:

It's got that mud look: 😦

[hardy pebble]

why should one install comfyui?

[surreal whale]

some more versatility i guess, tends to be harder for some (computer stuff)

[hardy pebble]

should i install it?

[surreal whale]

you used another app before?

i would, its good to learn, lots of development around it

[hardy pebble]

I use normal version

umm the webui?

is it what we call it

[surreal whale]

A1111

[hardy pebble]

Will it increase the quality of work? or better images will be created?

[surreal whale]

not particularly

[hardy pebble]

I do not collaborate or share projects. I read it like it makes it easy to share workflows. But i dont really share any.

I am confused at this point.

also I am not really into much how developing it works or whats its inside process

So do i still need it?

[surreal whale]

no

it does the same thing as what you used just diffrently

[hardy pebble]

hmm

[surreal whale]

a little more complex sometimes

[hardy pebble]

i think it maybe even make it hard for me

complex.

[surreal whale]

right

[hardy pebble]

ye you are right 🙂

thanks

^-^

[surreal whale]

np

[obtuse latch]

I personally came from auto 1111 and prefer Compy, you can easily copy paste images (that are embedded) or raw json for workflows.

[hardy pebble]

hm

[obtuse latch]

It's a little more work to learn it but it gives you more control over what you can do, in terms of customizing.

If you're just doing basic prompting, or img2img and have the plugins you want with 1111 then there's no real difference.

I'd only switch to comfy if you want more finetune control.

[hardy pebble]

I only use like basic stuff text to text and ima2img

and dont really need much finetune or specific details like perfecvt blurr

[obtuse latch]

Finetune would be more around giving you access to nodes without having to write them in python.

[hardy pebble]

oh

[obtuse latch]

A component / plugin in 1111 would be at the mercy of whoever programmed it.

If there's a property they didn't expose you either would have to write it yourself or deal without it.

[hardy pebble]

damn

i am not smart enough to write it

[obtuse latch]

But there's little chance you would need it.

[hardy pebble]

imma leave it, it does same as a111

i am not too picky about stuff

just a normal noob guy

generating random images .

[obtuse latch]

[question] any ideas why I'm losing so much detail on upscale left (and not gaining more?)

[obtuse latch]

You should be fine.

[hardy pebble]

ty

[rose storm]

I would say your denoise is way too low. As higher the value as sharper the image (up to 0.5 denoise). Use tile control net to prevent changes in the image

note that every pass to the vae damages the image (it's a lossy compression), so you already need some minimum denoise to revert the damage by the vae

also, SDXL is not good in upscaling (in particular with SDUpscale) and often ends up in washed out and blur images

so rather use SD 1.5 (your favourite custom model) + tile control net + SDUpscale together

or use SDXL img2img on the enllarged image directly

[obtuse latch]

Thanks, would you say steps matter in an upscale?

rasied it to 0.5 on a different upscale workflow, ideas on fixing the tiles?

[rose storm]

increase tile padding as much as possible (e.g. to 256 or even 512)

steps can be less, as SDUpscale is not multiplying steps with denoise

so if you do 50% denoise you can just do 15 steps

but too many steps don't do any harm usually

[obtuse latch]

getting closer, thank you.

[rose storm]

you can try seamless tiling, but I don't think that will solve the contrast issues

if contrast issues remain and the image is not too large you can make another pass with sdxl (without upscaler)

[obtuse latch]

seamless tiling?

contrast is fine, just still getting tiles:

[rose storm]

I mean seam fix mode

(with say 25 - 33% denoise)

this is just doing inpainting om the edges

but I have the feeling that the tiles have different brightness, inpainting will just make it less visible but it will still be there

[obtuse latch]

it takes 20 min to generate this :D, so sorry for bugging you, this? or actually 25?

[rose storm]

yeah, I would also increase padding, but just try around

[rose storm]

uff, it's a very high resolution? 😅

[obtuse latch]

Image size: 1536x880
Prompt executed in 1196.41 seconds

the tile padding going to 1024, killed the time, used to be 2 min gens

[rose storm]

hm, okay, depends on your vram. For me the resolution still fits into my vram for sdxl 😁

[obtuse latch]

8gb

[rose storm]

use your last image, set step to 2 and denoise to 0.1

yeah, 1024 is a lot

[obtuse latch]

denoise to .1 doesn't give me as good detail

I'm almost there:

[rose storm]

what I meant is: you already did upscaling with details, you just want to fix the edges. Therefore, you can reduce denoise to 0.1 to speed things up

[strong flower]

a tile of 1024 is something for the 8GB VRAM GPUs I'd say (as it's what I have lol). If you only have 4GB, the other 4GB it needs will be transfered to the RAM, which is way slower and would explain why your generating time is raising up like mad

[obtuse latch]

Ah, I got you, yeah good idea.

[sly sundial]

I found that the 3m sampler isn't good at anything not human, and it's also better with exponential or karras scheduler. Try using 2m karras, lower CFG to between 2 and 5 and increase the denoise to 0.25 to 0.35

I should've kept reading before replying, I see you've already had help 🤔

[sly sundial]

Better?

[deft jetty]

how do I fix this?

imported a workflow

[lusty pewter]

Download and install ComfyUI Manager, then use the included feature to install missing nodes.

[slim root]

Install the missing nodes using the manager. Google how to install ComfyUI manager.

[deft jetty]

Thanks mates

[sly sundial]

https://github.com/ltdrdata/ComfyUI-Manager

[deft jetty]

got it, anyway thanks

hmm how to upscale? I mean I can upscale but its always upscaling 4 time

so I did this but its just resize or smthing?

I need to upscale 512*768 2times

[obtuse latch]

I'm out of ideas why img2img is making literal isometric tiles...

[obtuse latch]

Thank you, this actually did give me some help

[solemn pebble]

Could someone help me figure out why I'm getting so much distortion in this workflow? It seems to be coming from the second ksampler prior to the final upscale but I can't figure out why.

I'm playing around with switching to using DPM++ 2M Karras from Euler or Euler A but having a lot of trouble getting similar quality results.

[olive yoke]

you are using 960x544 which isn't a resolution that sdxl is trained for

you are using lora weights that are higher than 1, which is a bad idea

and you are using an add detail lora, which is not a real thing

you are also trying to combine loras, which whether or not you can do that is very specific to the loras and not generalizable

[olive yoke]

you should upscale to the parameters it is trained for

then, if you need a different, smaller resolution, adjust it in photoshop

[solemn pebble]

If I'm trying to end up with 4k wallpaper size, how should I approach that? I had been using 960x544 which at 4x is 3840 x 2160 (or near enough).

[olive yoke]

4x upscalers do not go from 1:4, they go from 1:2

you would start with a supported resolution like 1344x768, then ML upscale to 2688x1536 and call it a day.

[solemn pebble]

I have another 2x upscale in the flow before the second ksampler

[olive yoke]

this stuff can only do what it's designed for. there's no free lunch

there aren't any upscalers for 1024x1024 to 2048x2048. as far as i know, they are all trained on 512x512 to 1024x1024. so unless you are doing tiled upscaling, the results can be very iffy.

[solemn pebble]

It was "working" with euler but I (most likely) don't understand if the flow was really working like I thought.

[olive yoke]

i don't know what to say besides it can be kind of random when you are diong a bunch of flawed stuff

[raw maple]

best upscaling quality and sharp workflow someone know? have? or see? I need some good upscale workflow please 😉

[rose storm]

I assume your background is "too strong". It forces the generated image to have this kind of gray-ish background, too. So this forces the scene into a isometric one

[rose storm]

as doctorpangloss said, you should use a resolution of at least 1megapixel for SDXL, otherwise, image quality will be bad. You doesn't have to use exactly the resolutions sdxl is trained on, but it might help a bit. For upscaling, you do not need an upscaler method. Honestly, I never found an upscaler that worked good, they all were so bad I need another SD upscaling pass anyways. So just upscale without model to the size you want and do img2img on it

[queen temple]

honestly, i'm getting good enough results, by genning 512x512 highres fix x2

is still very slow, but the results are very impressive

[silver nebula]

Add Detail LoRA can help increase contrast, as SDXL can be a little easy on contrast.

[lone cape]

Hi, finally I developed a detailer node which support normal, LCM and Turbo modes if the model_concept and concept sampling inputs are correct, and follow the concept requirements. I using for face, eye, hand and mouth restoration. I will publish this new node soon. Somebody wanna do preview test?

[queen temple]

why not? let's see what it does

[surreal whale]

ive been messing around with upscaling like this as things like SD ultimate upscale tap out my memory, im just an overwhelmed dabbler but does it look reasonable or trash it and scour the web for more ideas

[obtuse latch]

I found serge work flow on github is pretty much the best one.

Multiple loras, upscalers, and two passes.

[surreal whale]

found Searge probably what your talking about, ill check it out

[raw maple]

Can YOU SHARE THIS WORKFLOW, SCREEN OR FILE TO IMPORT MY COMFY?

[sly sundial]

https://github.com/SeargeDP/SeargeSDXL

[raw maple]

wtf thats big layout workflow

[sly sundial]

It's monstrously huge 😄

[raw maple]

how much vram need for this workflow?

[sly sundial]

🤷🏻‍♂️ I don't use it

If you want something a little simpler https://civitai.com/models/119257/gtm-comfyui-workflows-including-sdxl-and-sd15

[raw maple]

but @obtuse latch use it

[sly sundial]

What's the relevance?

[raw maple]

he know how much needed vram to work on this layout 😄

[sly sundial]

The workflow doesn't determine the amount of vram, but what you do with it does.

All depends on what settings you change.

[raw maple]

yeh ok

[sly sundial]

I'm pretty sure the ones I posted will use less, as they're not doing so much complex stuff and using so many models.

Easier to understand and use too.

[raw maple]

yeah and this is fo sdxl and sd? because u can see what version?

[sly sundial]

It'll work with whatever model you use, but you'll need to tweak settings to allow for that

"cleanUI_vx.x" are the latest iterations

[raw maple]

I understand, but I'm a beginner and I'd like to see the differences between them

😛

[sly sundial]

If you're a beginner then you may implode trying to understand Searge's workflow 🤣

[raw maple]

no thanks just tell me about his 2 or just download this clean version

[sly sundial]

Just download the latest (cleanUI) one, because each one should improve on the previous.

[raw maple]

ok

[strong flower]

SDXL on stock settings should allow base image and tiled upscales with no problem on a 8GB VRAM card, with less you'll probably have to enable some optimizations

[raw maple]

have 4070 12gb

[strong flower]

I have a 3070 8GB so you're good 🙂

[raw maple]

ok thanks

[raw maple]

Can someome tell me why i get this mark? in N prompt is no logo text etc,

[haughty edge]

This is usually fixed with “watermark” in the negative prompt. Or Inpaint it out in the source image

[raw maple]

i take this sample rom github tutorial

[raw maple]

not helps

[haughty edge]

The style looks totally different, so the model or the seed are different from the sample. Sometimes it’s not straightforward:)

[raw maple]

yeh i know is different because i change it to do something else, but that dosent mean i have before and after watermark 😄

[wintry fox]

u probably used a motion module that was trained with modelscope,try with another motion module

[raw maple]

ok

ok so model what i downlaoded via comfy manager mm_sd_v15 have watermark

[wintry fox]

try with v3_sd15_mm.ckpt

[raw maple]

its cool, nice.. thanks i have few models also mm_sd_v15_v2

[jagged trout]

can anyone help me with this node, seems its not available, I tried installing it with missing nodes

[sly sundial]

Just replace it with this one

[jagged trout]

on it

mine is not like yours, I mean there's no option to connect nodes

[jagged trout]

help???

[sly sundial]

Right-click and select "convert xxxx to input"

In this case, text positive and negative

You probably only want to have "style_positive" set to yes, and the others to no

[jagged trout]

worked, Thank you soo much

[strong flower]

it reminds me that it would be cool to be able to move the node connectors because sometimes (often) we can't place nodes in a way where it doesn't intersect each other, but could if we could move the connectors up/down in the node.

we could also move it on the up/down side of the node

[sly sundial]

Just use these and never need to have them connected at all... https://github.com/chrisgoringe/cg-use-everywhere

[strong flower]

(I think that the video editor "Davinci Resolve" does that dynamically)

[olive yoke]

because the model is trained on video scraped from shutterstock previews. you would have to crop it out.

[olive yoke]

the benefits of the community workflows are extremely marginal. try to use the simple comfyui examples workflows first, and see if get results you like.

unless you find worthless IT faff really intellectually engaging

[strong flower]

Agree, look at some excellent tutorials from the also excellent Scott https://www.youtube.com/@sedetweiler/videos he starts workflows from scratch and describes what each added node does, because starting from hyper complicated workflows won't allow you to know what to change and where when you won't get the reasult you want.

[olive yoke]

yes this is excellent

[slim root]

Yeah his videos are great I actually subbed to his YouTube.

[obtuse latch]

Why are you yelling at me?

[raw maple]

sorry capslock 😄

[swift dust]

Hi All! I am trying to run comfyui in an HPC Cluster. Though my problem is, I cannot visualize the webui on the given link http://127.0.0.1:8188/. Is there any solution for this?

[naive brook]

start with --listen [IP] Specify the IP address to listen on (default: 127.0.0.1). If --listen is provided without an argument, it defaults to 0.0.0.0. (listens on all)

you can get all options with python main.py --help

[obtuse latch]

You should be able to just run headless, you can load workflows and add / remove / trigger prompts easily.

[swift dust]

This I've done and still shows unable to connect when I run the link in my laptop's browser. I am accessing the HPC remotely from it's IP address. I'm thinking if I need to do SSH tunneling?

[swift dust]

Forgive me but what does this mean?

[naive brook]

This is what I do at work. A firewall at play?

[swift dust]

there is

[olive yoke]

it sounds like you are an advanced user. is your hpc cluster orchestrated? with kubernetes? slurm? or VMs?

or is this a single machine with interactive login?

i run comfyui clustered

[swift dust]

We are using SLURM and on bare metal

[olive yoke]

what is your objective? generating synthetic image data?

big picture objective

[swift dust]

Yes, you can say that.

[olive yoke]

okay. and you want to do some kind of fine tuning first. can i ask what it is on?

you are in great luck

some stuff you have --> you imagine some kind of fine tuning --> then you use controlnets or some other bs --> then you get a bunch of stuff that you want that is a mix of your trained content and new content

right?

i really don't think this is sensitive

it's totally and utterly obvious so i wouldn't get hung up on the cleverness of it all

[swift dust]

Well, yeah. But before I come to that, I need to access the webgui first, which is what I am needing help for. I just can't directly run http://127.0.0.1:8188 in the HPC

[olive yoke]

i think when you have a budget and you are looking for free help, it is totally within bounds to give and take. all i am asking for is an honest answer for what the application is

[naive brook]

open ports, change listening ports if that helps or as you said do some kind of tunneling if that's needed.

[olive yoke]

it's your prerogative. i really don't think it's sensitive

[swift dust]

I mean, I just started playing with SD for the past 3 weeks. lol . Application - to create some cool stuff?

[olive yoke]

think very deeply about what i said

what is the application?

@swift dust my goal, besides helping you, is to direct your resources to something valuable that i hope, it will turn out, you will share with the community

[swift dust]

Don't you worry, it's all SFW and yeah I'll share to civit whatever I can come up with

[olive yoke]

i really don't think the application is sensitive

i don't care about SFW versus NSFW

there isn't something clever about what is going to be made

do you know what i mean?

it's really not sensitive

you can also direct message me

like even sharing a specific idea will help a lot

[swift dust]

I'll use it to fine-tune a model based on arabic and indian cuisines, which the current available models doesn't do so great for these cuisines.

[olive yoke]

yes but why

what is the idea

[naive brook]

to learn

[swift dust]

lol nevermind. it's 3am here anyway

[olive yoke]

okay well... listen, you have a DGX machine, you know what SLURM is

you're not a hobbyist

i mean it's possible that you are a hobbyist, i understand there are circumstances where that may be true

but i think this stuff is give and take

[olive yoke]

so it was a small ask.

[olive yoke]

i can only help you with stuff that makes sense

you can't make up an application. it will not make sense, and i will know that

i don't know if you did make up an application, i'm just asking you to be honest and complete

but you've definitely piqued my curiosity

[swift dust]

it'll be SaaS

[olive yoke]

lol

i have written something that will conquer the hardest part of what you are trying to do, and you're still not just saying what it is!

just say what it is

trust me

it is NOT sensitive.

but if it turns out to jsut be "clipdrop"... you know

that's why i'm saying you should just write what it is

[obtuse latch]

Firewalls don't really do anything on local networks when running websockets like this.

[obtuse latch]

You don't need a gui to run comfy

It sounds like you should just write an API, I don;'t know why you're trying to do it the way you're doing it.

[olive yoke]

on a vanilla ubuntu install, you will use ufw to disable the firewall for port 8188 and comfyui --listen in my fork (or python main.py --listen in vanilla) to get everything going. on RHEL, consult ChatGPT for the alternative to ufw. nftables commands aren't permanent. in openshift it is more complicated. your networking is probably managed elsewhere. @swift dust

you had better get familiar with file systems. unless your vendor gives you a clustered filesystem (like ONTAP), you will eventually cross a bridge to run kubernetes.

[naive brook]

it does not have to be a firewall local to the machine. it's just something that's between his client and the datacenter. could be at several places in the network path too. he have to figure it out himself. issue is not comfy.

[obtuse latch]

If it's a SaaS it should be an API.

[naive brook]

and the api is exposed how..

[obtuse latch]

On a webserver?

Nginx...

[naive brook]

Yeah, a tcp port in this case.

[obtuse latch]

You have frameworks handle that, it;'s 2024 😄

[young wadi]

hey guys i succesfully installed ComfyUI on my pc but when i try to install some custom nodes it fails ,precisely (((ComfyUI's ControlNet Auxiliary Preprocessors ))do i do anything wrong?

[surreal whale]

hit the try fix button first

[young wadi]

not working i tried

[young wadi]

also i get this error

[obtuse latch]

install the ui manager

[young wadi]

i did

only this is causing the error

[obtuse latch]

Then you need to install the missing nodes with the ui manager

[young wadi]

i did but this is the only issue i have ,i tried to update install it again but i keep getting the same error ,and i dont know why

[obtuse latch]

I can't really see anything the4e.

You're saying you're press g install issuing custom nodes?

[strong flower]

when installing custom/missing nodes don't forget to relaunch ComfyUI as there will probably be additionnal librairies to install before it gets usable

[silver nebula]

My ComfyUI starts up and gets to "Using Xformers Cross Attention" when these two dialog boxes crop up?!?!?!?

I can proceed to use ComfyUI - but when it develops other problems - I am then not sure as to whether this Python/Tensor difficulty is the problem in the first place?!

[strong flower]

You can try to use ComfyUI through Pinokio, it will install it in some sort of virtual environment with its own Python version etc.

[karmic creek]

I had a similar issue and could not fix it but made a completely new standalone install and worked great after that

[silver nebula]

It suggests that ComfyUI will be OK unless "The specified procedure could not be found'If you don't plan on using image functionality from torchvision.io, you can ignore this warning. Otherwise, there might be something wrong with your environment. Did you have libjpeg or libpng installed before building torchvision from source?"

So delete libjpeg and/or libpng - rebuild torchvision from source - and then install libjpeg and/or libpng

[silver nebula]

Github is down??? At least here in London UK ...

[silver nebula]

How do I CLEAR CACHE in ComfyUI?

I found an answer

[tender stirrup]

Every time that I start a runpod with comfyui, the video combine node has no h264. I solved it a couple of times with a pip install ffmpeg but the problem keeps popping, and now not even that is working. Any idea how to solve it?

[silver nebula]

PhotoMaker in ComfyUI

[slim root]

I know I'm just a shill for StableSwarmUI but have you tried installing that? It uses Comfy as a backend and does a good job making sure dependencies are installed. I also don't know how runpod works if it allows you to install whatever you want.

[silver nebula]

I've just installed Stable Cascade ... a new toy from Japan

[naive brook]

It's from Europe

Germany and Spain to be specific

[wintry fox]

trained on taiwanese chips using world wide images datasets

[silver nebula]

https://github.com/Stability-AI/StableCascade

[strong flower]

What is VRAM consumption ?

[silver nebula]

Memory on the Graphics Chip

If VRAM is too small, no output of ai images

[strong flower]

yes I meant how much does it consume lol

I heard around 20GB

[silver nebula]

Not tried it yet ... 🙂

[silver nebula]

Some more from ComfyUI/Photomaker - I call these Swaylr Tift!!!

[copper gale]

Sequential Image Loading suggestions please......

QUESTION:
what is a way/node/customnode etc.. to handle a continuous sequence of images feeding into process/model tree ....
ie.. i have a short clip of 100(or more) frames in sequence like ;

someseqname.%04d.jpg

and want to play that through in sequential frame number order to a model and do IMG2IMG, or also feed to controlnets etc...

whats out there thats nice to play with ?

be great if i also had multiple loader/readers that could stay in sync, as in,
I have a rgb image loader and another alpha or depth loader and they should stay in sync when feeding into rest of the process tree, ;

someseqname_rgb.%04d.jpg
someseqname_alpha.%04d.jpg
someseqname_depth.%04d.jpg

keeping them loaded in order and all 3 in sync during processing.

any help would very welcome as I know this space has so many options constantly changing.

[sly sundial]

Does it even run locally?

[waxen harness]

wondering the same thing, is there any workflow to properly use it/is it even supported by anything yet?

[distant pendant]

it works with diffusers

[strong flower]

It's now available in Pinokio, it takes around 15GB VRAM so patience if you have only 8GB because estimated time is 45mn lol

[sly sundial]

Never heard of Pinokio! 😄
I have a 4090, so should be ok.

[strong flower]

it's a great tool by CocktailPeanut that allows you to install multiple AIs in 2 clicks 🙂 it manages everything for you

[sly sundial]

Thanks, but I'm not keen on "auto" installers and will see if I can get it working in ComfyUI.

[strong flower]

I wasn't until I changed my hard drive and had to reinstall everything lol

[sly sundial]

From Comfy... "it's probably going to be natively supported in ComfyUI before Saturday"

[strong flower]

I could never wait to open my christmas presents on the 25th morning...lol

and it's just to test it, I now know it's out of my RTX3070 league until the optimization magicians put their hands in the hat 🙂

[slow yoke]

you celebrate on dec 25th why no jan 7th

[obtuse latch]

You shouldn't need to manually manage packages, outside of custom nodes, Comfy builds it's own env with everything it needs.

Specifically torch and any core modules like that.

[rough portal]

Probably been asked a thousand times now but is there a way to change the output location in Comfyui? I'd like to move all of my outputs to a Working Project Drive instead of them eating up my main drive.

[buoyant iron]

how do i create padding for an image in a comfy node? my code gives me an error:
TypeError: color must be int or tuple

[olive yoke]

try debugging your code. you will see that an image is a torch tensor

[olive yoke]

you can set the output directory using a command line argument

in windows, you can also mount a junction

[buoyant iron]

so i have to use image.permute in some way? (i'm very new to this - only learing for some simple comfy nodes)

[olive yoke]

your best bet is to look at the code in PreviewImage and SaveImage

[surreal whale]

no i havent used it just saw it https://github.com/kijai/ComfyUI-DiffusersStableCascade

[buoyant iron]

thanks. got it working:
crop + padding

[silver nebula]

Venv in ComfyUI Standalone only

[raw maple]

I would like to connect the inpainting window to the final save image window to create a mask on the save image/preview image photo and pass it on to the next save image window because I do not want to load the photo into the load image window can someone help how connect that?

because without this i must first save image than upload image use mask and again save image

[rose storm]

I always use Preview Image. You can then rightclick on it and "copy clipspace" and then "paste clipspace" into the image loader

[raw maple]

can u share your workflow with this?

[obtuse latch]

would you consider this prompting better:

Temperature Type: Warm
Moisture: Oceanic
Weather Condition: Clear
Accessibility: Easy
Biome Type: IceCap
Vegetation Density: Abundant
Elevation: Hill
Wildlife Presence: Moderate
Magic Intensity: Absent
Resource Abundance: Abundant

Than
A warm Icecap on a hill...

[copper gale]

hi was wondering if for loading sequential png images this is best way ?

WASSasquatch has the 'Load Image Batch' node

ComfyUI-Manager ID23
https:// github.com/WASasquatch/was-node-suite-comfyui

One trouble with this for me is ideally need the concept of a scrubble timeline as this node loops and if i load multiple copies for rgb, depth, alpha etc.. then they also tend to fall out of sync... does anyone have ideas/workflow/nodes that might be even better?

[obtuse latch]

You can write a script / app to hit the api that just runs whatever you want.

[west ember]

what is the minimum and ideal GPU VRAM to play with comfy UI? I got a RTX 4070 and A4000

[surreal whale]

Quite a wide range and depends on what you are doing, People have gotten it running on as little as 4GB

you got 16gb cards that's a comfortable amount

ideal is 4090, 3090 24GB is up there also

[west ember]

4090 is overkill

[surreal whale]

its about the memory, can load bigger/more models, higher resolutions easier

technically ideal is A100 i believe 🙂

[raw maple]

i have 4070 and no problem use SDXL

[sly sundial]

https://tenor.com/view/no-donkeys-shrek-gif-1107972158425297474

[sly sundial]

If anyone is looking for Cascade wokflow ideas...

This will of course change when the official Comfy nodes are released.

[silver nebula]

Is this done locally or via a notebook? It is excellent, whichever way!

[sly sundial]

Thanks, locally

[silver nebula]

How do you run this locally at all? I have it in a folder on my C: Drive, but no obvious bat/main.py file?!

[sly sundial]

Do you have ComfyUI?

[silver nebula]

Yes

Just put your PNG onto the desktop?

[sly sundial]

You need to install the node from here and install the diffusers: https://github.com/kijai/ComfyUI-DiffusersStableCascade

If you wait a couple of days, it should be in ComfyUI natively by then.

[silver nebula]

Into my ComfyUI Folder? Or a separate standalone folder?

[sly sundial]

Go into the manager, install from URL "https://github.com/kijai/ComfyUI-DiffusersStableCascade.git"

Do you use standalone version of Comfy?

[silver nebula]

Yes

I have installed the Diffusers, ComfyUI is now re-booting

OK, now that the Diffusers are installed, how do I run Cascade at all?

[sly sundial]

Then from command line you need to install diffusers, in the ComfyUI directory, run .\python_embed\python.exe pip install git+https://github.com/kashif/diffusers.git@wuerstchen-v3

[sly sundial]

ok! 🙂

[silver nebula]

I don't have python.embeded as it is not Standalone

[sly sundial]

You said it was when I asked...but never mind 😄

[silver nebula]

Oops my mistake

[sly sundial]

Then just add the cascade node and connect a preview node...done!

Optimal resolution, that I found, is 1536x1536. Better results than 1024x1024

[silver nebula]

Which GitHub repo are you using for Cascade?

And does it install in the Custom Nodes folder? For some reason, I have installed it as a Standalone package!!!

https://github.com/Stability-AI/StableCascade.git

This one?

And I should install in Custom Nodes?

The Setup Page specifies a setup for a Notebook ... little info on ComfyUI

[sly sundial]

No need to do anything other than what I said earlier

Those instructions are the only option for comfyUI right now

That I know

[silver nebula]

pip install git+https://github.com/kashif/diffusers.git@wuerstchen-v3 in the ComfyUI Directory? OK - and not Custom Nodes? Gotcha

[sly sundial]

Doesn't matter where you run that part, if it isn't standalone ComfyUI

It will just use your local Python anyway

[silver nebula]

...its building wheels...

Finished

It upgraded diffusers 0.24.0 to 0.27.0

[sly sundial]

The models are automatically downloaded when you start ComfyUI

[silver nebula]

OK, I'm gonna give it a go ... 🙂

[silver nebula]

I have downloaded and installed Cascade - but cannot find the node on the ComfyUI desktop?

[sly sundial]

Any errors on startup?

@silver nebula

[sly sundial]

Need the details or can't help

[silver nebula]

I am going to wait until saturday 😉

[silver nebula]

Error occurred when executing StreamDiffusionCreateStream:

Cannot load because down.weight expected shape tensor(..., device='meta', size=(64, 640, 1, 1)), but got torch.Size([64, 640]). If you want to instead overwrite randomly initialized weights, please make sure to pass both low_cpu_mem_usage=False and ignore_mismatched_sizes=True. For more information, see also: https://github.com/huggingface/diffusers/issues/1619#issuecomment-1345604389 as an example.

[sly sundial]

I don't think 8GB is enough to run Cascade. If anyone knows different...?

[sinful raven]

i'm using modified comfy node and it shows 7,5Gb Vram usage: https://github.com/ccvv804/ComfyUI-DiffusersStableCascade-LowVRAM

[sly sundial]

Oh right! Didn't know there was another one. Maybe @silver nebula wants to try that.

[raw maple]

what is difference betweent his and normal SDcascade from github? only better for low cram or this also do worst quality image?

[sinful raven]

it downloads same models, so i assume should be no difference

[raw maple]

ok but if i install before other cascade how remove full instalation previously and install new this?

i dont want errors and 2x more space install second cascade

[merry ermine]

comfyui uses 10% gpu in idle, is that normal?

i narrowed it down to this custom node, crystools causes idle gpu usage of around 10%

[slim root]

Maybe try what they are talking about here?
#🐝｜swarm-ui message

[merry ermine]

try what

can you explain how this is related

[slim root]

What Comfy was talking about. To clear vram offload the models to ram. Dident know if that would help your issue or not.

[merry ermine]

my issue is that crystools causes idle gpu usage for no good reason

it shouldnt do that

it updates the performance metrics twice a second, thats not enough to cause 10% idle gpu on a 4090

either this stuff has a crypto miner hidden in it or there is some other massive inefficiency in it

[fleet thorn]

helppp

[slim root]

o sorry, I assumed without really reading your issue. I am definitely not having that problem but I also have it installed using Swarm.

[merry ermine]

you should check your idle gpu usage with it installed

[silver nebula]

Saturday!!! 😄

[obtuse latch]

you understand that all modules written are done by people that may or may not have much experience.

This is what you get when you use a open source application.

It's all in pure python, you can read through the code.

[merry ermine]

how does someone manage to use 10% of a 4090 just to read basic parameters such as vram and usage metrics

https://tenor.com/view/huh-cat-huh-m4rtin-huh-huh-meme-what-cat-gif-27377993

[obtuse latch]

Read the code and find out.

[merry ermine]

I did

[silver nebula]

I am PINOKIO-ing my way into CASCADE

[obtuse latch]

And?

Did you find a crypto miner?

[merry ermine]

no but i found out that each time gpu status is obtained the getstatus function enumerates all devices and rebuilds the lookup tables in a loop EACH CALL

also it uses the same loops over all devices and calls the low level vram function instead of simply using the inbuilt already used torch function to get vram

also even if the flags to not monitor the gpu are set to true, it still monitors the gpu and does these api calls

[obtuse latch]

Yeah I wouldn't expect people submitting to open source plugins to put much effort into optimzing.

[merry ermine]

so in idle it basically just spams the GPU to get all of the info over and over, like handles etc

instead of using the already active torch vram functions used by comfy it uses the low level nvidia api

so it adds a lot of redundancy, overhead and is not optimized at all

to improve it the guy could just get the device handles and index in the class initialization instead of calling it every time the utilization is called from gpu

[obtuse latch]

You can submit a PR

[merry ermine]

i can also just complain

https://tenor.com/view/byuntear-sad-sad-cat-cat-meme-gif-25617057

[obtuse latch]

Yeah, I guess you can.

I'd at least make an issue.

[merry ermine]

done

complaining complete

professionally

easiest fix would probably only start monitoring when a queue entry is made and automatically terminate it when queue is empty

but still this monitoring impacts performance

https://tenor.com/view/crazy-cat-insane-cat-help-my-cat-is-insane-help-my-cat-cat-gif-21688581

[olive yoke]

i don't think this shows you what you think it does

if you want to query GPU usage on nvidia, use nvidia-smi

[obtuse latch]

What is randomize the orientation mean?

[distant pendant]

idk if there's any node for that but you could make a very simple switch with a seed

[distant pendant]

I didn't try but I think this should work

import random

class Switch:
    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "a": ("INT", {"max": 8192, "forceInput": True}),
                "b": ("INT", {"max": 8192, "forceInput": True}),
                "seed": ("INT", {"max": 2**64 - 1}),
                "enable": ("BOOLEAN", {"default": False}),
                "randomize": ("BOOLEAN", {"default": False}),
            }
        }

    CATEGORY = "Switch"
    FUNCTION = "switch"
    RETURN_NAMES = ("A", "B")
    RETURN_TYPES = ("INT", "INT")

    def switch(self, a, b, seed, enable, randomize):
        if not enable:
            return (a, b)

        if randomize:
            random.seed(seed)
            enable = random.choice((False, True))

        return (b, a) if enable else (a, b)

NODE_CLASS_MAPPINGS = {"Switch": Switch}

can also remove forceinput if you just want to input w/h in this node I guess

is there a random list node by default?

can't recall any

I think the idea was just to switch between portrait and landscape randomly, as in like use width as height for latent or the other way around, but that probably works too, idk

[obtuse latch]

I don't see why you don't just que up 50 of each, sorry meant to ping jab

You're spending time trying to bruteforce a simple solution.

[buoyant iron]

what value (something like NULL) do i have to set as mask output from my node, so the receiving node behaves like the optional mask connection was not used at all?

[distant pendant]

probably just None?

then you check if mask is None:, not just if mask: cause that's ambiguous on tensors

[buoyant iron]

so i use:
mask_out = None
return (mask_out,)
?

(sometimes i want to return a mask)

[distant pendant]

that should work, in theory

[buoyant iron]

and None is the correct form to specify NULL in python?

[distant pendant]

also in the receiving Node, use mask = None, in the function definition if you have it as optional

yes

[buoyant iron]

i can't control the receiving node - only my sending node - but for the receiving node the mask input is optional

[obtuse latch]

You'd have to add it to a custom node

[distant pendant]

you can save it as whatever.py and put in root_comfyui_dir/custom_nodes

[obtuse latch]

It's just python.

[distant pendant]

just give it a try and see what error you get I guess, but it should be fine

yeah, you can put loose files in the custom nodes dir as long as they have NODE_CLASS_MAPPINGS

[obtuse latch]

Hey lads, anyone know if this would prompt better?

[surreal whale]

im a bad/lazy prompter, that would be too much typing for myself but give me some dropdown boxes and im selecting stuff

[distant pendant]

I don't think any model was trained with captions like that, natural language or tags likely work better

[untold tide]

hi, was wondering can comfyui train lora? i recently switch from a1111 to comfyui, so i'm back to newbie status 😛
if anyone got a workflow for it or good tutorial, please send me a link, appreciated 🙂

[eternal idol]

hi, could you give me some help with this error? I can see it's about memory, but can't understand what to do with that. I'm running the ComfyUI on RunPod with RTX 3070

[strong flower]

are you on a recent Nvidia driver ? Since a few months it allows data to be shared on VRAM and RAM when you go over your max VRAM, it gets damn slow, but at least it works lol

[eternal idol]

don't know, I run ComfyUI on RunPod using the Jupyter Lab template..

how can I get that info?

[strong flower]

Oh it's not a local install, so you'll have to know how much VRAM is available and try to not go over it I guess :/

[eternal idol]

and is there a way how to go around in terms not going over the VRAM? it appears when trying to use IPAdapter based on this workflow https://www.youtube.com/watch?app=desktop&v=oYjEFHb--RA

[distant pendant]

ipadapter was pretty vram hungry last I checked, and it adds up the more images you use

probably not much you can do with 8gb if it's sdxl

it's more of a clip vision encoder issue really, maybe if you run with lowvram or something it will fit

[eternal idol]

is there a way how to add that param when using the RunPod?

[distant pendant]

no idea how you start comfy on runpod but probably

[eternal idol]

it uses the AI Dock.. I'll try to find the launching file.. could you give me the exact param how to place it there?

[distant pendant]

iirc it's just --lowvram, if you start main.py with -h you can see all the args

[eternal idol]

I haven't found out how to start it with that arg but it seems it works just fine with 1.5 model so I'm gonna use that for now..

[vocal grail]

is that a diffrent performance hack? is it better than xformers? I am trying animate diff and getting out of memory so I want to try something different

[sly sundial]

No, just something I'm using couldn't use xformers. You can compare the difference yourself, but I haven't

[vocal grail]

ok ty

[slim root]

Any videos by Scott https://youtube.com/@sedetweiler?si=b_NlYr5ujoeiKg7s

If your coming from Auto you may want to check out StableSwarmUI it's backend is comfy so you can still do the normal workflows but front end in a nice clean UI. #🐝｜swarm-ui

[untold tide]

i appreciate the info ❤️

[distant pendant]

you can't train loras inside comfyui itself (well, you probably could if you hooked up kohya's scripts/diffusers to it but there's not much point staring at a node with a progress bar, might as well stare at the console instead)

[plush steppe]

hello everyone, i wanted to ask if anyone has already loaded the new stable cascade model, inside comfyui? i am anxious to try it and i still can't find anything that can help me.

[tacit fulcrum]

https://media.discordapp.net/attachments/1002602742667280404/1208086880801787974/image.png?ex=65e20195&is=65cf8c95&hm=7c5bd71ecd40a06309083f91fa026fedf5fee71ff43414d8efe64a76805acbb1&=&format=webp&quality=lossless

is a there a compatible clip vision model for the face plus sdxl model for ipadapter?

my current combination doesnt work

if I use the non face model it works

my checkpoint is based on sdxl so I am stuck here

[slender marsh]

hey ive finally got facereactor node to load but ive noticed when its processing the swap portion of the workflow its only using my cpu

any idea why

[untold tide]

i want to know this as well, i want to use stable cascade using comfy. i'm using the cascade demo UI, the result is pretty good, but there aren't many function, so i was hoping to get that in comfy with controlnet and lora and upscaler etc. if you found out how to do so, let me know as well 🙂

[slim root]

I'm sure Comfy will be releasing something official soon. So far a few days after a model release official support gets published.

I think built in SVD support was like 2-3 days or something like that.

[sly sundial]

Take a look at my Cascade workflow on civitai

https://civitai.com/models/119257/gtm-comfyui-workflows-including-cascade-sdxl-and-sd15

[pulsar ivy]

Hey guys, I'm working on a commission project and was hoping someone could point me in the right direction.

The project is to essentially take a photo of a fish (with a plain white background) and make it more realistic, detailed, and have a specular/wet effect while maintaining details like the scale pattern and black dots.

I've tried using ksampler with low denoising, using controlnet tile and IP2P, both with and without an IPAdapter of the original image. Admittedly controlnet may have been set up wrong as I was getting nonsense, but I couldn't see any obvious issues. Regardless, no luck with either option.

There was some luck with upscaling and sharpening, but that didn't really help with making it realistic or adding the required effects.

I'm starting to think a custom lora may be required but hoping to avoid that if it's possible. Appreciate any pointers you all can provide.

Also, mostly unrelated, if anyone wants to take on some paid commission workflows, shoot me a message for more details.

[thorny lily]

outdated and silly

https://gist.github.com/comfyanonymous/0f09119a342d0dd825bb2d99d19b781c :D

paynomindtothisnotexistingatthetimeyouposted

comfy's awake and implementing things :D

[jagged trout]

can anyone help with this??, onyx error

[jagged trout]

D:\a_work\1\s\onnxruntime\core\session\provider_bridge_ort.cc:1193 onnxruntime::ProviderLibrary::Get [ONNXRuntimeError] : 1 : FAIL : LoadLibrary failed with error 126 "" when trying to load "E:\comfyui\ComfyUI_windows_portable\python_embeded\lib\site-packages\onnxruntime\capi\onnxruntime_providers_tensorrt.dll"
when using ['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider']
Falling back to ['CUDAExecutionProvider', 'CPUExecutionProvider'] and retrying.

[sly sundial]

Outdated by a few hours, but certainly not silly. If there's a way to get better results from Cascade then please let us know. At the moment I have to refine it with SDXL, because the Cascade output is never detailed or sharp.

Some before/after examples

All of these examples are now using the new Cascade nodes and recommended flow from @storm folio.

I've increased steps to 80/20, but it still needs to be refined.

[jagged trout]

someone help???

[strong flower]

I'm interested on testing it if you don't mind 🙂

[sly sundial]

I've just published it! 🙂
https://civitai.com/models/119257?modelVersionId=347573

[sly sundial]

Updated to v2.1 to get rid of the discontinued Fooocus KSampler nodes.

[naive brook]

I think Cascade broke after commit 6c875d846b54ee34eb032fdb790a2fd1621d18f4. Not sure about other models. It bombs with: RuntimeError: The expanded size of the tensor (20) must match the existing size (2) at non-singleton dimension 1. Target sizes: [2, 20, 77, 77]. Tensor sizes: [2, 77, 77]

checkout 805c36ac9c0611a483b1f494e2dfe6b67f09fe36 works.

[storm folio]

should be fixed

[elfin bison]

Hi there, is it possible in stable diffusion + comfyui to set up natural light instead of a photorealstic studio like one?

[whole fossil]

2048x2048, from last night’s workflow

[untold tide]

hi galaxy~~~ i check out your workflow yesterday about cascade, and saw you update a new one today. was just wondering which nodes i should be adding after the cascade image and to link it to ultimate sd upscaler? or can i add the upscale in between?

(sorry i was using a1111 for so long and just change to comfy, sorry if this question is too stupid T_T)

https://github.com/ssitu/ComfyUI_UltimateSDUpscale?tab=readme-ov-file

[sly sundial]

I'm about to go out, but the image will need to be encoded and then an SDXL model used with Ultimate SD Upscaler.

[untold tide]

i see, i will check if there is a related nodes for it to encode i guess, thanks

[sly sundial]

Just a VAE encode, attached to the sdxl model vae

[strong flower]

it seems much more difficult to do hybrids compared to stock webui, but in Comfy I'm using FP16 models not sure if it's related. On webui I got a turtle and human hybrid on 1st try (maybe luck ? didn't try more), in Comfy I can't do it for now, no matter the resolution

[lone cape]

I have no success with Comfy's cascade workflow. Simple prompt with 4-5 words working well, but if the prompt longer and more difficult I always got unuseful result. Cant figured out how to solve. Do I wrong something? The test prompt working well in all other workflows (pic 2).

[distant pendant]

https://github.com/comfyanonymous/ComfyUI/commit/11e3221f1fd05d49b261ccec7dd99b704a86a89f
are there fp8 weights for cascade somewhere or is there some arg to enable fp8 in general? can't really find anything

oh wait, is it one of these?

fpunet_group.add_argument("--fp8_e4m3fn-unet", action="store_true", help="Store unet weights in fp8_e4m3fn.")
fpunet_group.add_argument("--fp8_e5m2-unet", action="store_true", help="Store unet weights in fp8_e5m2.")

[wraith prawn]

https://github.com/comfyanonymous/ComfyUI/issues/2157

about fp8 cascade
https://github.com/comfyanonymous/ComfyUI/issues/2785#issuecomment-1950300724

[distant pendant]

thanks

[thorny lily]

the diffusers hack was silly -- that message was mostly just a funny way of sharing the comfy cascade support link here lol

[jagged trout]

pld solve this

[raw maple]

Guys, can someone share some good workflow for

1. Text2Video where i can put frames "0", -text "25". text etc....
2. v2v also + qrcode and upscaling?

Everywhere from YT reddit or civii they create very big layout with thousends windows:D
Ty!

[gusty bridge]

hey friends, iam currently switching to Comfy UI and get this error while trying to start the interface

any idea?

seems like it cant find the python folder. but i have it installed. anyways, can anyone provide the current python module for download? thx in advance mates

[elfin bison]

Hi there, i have a missing model named "morgan.safetensors" . Where can i find this missing model? Thanks

[silver nebula]

Got to system environment variables and add Python to PATH

All of a sudden, ComfyUI startsup with dark screen mode enabled - but no graph, no manager, and no nodes - just a blank screen!!!

[tacit fulcrum]

is german too

[silver nebula]

Drucken ... Taste means press any key

[gusty bridge]

https://tenor.com/view/annoyed-fuck-frustrated-gif-15331740

[gusty bridge]

can u screenshot

shit still isnt working

[gusty bridge]

i already did that

"automatically..." right?

[sly sundial]

Why don't you just use the portable version? It has its own Python environment included.

[gusty bridge]

iam using that

[silver nebula]

Delete VENV and start again

[silver nebula]

cu118 is better

[gusty bridge]

?

[sly sundial]

Oh yes, sorry, I didn't read the details properly.

[gusty bridge]

np

[gusty bridge]

why is the prev one better

[silver nebula]

Got to ComfyUI folder - see a folder named VENV - delete that folder, and when restart, ComfyUI automatically rebuilds the VENV folder

torch 2.2.0+cu118 is the standard

... I believe?

[sly sundial]

torch 2.2.0+cu121

[silver nebula]

OK, but I always use 118

[sly sundial]

Are you using the start script within that directory to start it?

[silver nebula]

Go to pytorch.org - it tells you the latest implementation of Torch (2.2.0) and automatically chooses Cuda 118 to compile with it

[sly sundial]

You don't need to worry about it with portable version. Anyway, from ComfyUI manual install instructions:

NVIDIA
Nvidia users should install stable pytorch using this command:

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121

This is the command to install pytorch nightly instead which might have performance improvements:

pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121

[gusty bridge]

folder isnt existing

[sly sundial]

Have you manually opened it to see if it's there?

[gusty bridge]

yep

[sly sundial]

and?

[gusty bridge]

ita not there lmao

[sly sundial]

Then it sounds like you've not installed it properly....?

Corrupt download? Have you tried deleting and starting again? Do you see the misssing files in the zip file?

[gusty bridge]

yeah about to do that

[gusty bridge]

working now - i had to give extra premissions to the folder, even tho i was already the admin

lol

[silver nebula]

Haunting Steampunk, Shaun2 and Modernart LoRAs, TurboVisionXL_Superfast as the checkpoint, Reactor Face Restore

[gusty bridge]

omfg Comfy is so fucking siiick! holy shit! i was using the fooocus interface before

it has its + but comfy is something different

holy shit

[silver nebula]

Glad you're impressed

[gusty bridge]

thx mate for helping yesterday

appreciated

[gusty bridge]

In Fooocus, I could see how the images were generated and I could observe the batches that had been rendered so far while the others were being completed. Does it work the same way here?

and if yes, where can i activate it

[silver nebula]

Set up an Image Preview node

[gusty bridge]

thx, yeah i havent checked all the nodes yet. iam on it rn

[eager escarp]

haha, just updated comfyui for the new stable cascade stuff 20 min ago, and now I gotta update again! thank you comfyanonymous for your hard work!

[worthy ember]

Any Comfy UI top level expert ?. can you send me friend request. [Paid Task]. need small workflow that is from SD ui

[eager escarp]

might want to just give a quick summary of your question/request?

[worthy ember]

I am using some specific settings in A1111 with 4-5 ControlNets. But A1111 too slow. Wanna shift to Comfy ui for faster generations But I know basics only. Basically I need img-img workflow in comfy UI with same control net settings that I am using in A1111

[eager escarp]

what's the workflow? the sequence of controlnets

[worthy ember]

reference, tile+color fix, canny, lineart realistic, tile

Please don't ask me why😅

[eager escarp]

you're all good. fyi, sadly reference doesn't exist in comfyui. there's a knockoff version of it but it's really not at all the same. problem is it's really a preprocessor not a controlnot

[worthy ember]

ohh. Is it possible to implement?🥹

[eager escarp]

people have hoped it would happen for months

i'm sure it's possible

[worthy ember]

can you DM me please

[thorny lily]

it's not a preprocessor either, that's just where the auto webui extension shoves it

it's actually its own thing entirely

[eager escarp]

ahh gotcha

i just heard someone say that and never looked into it tbh, thx for the clarification! 🙂

[untold tide]

how to not save the workflow in the generated picture?

[tacit fulcrum]

Hi, are there any xl models for faceid

all I see are 1.5 based

I saw there were lora models in comfy manager

but no ipadapter model

are there any?

oh I see models now

they require 1.5 encoders

[lone cape]

Over the weekend I tried out an old plan of mine, to try out if the node could be generated by reading an external file at load time, and it worked without a problem. The source of the node is a .toml file that I generated from my prompt collection, the node is generated by as many prompts as I have at loading. This was done with a back-end only. I'm still thinking, if selected one prompt, all the other combos should be deleted to 'None', because if not, 2 prompts will be mixed up. What I might add is that can also select the prompt source .toml file in a top list and re-generate the full node by the selected file.

[silver nebula]

ComfyUI - I just cannot get WAS Nodes to work!!

[lone cape]

I started to implement "emotions" as style by Wikipedia, and a little surprised how well the sd-comfy performs to the longer specifications of feelings. These are only development demo pics, the red framed the original without emotions, all others used same model, prompt and seed but added emotions, or combined two several. I will continue but little boring to copy details to .toml file 🙂

[silver nebula]

Cannot get WAS nodes in ComfyUI to work. Can anyone help?
It comes up in a red box in ComfyUI Manager ... !

[raw maple]

Is someone here who can help me mix this two workflow together? I m tried add group Adetailer to my workflow but something is wrong and dosent work. I upload image here what group i want add to workflow what i have below....Also u can see YT that workflow to understand more, i cant...:/

My workflow

This detailer is from https://youtu.be/J_fPCYD7vbc?si=oTTXn8p8AMQyDreb&t=280

[sly sundial]

Have you tried deleting the folder and re-installing them?

[olive yoke]

what is your objective?

[olive yoke]

sadly reference doesn't exist in comfyui. there's a knockoff version of it but it's really not at all the same.
what is it?

[eager escarp]

https://github.com/comfyanonymous/ComfyUI_experiments

[opaque escarp]

is there something different about the checkpoints linked from the comfyui page here https://huggingface.co/stabilityai/stable-cascade/tree/main/comfyui_checkpoints and the slightly bigger ones https://huggingface.co/stabilityai/stable-cascade/tree/main in this area? The diffusers initial node worked with the latter, but the comfyui implementation only works with the ones linked first.

[jagged trout]

can anyone help??

[surreal whale]

i totaly tried to click close on that error message

i dont know but it says you are mismatching SD versions

[jagged trout]

I'm trying to run Cascade model

[jagged trout]

try this workflow https://comfyworkflows.com/workflows/74b9638c-8162-43ba-92f0-3aee1290fc2d

[lone cape]

As I see the longer definition working much better than one word for example mood [something]. So this node output not the selected word but the longer definition. But usually im not a creator, im a developer, so I hope that working and useful for somebody.

[shy scarab]

Good afternoon, excuse the mess the render link mode is broken and i cant seem to fix it. ive been trying to use openpose the whole day and i cant seem to get it right, im trying to use it to generate realistic images however its not working

i tried to follow this guys guide and it worked for animated images however i tried to move it to another workflow and it doesnt seem to use the openpose https://www.youtube.com/watch?v=B6qilXLls0o&ab_channel=AIAngelGallery

anyone have some tips or other videos i can take a look at?

[olive yoke]

working much better in what sense? like a more exaggerated version of the expression?

you are trying to show a grid of moods

and you want the moods to be exaggerated?

or do you want more creative images generally?

[olive yoke]

i think if you want an expression to be legible on a face, you have to add exaggerated. there aren't aesthetic images in these datasets of "moods" because you cannot see moods. prompt stuff that you can see

left is "a woman in a bad mood" right is "a woman in an exaggerated bad mood"

words in the prompt that are not visible and not related to IP are noise

you might as well reduce CFG and steps instead

[sick badger]

anyone know how to clear these errors on comfy startup? not breaking anything, just annoying
i've readded them and then removed them, even removed the folder from custom nodes manually, error still appears

is there like a conf file somewhere that stores what node packs are installed i can remove lines from?

[merry ermine]

ComfyUI Manager -> Install Missing Nodes

[sick badger]

there not technically missing tho, i can add them just fine in the manager, but when i uninstall them the start up error is still present, would like to see them gone completely when the node is uninstalled 🙂

[merry ermine]

Could be nodes that are conflicting but i couldnt see how it would be 4 of them. Try to update all, if that doesnt work, delete each node folder and reinstall? Best I can guess at this point, honestly it doesnt even take long to reinstall comfyUI as a whole

cd comfyui
python -m venv venv
cd venv
.\scripts\activate.ps1
cd ..
pip install -r requirements.txt
python main.py```

Copy your models folder and models from animatediff / whatever other nodes you use

while your in venv might as well update to latest pip, python.exe -m pip install --upgrade pip

Or even easier, just use Pinokio and copy your models over

[lone cape]

Prompt and pics at queue, same seed, same model:
1: woman face
2: woman face, surprise mood
3: woman face, exaggerated surprise mood
4: woman face, (Surprise, the feeling caused by something unexpected happening.) -> from my style node

Interesting difference between "mood" and "emotion":
1: woman face, (Surprise mood, the feeling caused by something unexpected happening.)
2: woman face, (Surprise emotion, the feeling caused by something unexpected happening.)

[sick badger]

oh, i guess all comfyui installs show those 2 errors, installed on a new system that's never had comfyui on it, startup shows same errors, guess it's a basic system check or something

[merry ermine]

I'll update and try now.

[sick badger]

must be my ocd, not breaking anything just seeing the message makes me tweak lol

[merry ermine]

Oh, did you remember to update your directory in exports_whatever.yaml

I just updated, she's purrin like a kitten

[sick badger]

what does your terminal output after you start it?

[merry ermine]

i don't get what you're getting

[sick badger]

maybe it's a linux thing 🙂 everything still works great, i'm gonna chalk this up as a loss and move on LOL

[merry ermine]

seems to be running quite nice

[sick badger]

same 🙂

my brain gets hung up on console output messages, i see an error or a couldnt find and i freak lol, regardless of it actually being an issue

this is a production server so i tend to get even more so worried lol

[merry ermine]

there probably is an issue with conflicting nodes, but if your stuff is working, id ignore it

[olive yoke]

what are you trying to do?

what is your goal?

[olive yoke]

you can't see "feeling by something unexpected happening"

so it is just adding noise to the conditioning

[merry ermine]

several questions regarding instant loras:

1. can you use multiple instant loras at once?
2. can you save the lora generated locally, for later use?
3. if not able tosave locally, do you know other methods of saving an instant lora?
4. do instsnt loras work with trigger words or the classic lora:name:1 command?

[silver nebula]

excited, anxious, nervous, frightened - all intensive moods - and of course exaggerated, very, totally, completely, overwhelmingly ...

[dapper lichen]

Girl (Emily): Body shape: Slender and graceful figure, slightly thin but with a certain sense of vitality and agility. Appearance: Delicate and lovely face, big and bright eyes, soft outline, sweet smile. Clothing: Simple and comfortable spring clothing, which may be a dress or skirt paired with a simple top. The colors are mainly elegant pink, light blue or light yellow, reflecting youthful vitality. Hairstyle: Slightly curly long hair, or shoulder-length straight hair, simply decorated with a hairband or bow, showing a fresh and natural style. (╯°□°)╯︵ ┻━┻

[valid swift]

PhotoMaker Retro Ghibli style, a young girl with golden hair tied back into a simple braid.
wearing a humble dress.--nii 5character sheet, white background

[silver nebula]

ComfyUI using a @junior hull w/flow, NightVisionXL checkpoint, Underwater, Melting Candles and Future Cubism LoRAs

Rstore Face by Reactor does a good job on larger faces ...

[lone cape]

Interesting, for me the large faces good enough, I using size trigger to refine only small faces because the face restorer change too much of original.

[silver nebula]

I'm thinking of using Ddetailer - I wonder if it does all the faces? I'm putting ComfyUI into A1111 img2img for the ADetailer - it remedies all the small faces.

I then use Photohop and paint back in the main face

ADetailer in A1111 would have dealt with every face. ComfyUI Reactor Face Restore only does the larger faces

[desert sedge]

Hello everyone! I am trying to use InstantID in comfy and I am running into an issue... In the Cubiq github repo it says " you need to add it to your libraries together with onnxruntime and onnxruntime-gpu." but it doesent say what or where these onnsruntime things are or where i'm suposed to put them... The example workflow won't work without this, I have tried. Can someone give me a hand?

[lone cape]

I developed custom node for yolo models with triggers. For example if the face larger than the 3% of full area of input image, then ignore the refining process, because the face good enough no need to refine. Sometime I using triggers for lower value, if the refining needed for larger segment, for example for eyes and mouth. With these triggers easy to control which segments size must be refine or ignore. The first is too large, but two another example refined because was smaller than the trigger value.

[lone cape]

Just for testing can I drop one of your image to my triggered refiner?

[silver nebula]

I might d/load and try this, if I may?

[lone cape]

If you install primere nodes by comfy manager you can try, but the refiner's dev workflow where you can drop already generated images not public, but I can send you if downloaded the nodepack.

[silver nebula]

OK, will do, thanks

(CosmicLaca)

[lone cape]

Haha yes its me 🙂

[silver nebula]

Its d/loading all the face detailing items

[lone cape]

Yes if not downloaded yet.

[silver nebula]

Now Downloading: sam_vit_l_0b3195.pth

Done!

[lone cape]

I really dont like it, but users reported that they cannot use detailer because they do not have the segmentation models, so I made download instead of always copy the source and target paths. While all required models available in comfy model manager 😦

[lone cape]

Comfy manager -> Install models help im sure.

[desert sedge]

I will check

[chrome gust]

Hey guys, was experimenting a good upscale workflow based on new developed lightning lora. Found out when upscale the img, result have a lot of noise and unwanted details, the. Went on to latent upscale, the noise is a lot better, but with increasing denoise strength result start to get weird. When I try to use control net to provide some controlled generation it always give me an none type error, I used the SDXL version of linart control net model, my question is can you use controlnet for latent control at all? Maybe I should test it with regular sdxl model and workflow?

[heavy bough]

Hi there, I've been trying to figure out how to output batches of 5 images but each using randomly chosen ratios. Is it possible to have the ratio change between each batch generation? or do I need to generate the 1 image and then loop the whole process with a different ratio? I have confused myself greatly...

[lone cape]

Not for the ratio, but I using node for randomizing orientation. This not enough for you?

[heavy bough]

this could work, I'll give it a go, thanks!

[distant pendant]

I don't think you can have different dimensions in a single batch though

the shapes need to be the same

[heavy bough]

^ This is what I thought thanks for comfirming

[lone cape]

In comfy you can set Batch count under Extra options not just latent batch, so you can change anything if using batch under the Extra settings.

[distant pendant]

that is not a batch in the same sense, that's just querying multiple generations to run one after another

you can have different dimensions between those, yes

but you can't start a generation with a, say, latent batch of 4 and have the dimensions different there since the tensor shapes have to match

[lone cape]

I know, but the random size or orientation possible using query.

[distant pendant]

sure yeah, also an idea: you could technically have different shapes in the same batch if you only denoised some area of the image and then cropped to whatever you wanted, but would likely need some new saving logic

[heavy bough]

What I'm trying to get to is a workflow that can create a mood board, one prompt - 5 imasges different ratios that i either screenshot or drag the images into a prez editor etc

e.g. in Mid Journey I can use the chaos parameter

[broken spoke]

Hey can someone build a custom workflow. Asking for someone else

this person wil pay

[ancient bear]

Hi, I'm trying to install controlnet aux through comfy manager but it is giving some error, could anyone help out?

[slender marsh]

anyone got a workflow for lip sync?

[tranquil yarrow]

Using the Stable Cascade on ComfyUi I got this error:

Error occurred when executing KSampler:

linear(): input and weight.T shapes cannot be multiplied (77x768 and 1280x2048)

How can I fix it?

[distant pendant]

lip sync as in like face image to audio-guided video? something like sadtalker?

I don't think anybody ported sadtalker specifically but it sure would be interesting

[slender marsh]

Yeah image to guided by audio lip movement

[distant pendant]

don't think that exists for comfy but it does work in a1111

[slender marsh]

Also if there is a mov2mov lip sync where you can use different than original audio and sync to the new audio

[strong flower]

For lip sync to me the best solution right now is the wav2lip extension on A1111

[slender marsh]

For some reason when I do any mov2mov stuff in forge when I click the generate button it just hangs. Doesn't use processor, no errors in console, just does nothing

[slender marsh]

Thanks:)

[proper inlet]

Hey friend. This is my workflow.

Maybe I can try and help you figure out why it's happening?

The reason there are multiple clip text encoders in the workflow is because it needs to load a clip model for the cascade samples. That loader takes place before the clip text encoder, so a different text encoder that loads clip directly from the XL models.

That error may be occuring because clip is being routed from the wrong encoder into the XL models.

pulled those off to the side so you can see what I mean.

the bottom encoder routes to the ksamplers connected to cascade models, the top ones to the ksamplers connected to the XL models.

[proper inlet]

https://comfyworkflows.com/workflows/74b9638c-8162-43ba-92f0-3aee1290fc2d?version=578c2aeb-ddad-4891-a8f2-13298b4a8f78 Just updated with a very cleaned up v2

[crimson cloud]

Could someone explain how a work flow is transported from just copying a picture or dragging it in? Tried following YouTube videos but nothing happened ( have comfyui and manager installed) it's like I did not do anything at all. Is it only images that were made in comfyui that it will work from( tried images I made from automatic 1111 a while back.)

[lusty pewter]

If you're using ComfyUI, you can drag an image into the screen and pull the workflow in, however, that workflow has to be embedded in the image you're dropping in. If it's not, it won't open the workflow because the workflow is not known.

[nocturne garnet]

Only images made in comfyui have that feature

[crimson cloud]

Thank you for the explanation

[proper inlet]

I actually think some others do it too, but probably not to the degree that comfy can.

[thorny lily]

most UIs embed metadata in images - auto webui, comfy, swarm, ... all do

comfy embeds an entire raw workflow json graph, auto webui embeds a similar string of parameters, swarm embeds a detailed json of parameters

swarm ones if you pull into swarm you can import from swarm to comfy to get a raw workflow if you want that

[strong flower]

is there a way in comfy to output a png image with embedded workflow AND a jpg image without metadata ?

[unique grotto]

Some great work from Comfy (and the entire Stability team) to make a pre-release seem production ready. Created 50 images tonight using Stable Cascade in the time it took me to create 1 two days ago.

https://x.com/d_w_b7/status/1760926628568580489?s=20

Don (@d_w_b7) on X

A quick shout out to @ComfyUI for his quick work in making the @StabilityAI Stable Cascade research release feel production ready. Makes quick work of seeing the incredible prompt alignment and aesthetic quality.

https://t.co/Zl8Nl47jeb

Twitter · Feb 23, 2024, 7:16 AM

[unique grotto]

Search for references to "Exif" in the comfy code and available nodes. I found some, but didn't have time to look specifically for removing data.

[unique grotto]

Check the Workflow metadata field of the image using this tool. pretty amazing that comfy can store the entire WF in the metadata, but there may be limits.

https://exif.tools/

[wispy kelp]

So is this Cascade thing good? Is it the future? what's going on... Every week soemthign new....

[strong flower]

look what it can achieve with just a SDXL fine detailing post process workflow (the original Cascade image is almost this detailed)

[wispy kelp]

🤔

[sly sundial]

Yes, this is possible. If you look at my workflows it uses a node to save the image and you can choose what format and if it save the data or not. https://civitai.com/models/119257/gtm-comfyui-workflows-including-cascade-sdxl-and-sd15

[sly sundial]

These are just Cascade with no detailing

[strong flower]

Yes I'm at work and don't have access to my original images but it has great outputs, I just like the extra details of a little SDXL detailer because I'm the kind of person to open the images in a browser and zoom in 😅

[tidal agate]

how can i update to latest version of comfyui

[lone cape]

git pull on comfy dir

[sly sundial]

Are you using the portable version? I think that has a script to update it.

[untold tide]

is there a node or function that i can load the txt file to comfyUI that i generated before in A1111?

[errant glen]

Is it possible to change the default value for the "Save Image" node? I would prefer the save path to be in daily subfolders and set the default value to %date:yyyy-MM-dd%/o instead of ComfyUI

[olive yoke]

i was just exploring this as a patch. are you familiar with any other tool you use which lets you specify format strings?

the answer to your question is no*

there aren't any text manipulation nodes. So it's clunky to even pass an argument to it

[flat bobcat]

hi is there any update to use cascade control net models on comfyui !

also any idea how to compress model to ,75 as mentioned on cascade manual ! they tell givebetter result !

[olive yoke]

which cascade control net models? can you share a link?

[ivory ruin]

whats the best model for sd 1.5?

[calm solar]

as in comparison to these 4?
it depends on what you are trying to do... a trained model to your needs is usually the best model in my opinion

[boreal panther]

Hi friends, I'm beginning to immerse myself into the world of text to image generation and trying to get ComfyUI running locally. It's running, but it shows Device: cuda:0 AMD Radeon RX 5700 xt : native - as you can see, that last line looks a little contradictory - "CUDA" being related to the Nvidia cards. I do have a GUI, I haven't generated anything yet as I don't know how. I just wanna make sure I'm setup properly.

Specs:
Ubuntu 22.04.2
Ryzen 7 3700x
Radeon RX 5700 XT

[boreal panther]

I'm generating an image now, and not sure it's using just the GPU. It's taking awhile, and my CPU monitor shows different cores spiking to 100% usage methodically

[boreal panther]

and failed

[ivory ruin]

so im trying to use a lora which is called atmosphere photohraphy but it only kinda works with the epic realism model

[calm solar]

yeahhh using other peoples loras can be tricky :/ is hard to know what they trained it with or what keywords specifically and if they did it right :/ I did a google and saw some people also using juggernaut (non xl) for that lora so idk

[ivory ruin]

oh i see thank you that might work

[calm solar]

are u using the main branch of comfyui

[boreal panther]

latest stable that's linked on their github

[calm solar]

have u installed directml and running with the flag

[boreal panther]

I'm on Ubuntu

I'm trying again with HSA_OVERRIDE_GFX_VERSION=10.3.0 python3 main.py

It's running, hasn't failed yet. Not sure how long it's supposed to take. I loaded the Juggernaut XL .safetensors in the checkpoint folder, selected it in gui and ran with all the defaults

[calm solar]

ah alright, i havent used comfy on linux before sorry...

are u using the SDXL loader in comfyui or just the regular loader?
theres a bit of a difference between XL and non xl workflows

[boreal panther]

I'm not even familiar with that. So idk. Got any recommendations for reading to get up to speed from the ground up? I'm a computer programmer, so pretty savvy

I do see Requested to load SDXLClipModel in the terminal

[calm solar]

XL can take a little while, would recommend just using a 1.5 model to try get things working

and yehh that could be the different loader being the issue

[boreal panther]

Is there a flag I can use to tell it to run with non xl workflow?

[calm solar]

idk if i am allowed to share others youtubes but this guy is pretty good
https://www.youtube.com/playlist?list=PLH1tkjphTlWUTApzX-Hmw_WykUpG13eza

[boreal panther]

Sweet

[calm solar]

if you get a basic 1.5 workflow setup and still having issues
i use comfyui plugins for my XL workflows, i forget the name but dw about plugins until u get the basics working

[boreal panther]

Is there a flag I can use when starting comfyui that tells it to use the 1.5 workflow?

I'm realizing that question might not even make sense.... I'll do some reading and watching. Thanks

[calm solar]

its a node, like Loader Simple I think.. and SDXL Loader or something

u wanna just use the regualr loader for 1.5 and XL for Xl

i use "efficiency nodes" for XL
if u get the "comfyui manager" itll make it easier to get plugins and XL working

i might be lying about the loader and it might be like a SDXL Clip or something thats the difference, i havent done vanilla SDXL workflows

[boreal panther]

Cool cool, I've been playing around with midjourney and it's still mind blowing to me. I'm trying to replicate that quality locally. Might take some learning

[calm solar]

its alot of fun 😄 good luck

[gusty bridge]

i had the same "problem" at the beginning. but u will get there. i got there too

[silver nebula]

IPAdapter NightVisionXL

[silver nebula]

Apparently the server-power behind MJ will make desktop MJ hard to replicate - u will either pay $£$£$£$ - or wait for the technology to catch-up!

But I agree - it's MJ quality that I seek in Stable Diffusion. SD3 looks like it's going the right way ...

[merry ermine]

1 finger is going the wrong way on top pic

[echo spindle]

i have some problem with the face detailer pipe, keep getting this error : return self._op(*args, **kwargs or {})
any idea on how to solve this ?

[fallow jackal]

How didn't I know before that "Load Image" makes transparency masks automatically?

[echo spindle]

its because its a png, your mask is included in the alpha pass if i'm not mistaken

[manic moat]

Hello guyz how ru all please can anyone help me how to do like this please help me out I installed stable diffusion and comfyui both but don’t know how to fix like this https://www.instagram.com/reel/C3Pw0NAvNA6/?igsh=b3AzMzNxdGZ2amZl

aiart.mania

Transform your sweet memories into visual spectacle.

Did you miss the photoshoot for the past memories for your kids or for you or for your family.

Dm if your interested to get for your loved ones.

#sita #godess
#aiartmania #aiartportrait #aiart #aiphotos #AiPhotography
#aiimages #aiartwork #aistillsCustomPortraits
#AiAssistedArt #Preciou...

Instagram

[tranquil parcel]

Hey is there a Comfy node for live Voice inputs and is there also a node for streaming in text real-time from an external doc?

[silver nebula]

InstantID - prompt = comic character. graphic illustration, comic art, graphic novel art, vibrant, highly detailed - a swiftly tailored SuperHero Comic Champ!!! 😄

For comic "character" read "police officer aiming a gun"

This one was a surprise - I swapped out "comic character" and made it "kissing Joe Biden" - but Karly Kloss snuck in there somehow?!?!?!?!? 😄

[main osprey]

when i save the file, how to access the input filename?

[silver nebula]

Either just use ComfyUI - or if you want a date 2024-02-25/ComfyUI - or something else like IPAdapter/ComfyUI or 2024-01-25 InstantID/ComfyUI etc - these folders will end up in the output folder at ComfyUI

InstantID "MarJo BidRoe!!!"

Sorrrry about the fingerrrrrs

[jagged trout]

yesterday this workflow was working normally but now I'm getting thios error

[jagged trout]

is there anyone

[main osprey]

want to access the filename_text, how do i do that?

[silver nebula]

InstantID - "Baribifying" Margot Robbie

[fallow jackal]

Right click on "Save Image" node → "Convert filename_prefix to input"

[untold tide]

is there a function on SD and comfy, that have the midjourney "blend" function?
ipadadapter can use only one photo as reference, but the blend in MJ can use up to 5 pics
so i was wondering is that the same tech behind it.

[thorny lily]

ipadapter can take several, as can XL image prompting

[silver nebula]

You can blend at least 3 photos using IPAdapter!!!

See Scott Detweiler's Workflows, or even mat3o's

[thorny lily]

or just use swarm and drag images into the prompt box

and if you're curious about the noodles just import to comfy workflow editor

[flat bobcat]

https://huggingface.co/stabilityai/stable-cascade/tree/main/controlnet

[flat bobcat]

https://github.com/Stability-AI/StableCascade/blob/master/inference/controlnet.ipynb

[lost depot]

greets! is this the right place to ask a question about comfyUI and Animatediff?

[lone cape]

Hi, I just created node to switch between new concepts for fast result comparison. The node (and workflow) handles: SD, SDXL, SD-LCM, SDXL-LCM, Turbo, and Lightning if downloaded safetensors, Loras, and Unets to right paths, and with Lightning custom handle 3rd party Lightning checkpoints. Cascade not finished yet. The workflow image contains workflow, Primere nodes required:

[worthy ember]

I have small PAID TASK . Anyone interested (must know comfy ui, instant id, stable cascade things) ? you can DM me for details

[sly sundial]

Why is it that I can see these nodes in one window, but not in any other, even after a restart?

Other window

I only seem to be able to see them after dragging in an image that has them in the flow...eg.

They're from the Searge SDXL nodes

I can't even copy/paste them from one window to another, but I've been able to with other nodes.

If I "Clear" the window, that also causes the nodes to be inaccessible.

[deep pewter]

a dog

[sly sundial]

Nope, it's a Raccoon

[deep pewter]

okay

how can i use this

[sly sundial]

Run it locally or through a cloud service.

https://github.com/comfyanonymous/ComfyUI

[silver nebula]

InstantID

I know! The fingers just aren't good enough!!!

[sick badger]

anyone know how to better clean up the outer edges of a inpainting grow mask?

without chaning the mask quality?

[silver nebula]

I used Kohya and generated larger images than my 8Gb VRAM will normally allow.

[strong flower]

with Cascade you can go 2500x2500 with no problem on a 8GB 🙂

(push the latent compression to like 64)

[silver nebula]

I agree, but the smaller faces do go all mushy when you compress above 42

[merry ermine]

my string nodes are weird

its displayed like an integer and it doesnt expand

[limpid dragon]

horse (╯°□°)╯︵ ┻━┻

[olive yoke]

that looks like a real bug that you should ticket.

[olive yoke]

there's no set of nodes or workflow in comfyui that solves masking in the general case. additionally, all masking techniques involving the stable diffusion family of models are limited, because you cannot "feather" in latent space, so you cannot make a smooth transition from noisy-to-not-noisy in the latent space representation of a 2d mask. can you talk more specifically about what you are trying to do?

[olive yoke]

the hand repair lora from the sliders lora guy, which i think has been repackaged on civitai as "openxl-hand" or something, helps the most with hands

creatively not showing hands is still the best way, followed by that solution, and then maybe graphormer

[sly sundial]

With who though? Is it a bug in the nodes or comfyui?

[olive yoke]

i think comfyui

it won't hurt to ticket in both places though

[olive yoke]

there isn't really a reason to use the efficiency nodes. this appears to be a bug in the node. your workflow is probably missing an input somewhere, like a filename

[unreal reef]

I've been using Automatic1111 since SD was released. I'd like to see what all the comfy hype is about. Is there a good tutorial video out there for people making the switch? The comfy look is daunting.

[trail ice]

Hi first time using MultiAreaConditioning node by Davemane42 and instead of getting what is specified in the different prompts I am getting a sort of blending with what is specified in the background prompt. The only model that works is an illustration model. The other models I try which are more photo real fail. Has anyone encountered this? Is there a correct workflow to fix this? Any help? Thanks.

[next frigate]

I'm having issues with my hires sampler stopping several steps shy of the intended target

[winter estuary]

what i have i done

[olive yoke]

i'm not sure, but you should download it and drag and drop an example image into the website it asks you to use, and you can sort of go from there.

[sick badger]

Great explanation, appreciate that.
First I'm using a blender file via Python to layer user choices together and pull in data from json format and output the transparent pokeball image to a tmp file. After that completes, I'm bringing the transparent pokeball back into a comfyui workflow using the web socket API to add the background image from a user typed description. Everything works as intended all from my webUI but just not super thrilled about the final quality due to the non blending feathered edges as shown.

It's not terrible but could definitely be better once the tech is there. Upping the denoise a bit on the second sampler helps some bit distorts text if too high.

Looks ok if you don't zoom in lol

[olive yoke]

cool let's talk about this, thanks for the detailed explanation

[sick badger]

That's just using the example comfy version of inpainting, no ipadapter of controlnet of any kind

[olive yoke]

first you probably want my fully update to date packaging of comfyui - https://github.com/hiddenswitch/ComfyUI - if you are trying to make this into any kind of application

see the updated examples here - https://github.com/hiddenswitch/ComfyUI/tree/master/script_examples

the README is updated too