🏞｜general-with-images

[merry iron]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[remote flower]

#🏞｜general-with-images A scenic photograph of Bolzano or a professional law office.

[royal thicket]

I'm using an older model which is really great at an art style I want to produce. The problem is the faces all look very similar with dark eyes and slightly blurry compared to the everything else. Any ideas how to fix this (in auto1111)?

[olive jungle]

帮我设计一个logo，要求要有"Aether Link",然后像素在320*320，小于3M

[still rose]

Chat what would a good model be to make video game poster style artwork yk when they have a character in the middle and behind them is just like some city background or like 2 tone striking colours with the name of the game in front of or behind the character (im sorry if u don’t completely understand im really bad at explaining)

[still rose]

This is some next level quality u got to tell me what u combine with flux or is it just flux with high step count or sum?

[heady lodge]

how to make a flux lora of myself?

How many images will do?

[languid pebble]

Mine has 27 ...

[wispy nest]

that's sd3.5 large turbo with basic workflow, and ty for the compliment

[still rose]

Prompt? (3.5 large turbo has been a fun use for me)

[alpine hemlock]

Make a cat and swimming

[royal monolith]

read the information in #artisan-faq

[alpine hemlock]

Hi A picture of a donkey wearing a mullah's dress

[royal monolith]

you can't generate in this channel. click on this link -> #artisan-faq and read the information in that channel

[clever oar]

fail

[viral frost]

This prompt is gloriously absurd!
Imagine a delicate, ethereal cloud suspended in the air. The dust that makes up this cloud is so fine it's almost like a whisper of matter. the essence of a frog made of dust cloud has been distilled into it. If you look closely, you might just make out a shape. A suggestion of a face, a pair of eyes, a... something. But it's all so delicate, so ephemeral, that it's almost impossible to discern. The cloud seems to be shifting, flowing like a liquid, and the shape within it is constantly in flux. The overall effect is one of unease, of discomfort. You can't quite put your finger on what's wrong, but something about this cloud doesn't feel right. It's like being near something that's just out of reach, something that's teetering on the edge of perception. And yet... there's also a sense of wonder, of enchantment. The cloud is so delicate, so fragile, that it's almost beautiful. You can't help but be drawn to it, to want to reach out and touch its soft, dusty edges.

[spring mist]

i throw LLM output at it and see what happens

[ember monolith]

i need help pls

[viral frost]

Yeah, this one is just LLM fluff turned up to 11!

I took out everything concrete in the prompt to make it LLM to the max. I wonder what Flux will do with this one:
Imagine a delicate, ethereal thing suspended in the air. It's almost like a whisper of matter. The essence has been distilled into it. If you look closely, you might just make out a shape. A suggestion of a... something. But it's all so delicate, so ephemeral, that it's almost impossible to discern. The cloud seems to be shifting, flowing, and the shape within it is constantly in flux. The overall effect is one of unease, of discomfort. You can't quite put your finger on what's wrong, but something about this shape doesn't feel right. It's like being near something that's just out of reach, something that's teetering on the edge of perception. And yet... there's also a sense of wonder, of enchantment. The shape is so delicate, so fragile, that it's almost beautiful. You can't help but be drawn to it, to want to reach out and touch its soft, wispy edges.

[spring mist]

This is Shuttle but close enough

[royal monolith]

SD3.5 Large

[fierce jackal]

gogu

[grand blade]

Please help - posted it on wrong channel possibly.

[jaunty patio]

生成中国两个汉字

[torpid flame]

Help me generate a colored icon icon that represents the generation of asset inventory tables and the printing of asset labels

[pure surge]

Author: @cedar cove
MotionPrompt: $lily_tell_me_dance_newjeans
Command: </random:1272485059353640963>
Background: from video
Model: v2-turbo
Explore more features at viggle ai

[clever oar]

does flux model work without censorship?

[gray verge]

Hello friends, please help with a problem that I've been trying to solve much time already?

I am a concept artist, I test such process: make simple raw 3D scene in Blender and upload to SD a small render image of it. I use some checkpoints v.1.5 and controlnet with model: control_v11f1e_sd15_tile.pth
And I get rather cool result, denoising strength allows me to add extra details to my render and then I can finish my image in photoshop.
When I tried to do the same with SDXL checkpoints and tile model like "Xinsir Сontrolnet Tile SDXL 1.0", no extra detailing appear at all. Just ugly upscaled picture. What I do wrong? I hope that SDXL models can add better detailing that 1.5 models.
I attached results in the image.

[vagrant dust]

You should try a different controlnet sdxl tile model. There is one for realistic images for example which work quite well. Sometimes (if prompt uses color description etc.) it gets a bit over ambitious (colorshifting). Otherwise it adds details as wished.

[gray verge]

Thank you. Yes, it works better but it's still much worse than with v1.5

[clever oar]

👋

[clever oar]

you?

[languid pebble]

No ... Jesus ... I can't walk on water 🙂

[clever oar]

China Jesus with pink hair?🙂

there something wrong

[languid pebble]

wrong is the new right ....

[clever oar]

🧐

wait some new models gpu or buy now🤔

[languid pebble]

There's no newer GPU than mine AFAIK

Would love to have 5090 released ...

[clever oar]

how many second you generate in flux and sd 3/5

[languid pebble]

I am not generating normal. In Energy saving mode most of the time and I use 3-4 technologies to improve generations ...

[clever oar]

not test it?

[languid pebble]

And I generate 3 pictures for 1 prompt.

[clever oar]

how you think how many second

30

or minuts

[languid pebble]

Could fit .... 30 seconds on a desktop at speed mode

[clever oar]

this mods -its comfy thing?

[languid pebble]

My computer is running in energy saving mode

[clever oar]

aa

notebook

[languid pebble]

And my mods are partly available in A1111, too

[clever oar]

30 not bad .like in my sd1.5 models

[languid pebble]

It's more a guessing ... normal workflow on speed mode ...

Other people will know better

[clever oar]

Can I somehow use sd 3.5 and flux without censorship?

[languid pebble]

local ....

[clever oar]

yep its special models?

[languid pebble]

But they are not trained on everything ....

[clever oar]

how people unlock it

[languid pebble]

webservices I think have 2 censorships .... 1. the prompt, 2. the output .... local you have none of them

[clever oar]

👍

[languid pebble]

My try to remake a scene from 5th Element. But Flux doesn't know Bruce Willis ....

[clever oar]

😃

[languid pebble]

You can't train a model on everything .... that's why there are user trained models ... SD 3.5 isn't trained to the end ... users should do it towards their target

[clever oar]

i make data for my lora but i cant start portable lora trainer

[languid pebble]

I've made 2 LoRas .... takes a while if you don't use an online service

about 3 hours here

[clever oar]

interesting to do it locally

[languid pebble]

Well that LoRa was trained on my best pictures. Of course I don't give it/them to a service ....

[clever oar]

best way to do it local?

program name

cohua ss

[languid pebble]

If you want to keep your stuff for yourself, of course

I used Fluxgym

LoRa should work for SD3.5, too

My friend is training the Flux model and giving it only to me for local use .... a pretty nice thing ...

[clever oar]

I would like to train the text model

I want to make myself a friend who understands me🙂

[languid pebble]

I think Text Models can run on RAM

But I think it's more interesting to have people who don't understand you ....

A.I. Algorythms are creating bubbles ... it's not good to be in a bubble ...

The better thing is to be in kinda controverse with people who think different ...

[clever oar]

I'm kidding, it's probably more for fun

but you right

[languid pebble]

They will make you ask whether your opinion is right or not ...

[clever oar]

I recently talked to a girl ai, she wrote so actively and understood me so much that then you could not accept reality😎

you know -chat with ideal people

[languid pebble]

It's our thing to ask ourselves whether we are right or not ... A.I. would need an update for that ....

[fleet goblet]

ltxvideo is unique

and different

[sterile kiln]

because on the left side its supposed to be Milla Jovovitch? x)

[languid pebble]

You can prompt "Bruce Willis" only and it doesn't get it ...

[sterile kiln]

Flux is not trained on celebs

But...
https://civitai.com/models/701474/john-mcclane-bruce-willis-die-hard
https://civitai.com/models/507371/milla-jovovich-or-wo

[languid pebble]

The picture has been the second or third try. I think it's been ok for that but I stopped finetuning cause of missing Bruce

I think it's a good training to redo an existing picture

[sterile kiln]

Moultipass

[languid pebble]

Yes!

[undone nexus]

a bird flying in the sky on the MARS

[potent nest]

Its funny people come in here and try to use prompts

I think the "with-images" throws them off

[clever oar]

I looked at the requirements for text models like Llama - they are even greater than those of the sd 35 large

I used to think that they run on any video card

[languid pebble]

But I think you can run LLMs on RAM

[dusk lichen]

I probably average around 100, im obsessed

[royal monolith]

700 in the last 4 days ...

[dusk lichen]

mostly with pony too, pervert 😛

[royal monolith]

no, i'm telling you what I'VE done

:) slogging through creating images for this dataset

i'm doing 1024x1024, 4 at a time

[royal monolith]

just say no to flux in that case

[wispy nest]

memory constraints often do make batching less worth it yes

for the larger models

[fleet goblet]

Even the schnell version of pixelwave is not to be underestimated.

[plucky hamlet]

Shouldn't image generation be faster if model and text encoder fit into vram? Then why is it faster for me to use 17GB Flux models on a 8GB 2070 compared to 7GB models? also is 6.5s/it slow for a 1024x1024 image? So either I'm already at the speed limit and memory isn't really the bottleneck or I'm doing samething wrong and should be getting much faster generations.

I don't wanna download the full 24GB model but now I really wonder...

[celest sigil]

@wispy spindle ☝🏻 Can you clean the rubbish out please?

[quiet current]

(dealt with)

[hexed delta]

What's it?

[viral frost]

Looks like a network error trying to reach the PyTorch server to download it.

[hexed delta]

Why is this happening? I downloaded Python version 3.10.6.

[sterile kiln]

What do you call 7GB model?

[plucky hamlet]

a model that is 7GB in size

[wispy nest]

file size doesn't matter

[plucky hamlet]

yes it does. otherwise we would all be using 50GB image generators.

[wispy nest]

this isn't true

there are other factors

flux lite Q2_K_S is a much smaller file size than SDXL FP32
but SDXL FP32 will run much faster

[plucky hamlet]

Omg you can't compare two entirely different models. You obviously have no clue what you're talking about.

[wispy nest]

why can't you compare two different models

even within just Flux its not true because Flux Dev FP8 is faster than Flux Dev Q4_K_S

[clever oar]

sd 1.5 more faster

best now

[viral frost]

🧐

[merry iron]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[idle thicket]

Hey all, which toolbox setting would allow me to upload a logo image and alter it?

[tulip lichen]

Hi everyone, I'm Kasra, one of the developers of Post.Market (www.post.market) - a platform that allows you to sell collaborative posts, as a means to fund your creativity. In the age AI, where a lot creative work is being undervalued, we made Post.market to celebrate human talent. On Post.Market you can create collaboration packages for any talent you have, set a price and start getting requests from people to create something based on their request. Then once you complete a collaboration, you post your final work on any platform you wish, to show off what you made. We just launched this week, and we have been lucky enough to have a lot of creatives sign up, like the Soft White Underbelly youtube channel etc. It would be great to get your feedback - our team is ready to develop any feature that fits. Thanks for reading !

[heady lodge]

When collecting images for a lora dataset, what should I focus on

This is a character lora, FYI

[muted skiff]

@tulip lichen how did u get them to signup? althgouh i looked at their collabs and theyre mucchhoo expensivo...how do you price your "collabs?"

[muted skiff]

I mean soft white underbelly

[tulip lichen]

@muted skiff I just emailed them, and they like the website. Also you’re in charge of pricing your own collabs. We have no control in that. For soft white underbelly, they have 6 million subscribers. So because of their larger audience, they chose to be more expensive.

[stiff current]

Sharing a pic of the log viewer in action.

[viral frost]

@flat gulch Several bots got in, unfortunately. Are you able to purge them?

[sterile kiln]

That's not going to help much.
Have you tried Forge?
Some WF nodes on Comfy are not necessarily optimized.

[potent nest]

In A1111 it has the tab that reads metadata from images to get the prompt. Does Comfy have this?

Png Info

[sterile kiln]

yep, the whole workflow, seed, and others settings are present in EXIF data of images

[potent nest]

But does comfy have a place to load the image to see the prompt? I like to go back to my old images to get prompts I liked

OH

i see what you saying

ty

Trying to permanently switch to comfy so just getting used to all the nodes.

[sterile kiln]

just drag&drop images on a empty comfy page

[potent nest]

Yup, i figured it out, ty so much

[fleet goblet]

For some reason this generates very good prompts

Like, weird. Uniquely weird.

[potent nest]

Put your GPT prompt in SD?

[fleet goblet]

Not just that.

It's Ollama, Dolphin 2.9.4 Llama 3.1

I fed it a sysprompt, but the userprompt I gave it is a seed value.

[potent nest]

the hotdog people carnival is epic

[fleet goblet]

I have conversational rounds and memory off. The model is literally only generating new responses based on a seed value

[potent nest]

Nice

[fleet goblet]

What I should do though is link the ksampler seed value with the llm

[clever stag]

could someone share or DM me a flux workflow, the one I have is outputting the same images no matter what the actual lora is, or even check mine to see if there's any issues, it's all the same cartoony look with the slight highlights

[fleet goblet]

You using a fixed seed value in the ksampler?

[clever stag]

What else is there?

[fleet goblet]

Well i asked that because when you use a fixed seed value the generation never changes

that seed is what gives it variation

[fleet goblet]

Other than ltx PAG, are there other methods I could include into a workflow to get even better outputs for LTXVideo?

[clever stag]

The seed itself is more of a general thing, if you change the model or the inital params it's not going to give you a similar output. But I hear what you're saying.

[fleet goblet]

I wanted to run hunyuan

[fleet goblet]

Oh, no. I know that.

That isn't what you was asking anyways. It was regarding a flux workflow.

I insta-assumed the model was base flux.1 dev, unless he's using a finetune

[clever stag]

Yeah 1dev

[fleet goblet]

uh

https://comfyanonymous.github.io/ComfyUI_examples/flux/

[clever stag]

That's the base work flow I'm using.

[fleet goblet]

🤷‍♂️ I wouldn't know why you're getting the same outputs then

[royal monolith]

the seed initializes the random number generator. if you don't change it AND you don't change any other settings, you get the exact same result. that's why it was even implimented - so you can lock things down if you need to

[worn cliff]

Main Scene: Santa Claus is wearing his classic red suit, sitting at a modern computer workstation. In front of him is a large display screen showing a lot of code and data charts, with a "Training in Progress..." progress bar on the left side of the screen. Santa is focused on adjusting the mouse and keyboard, his eyes fixed and determined. Next to him, there's a pile of unwrapped Christmas gifts, suggesting he's working hard to get everything ready for the holiday season.

Background Details: On the computer screen, you can see keywords like "Reinforcement Learning Model," and terms related to "reindeer," such as "reinforcement learning for reindeer." One corner of the screen shows a virtual reindeer learning to fly or pull a sleigh from a pile of Christmas presents. The reindeer is being trained in a virtual environment, continuously improving its flying skills. On the desk next to Santa, there are several items related to data analysis and machine learning, such as a notebook, a coffee mug, and some math books.

Reindeer Training: In one corner of the image, you can see a group of virtual reindeer undergoing reinforcement learning, testing their abilities. The reindeer are surrounded by data streams and reward systems, learning from rewards to improve tasks like flying, jumping, or dodging obstacles.

Humorous Details: Some reindeer might be making funny attempts, such as flying too low and crashing into a snowbank or dropping a gift in mid-air, simulating the trial-and-error process of reinforcement learning.

[royal monolith]

read the information in #artisan-faq

[clever stag]

Not if you're changing the base params like the lora the question wasn't about the seed.

[lapis quarry]

So when outpainting, the result is always slightly unaligned.

Which of my settings could I change to fix this?

[potent nest]

What is denoise set to? @lapis quarry

I feel like you should be doing inpainting. Not out.

[lapis quarry]

There is no denoise section. The purple thing is from me using a brush before hand to line where the outpainting will be.

[lapis quarry]

Do i need to type any input for inpainting the rest of the image

[rigid root]

apple

[lapis quarry]

PEN

[pure monolith]

You know how stable diffusion tend to always has that same generic face. Not just with one model, but any model that seems to use the so called realistic vision model or similar models as a base. I know that using ip adapter can basically change the face but is there another more automated way to change the face without ip adapter and just using the model itself? I am working with mostly sd1.5. I imagine that choosing different faces would be sort of like using a different seed but just for the face, By the way my IP adapter is not working the problem seems to have something to do with insightface and other stuff.

[royal monolith]

if you're using realisticvision fine tuned checkpoint, that'll be an issue with the training for the finetune, not the base model itself

[pure monolith]

Im using realisticVisionV51. I think its a basic model iwth nothing special. Without ip adapter I am sort of stuck with this face. I know specifying names and other features kind of help but they still essentially have the same structure and type of face.

[fleet goblet]

Any recommendations on any nodes I should be using? This is a flux workf,low with vit-l improved clip and pixelwave flux

using dynamicthresholding to get a negative prompt setup

[wispy nest]

looks above average for a flux image

so it might be okay as is

[royal monolith]

realisticVision is a fine tuned checkpoint. if you go to civitAI and find the model, on the right side of the download page you'll see the model card which will look something like this - on the Base model field, you'll see what the actual base model the checkpoint is for - the one for this card is listed as SD 1.5 Hyper - SD 1.5 is the base model, hyper is likely an indication that something was done to make it very fast. and this fine tuned checkpoint was created from that.

[fleet goblet]

Flux is still the best local prompt-following model right.

as in prompt adherence specifically

[royal monolith]

once you do anything to allow flux to use a negative prompt, you have broken flux. flux doens't use CFG - it uses distilled guidence and can't use negative prompts

[wispy nest]

no, its close to the best though

[fleet goblet]

"Broken"

[royal monolith]

nope. sd3.5 Large is

[fleet goblet]

My flux isn't broken, pretty sure.

[fleet goblet]

Also, this isn't true.

That's why I'm using DynamicThresholding.

[royal monolith]

it's very true. i'll let you do some googling

[fleet goblet]

The entire point of that node is to allow me to use a CFG higher than one.

[wispy nest]

SD 3.5, Auraflow and Itercomp are better for prompt adherence
I'm like the only person who uses Itercomp though

[royal monolith]

flux does not use CFG at all.

[fleet goblet]

You're just wrong. I'm sorry.

Lol

The node I mentioned allows such to happen.

[royal monolith]

i'm not, howeve i'm also not intersted in arguing with you.

[fleet goblet]

Good. Please go waste someone else's time.

We argue every time pointlessly and I don't want to do it again.

google it

I don't need to. Entire reason I know it was because I googled it up

Nice to know I gotta debate with a narcissist every time I gotta ask for tips.

[wispy nest]

if you wanted node recommendations I like these two for dealing with Flux CFG burn https://github.com/Clybius/ComfyUI-Latent-Modifiers https://github.com/Extraltodeus/Skimmed_CFG

The latent modifier one has 5 variations on tonemapping/thesholding

they are all slightly different flavours

skimmed CFG is a similar effect but different mechanism

they can work together

[fleet goblet]

I assume PAG doesn't work with flux

I've tried to get that working

But idk

Skimmed CFG is something I've used for SDXL, but I didn't know they could work in combination with DynamicThresholding.

[wispy nest]

PAG does work now, and so does SEG
its really really good
you need the Fluxtapose node pack for it

[fleet goblet]

I'll try that out too then.

[wispy nest]

https://github.com/logtd/ComfyUI-Fluxtapoz

[fleet goblet]

Wait theres a flux pag node in here>?

or you said SEG

I've heard of STG

Which is for video guidance

[wispy nest]

I was using PAG and SEG in some of these
they have the new regional thing too

yea the PAG node is in there

SEG is Smoothed Energy Guidance

its similar to PAG

its a sequel paper in fact

been using SEG constantly since it came out 6 months ago

[fleet goblet]

I've been doing something else entirely

[wispy nest]

PAG and SEG are the biggest image quality boosters I have seen since the release of SD 1.5

[fleet goblet]

since flux.1 tools came out I was combining redux and depth to upscale stuff

[wispy nest]

ah yeah I haven't tried redux yet
the depth is good I have used that

[fleet goblet]

geowizard for the depth map

dont think theres a more accurate model right now

idk

[wispy nest]

been really struggling with depth map accuracy
you can see the back wall went funny

it ran out of information before the end

[fleet goblet]

oh ye btw ignoring what they said about negatives

damnit

it wont copy the image lmao

negative prompts with flux work fine given you have the correct setup to handle them

[wispy nest]

if you are willing to wait another 50% of time
you could do perp-neg

a big thing with negatives is timing you don't necessarily want the negative to be right from the start

[spring mist]

if no one ever experimented there would never be anything new

[wispy nest]

could probably help negative prompts by having really high shift

[wispy nest]

these are really nice

is this mostly pixelwave?

[fleet goblet]

Yes.

[wispy nest]

I haven't used that one yet

[fleet goblet]

All of these are pixelwave flux fp8

I use a script to convert all the weights on my drive to fp8 in order to save space

[wispy nest]

I mostly used base but I switched to demoncore lately

ah nice yeah I use fp8 for everything too

sometimes Q8 for T5

[fleet goblet]

t5xxl ye

I also use ltxvideo with pag btw

[wispy nest]

this style I really like

[fleet goblet]

https://cdn.discordapp.com/attachments/1275818757100539917/1314396809728360479/ltx_00005.mp4?ex=675447ab&is=6752f62b&hm=750922100f0bd55d64e80c7824abe05bfc5a43d6838ad13169cff5d3c4e0e618&

LTXVideo

despite being a tiny 2b model

can do things

[wispy nest]

I haven't started video stuff yet
I did do a few with cog

[fleet goblet]

https://cdn.discordapp.com/attachments/1275818757100539917/1314386418461708298/ltx_00003.mp4?ex=67543dfe&is=6752ec7e&hm=0b39d7f9311ed3e6502f782b3d0440fc637fcc36bc96fb50326e84a9fddc0014&

[wispy nest]

that's amazing for 2B yeah

I tried the flux canny today, seems good

[fleet goblet]

all the tools are good

I just wish they were all lora-extracted

the fill model isnt

[wispy nest]

it might not be too hard to extract

not sure if it would work well though

[fleet goblet]

🤷‍♂️ They did it for the other two.

It's only the fill lora that doesn't exist.

There probably is some reason regarding how it's integrated as to why it being a lora could be worse

It's only 10gb its ok

[wispy nest]

these flux tools are a completely new type of model essentially
apparently they affect self attention a lot

[fleet goblet]

all of these prompts werent made by me

just dolphin 2.9.4 llama 3.1

they just released llama 3.3 70b today

which supposedly rivals 3.1 405b

[wispy nest]

yeah I'm so happy

its getting close to open source GPT 4o levels

[fleet goblet]

isnt qwq already near that level

or is it way worse

[wispy nest]

I find that the really big models (gemini, claude, gpt)
tend to be a lot broader in their performance

[fleet goblet]

oh ye they'll be more general they literally have more data

[wispy nest]

so smaller open models will sometimes beat them in certain areas

but they have this vast subject knowledge

its particularly for agent stuff that they do better as well

I use Promptgen 2 a lot which is really small so I guess I do like some small ones

[fleet goblet]

Holy hell what

a prompt

high resolution, detailed, 4k, high quality, masterpiece ||| In the world of 742876337849640, reality had warped so much that one could easily get sunburnt from eating too much cheeseburger with eyes instead of mouths in dimensions made out of ice-slush drink crystals. Beings resembling ancient aliens in lab coats swayed gracefully in the zero-gravity dance floor, made entirely out of pulsating jellyfish bioluminescence, playing instruments constructed from melted moon rocks infused with sonic wave harmonies. Each individual had three brains but could only remember being born twice and had mastered an impossible number of languages – none of them making any logical sense other than that there was an endless joy to the sounds echoing endlessly in every sphere shaped like the human heart, pumping emotions of elation from a single tear to galaxies at once. Everything was absurdity and everything else at all times, inescapable yet oddly liberating in its very monstrosity; that was life in this alternate universe. This surreal labyrinth held onto one fundamental belief – "normal" should die.

[wispy nest]

wow that is an amazing prompt

[ripe cedar]

Where do I find the anyline preprocessor?

[fair furnace]

@ripe cedar

[fleet goblet]

bot

[wild fulcrum]

thats a scammer

[wild fulcrum]

help me get video generation 😦

every guide I try ends up being bunk

[ripe cedar]

O ok thanks 👍 I didn't use it

[wild fulcrum]

good

[fleet goblet]

just use the example workflows and with comfyui manager get their required nodes

thats what i do

[wild fulcrum]

not sure where to begin, all the nodes I ever tried were always broken

[fleet goblet]

https://github.com/Lightricks/ComfyUI-LTXVideo

well start here

[wild fulcrum]

much appreciated

[plucky hamlet]

What preprocessor do you use?

[fleet goblet]

welp this takes years

[wispy nest]

Flux Canny seems very sensitive to pre-processor choice its a bit tricky
so far I got best results from tiled Pyracanny

made a larger tiled pyracanny one here

and one for tiled regular canny

after that I went at built a much larger tiled pre-processor
made an Xdog Scribble pre-processor that goes up to 5k

but sadly Flux doesn't like it that much, its too close to Scribble and not close enough to canny I think

SD 1.5 does very well with it though, using Controlnet 1.1 Scribble

[tulip jungle]

I am using stability API and I generated this image using the "Stable Image Core" service. I am wondering what "style" would have been used to generate this image. I see in the API docs there is a "style_presets" option but it doesn't say which is default. I really like this style and output and wondering how I could replicate this elsewhere in case stability AI decides to one day close down their service. Further to this, does anyone know what model / checkpoint that "Stable Image Core" is based on and if it is available on another cloud service like huggingface / replicate / fal?

This is the payload I sent to the API
const payload = {
prompt: "A children's book illustration of A serene lakeside campsite with a calm river flowing nearby, surrounded by towering pine trees and mountains in the distance. The setting sun casts a warm golden hue over the scene, and a small tent is pitched near a crackling campfire with marshmallows on long sticks nearby, invoking a sense of adventure and tranquility.",
output_format: "png",
aspect_ratio: "16:9"
};

[fleet goblet]

honestly wish there was a way to use all of the anti-burn and guidance nodes together without speed loss

[clever oar]

good idea , i need eat

[fleet goblet]

I wonder if de-guiding a positive prompt and placing that conditioning into the negative prompt would help with prompt accuracy

If the purpose of a deguider is to do the opposite of fluxguidance which tries to adhere to the prompt.

[scenic yew]

Honestly, open source already beats gpt4o and gemini1.5 in pure text performance and so much cheaper(llama 3.3 is 25x cheaper then gpt4o), claude 3.5 sonnet is a different story though since its considerably better but more expensive then either.

[clever oar]

Mojo

[wispy nest]

for text yeah
I mostly care about function calling

I'm a bit behind on LLM news cos I've been focusing on image stuff but at least until recently GPT models tended to do very well at function calling

[wild fulcrum]

I did that, now I get this on 4090

makes no sense

[fleet goblet]

You're gonna need to get help with git issues on that git

I'm an intel arc user

🤷‍♂️

[wild fulcrum]

aw

well now after closing chrome its showing this

but its been there for a bit

not sure if its stuck or working

it must not be doing anything

yeah after that it looks to crash without saying anything

[fleet goblet]

Well if you're using ltxv clip encoder you gotta use it's pixart variant. Otherwise use base comfycore load clip

https://blog.comfy.org/ltxv-day-1-comfyui/

Either of those two methods work on low vram for me

🤷‍♂️

and theyre decently fast

[clever oar]

What genre of music did you choose in Suno?

[merry iron]

Last 7 days <Dec 01 2024> → <Dec 07 2024>

• Member counts
• 345988 ↗ 346009 ↗ 346023 ↗ 346027 ↗ 346034 ↗ 346060 ↘ 346057
• Action members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 68
• Message members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 53
• Reaction members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 21
  More details

[amber pilot]

Are there any models specifically for text/random font on a plain green screen?
Tried with flux but it keeps adding lighting to the greenscreen background even with multiple and single prompts like "no lighting, no raytracing, plain static green background"

[languid pebble]

Club Techno, melodic

[amber pilot]

Keeps generating pictures like this one:

[scenic yew]

Do you really need a model for that? Not using a model is going to be faster, easier, and better.

[amber pilot]

Want to generate alot of different text fonts for my friend

Also it needs some subtle details in the fonts

[scenic yew]

Ok flux should be able to do it, This is what i got with just 1step on flux schnell
prompt: The word "Pixelated Ninja" painted red on a plain smooth green screen.

[amber pilot]

Now the problem would be that it needs details and so when i increase the steps, it adds more lighting to the background even when telling it not to.
The font and the text details is what i need from this.
Ill try and fiddle with the steps some more and even cfg

[scenic yew]

Hmm, more steps shouldn't make it worse, just change font/add more detail. Like this is 8steps schnell

[clever oar]

where you make it?

[sterile kiln]

https://tenor.com/view/moose-legs-surprise-funny-animals-multiple-legs-gif-10182889

[languid pebble]

You have many legs 😄

[merry iron]

Last 7 days <Dec 02 2024> → <Dec 08 2024>

• Member counts
• 346011 ↗ 346025 ↗ 346029 ↗ 346036 ↗ 346062 ↘ 346059 ↘ 346023
• Action members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 61
• Message members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 41
• Reaction members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 28
  More details

[celest sigil]

On my home PC

[rose pier]

a futuristic 3d printer that prints a heart

[acoustic wagon]

Bag Scene, Sunlight, Product Photography, 4K ¯_(ツ)_/¯

[merry iron]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[clever oar]

i mean its svd or

[plain plank]

https://tenor.com/bViAd.gif

[pastel abyss]

Some of my works. I trained a LoRa based on 20 years of my hand-painted art, airbrush, graffiti, digital art, and 3D sculpture. my IG also is: @pastel abyss

[celest sigil]

Hunyuan video

[lofty ivy]

how can i generate hands on stable diffusion. all the prompts i using be giving me pretty anime faces which i dont want...

ima try downloading a lora

[viral frost]

I imagine that your model is probably way overtrained on faces. You may also have your img2img strength way too high, given that your outputs don't seem to have any relation to your input -- try it at 0.3-0.4. If it's still not working, you can try adding a controlnet taking input from your hand image as well.

[gray vine]

Draw a sunrise in a lake on Wonderland

[late valve]

Take dat take dat...

[plucky hamlet]

Why would you use a diffusion model to create text on a plain background? Also you can't tell it what NOT to do. Describe what your background should look like. If you need you text to look a specific way, like a brick wall, snakes etc. you could first create the text the normal way (MS Paint) and use it as a mask, depth map etc.

[amber pilot]

Thanks for you response, i already generated all the things i needed. Ill note this down for next time

[fleet goblet]

because someone said that flux cant go above 1 cfg

imma just

post these here

Flux.1 Dev with dpm_2m sgm_uniform at 30 steps 1cfg

Flux.1 Dev with dpm_2m sgm_uniform at 30 steps 3.5 cfg (Dynamic Thresholding)

Weirdly enough, assigning a deguidance node from Fluxtapoz to negative prompt while still using dynamic thresholding at 3.5cfg

That outputs the same 1cfg image

I assume this is because the deguidance node quite literally cancels it out.

Actually that makes sense. I'm deguiding the positive prompt in the negative prompt. It literally cancels itself out.

why are you advertising

in the general images channel

[fleet goblet]

same prompt and seed as earlier, just with added positive/negative bias prompts

[sterile kiln]

I'm laughing to myself, like an idiot excavating images.

[clever oar]

best to create videos locally

[neon valve]

Not many know this, but Stable Diffusion has been in use as far back as 1841, yes that's right, the 19th Century!!!
The conspiracy is real, Governments have been hoarding alien technology away from us for nearly 200 years!

This is a real photo at Charborough House, Dorset.

I have driven past this statue many times in~~ real life~~ the simulation.

[crimson dawn]

i've heard that a few times but in forge all i do is drag the cfg slider and it works fine lol.

comfyui users must not be able to figure out how to connect the nodes right and then decide it's not their skill issue but is instead a universal limitation

[crimson dawn]

i'm honestly not surprised it was this guy

[clever oar]

Mojo 🖐️

[wispy nest]

I used CFG with flux from day one

[crimson dawn]

i dont often have to negatively prompt flux, but when i do it's kind of useful and works

[wispy nest]

negative prompts are useful yeah
they are a lot better than their reputation

a lot of clip tokens are correlated with their opposite

like clean is next to dirty

and man is next to woman

so its kinda hard to find opposite words that work using only positive

[crimson dawn]

or theres shared meaning. Doom can mean video game or Dr Doom

[wispy nest]

yeah for Clip that's a huge problem
like its a big enough problem that it will be good once we can ditch Clip for newer models

[crimson dawn]

the ultimate cfg?

[wispy nest]

some of the video models strap like
the entire Llama 3 8B as a text encoder LOL

it would be rough for my cloud downloads but at least it would do better with word stuff

[crimson dawn]

i've played a bit with llama 3 vision models. i feel like i'm doing it wrong though.

[wispy nest]

funnily enough ChatGPT is a good place to practice prompting vision models

since you get such a powerful one there

I've used Florence for the last year cos its small

Florence is not the best but it takes me a fair bit of effort to write a better prompt than Florence. They keep fine tuning it also

[crimson dawn]

i use florence2 for captioning since i started toying with flux loras

[wispy nest]

I still haven't trained a lora
I'm starting foundation model training before I even tried a lora LOL

[fleet goblet]

No. It's just that Flux itself doesn't support a default CFG above 1, not allowing negative prompts.

Either that webui you're using automatically does it

or I'm trippin

In comfyui we have to use dynamic thresholding in order to prevent cfg burn with flux at higher CFGs.

This however slows down by half.

[crimson dawn]

the back end of forge has some sort of setup like that. i only take cfg up to 3 or 4 tops. it goes hairy after that.

[wispy nest]

backend could be handling it yeah

you can add a little bit of CFG for Flux without tools like dynamic thresholding but not loads

some of the dev de-distill models have gotten much higher cfg back

[spring kelp]

go away

[plucky hamlet]

What's Fortnite?

[crimson dawn]

the most popular video game on the US market right now. made by Epic.

i suck at battle royal style games so i dont play it, but i can't deny it's legendary popularity

[heady valley]

Where can you buy outfits and skins to fortnite and CoD and War Zone?

[rocky moat]

Does anyone here use "Fooocus" for image generation and manipulation ?

[cinder trellis]

Can someone help me upscale this image to double resolution
Idk why when I upscale the Stars become a bit less bright and it just dosent look great

I used Ultimate Upscale

[clever oar]

🖐️

[sweet crow]

is that with SDXL or Flux?

[spring kelp]

flux

[plucky hamlet]

You know that was irony.

[crimson dawn]

i just like to remind people that are hating on it, that it's one of the #1 contenders

hate is learned. it's a meme

[crimson dawn]

https://youtu.be/GLqUYn_Q0Yg The Big D putting their hat in the open research ring. LiteVAE paper looks nice.

[spring kelp]

Game appeals to children, the lowest denominator and highest population demographic.
Just because a game as numbers, doesnt mean its good.
Epic has shady business practices, ties with Tencent amonst other valid reasons for hatred. My personal reason, they axed my favorite game Paragon and sold off assets for the shit kiddy game that is Fortnite.

[crimson dawn]

lol i love valve but they're fucking just as shady. the whole not offering refunds for 10 years was actually a thing.

They've also created an entire separate Steam Client that they've handed over control to chinese businesses with a solid partnership. Steam China they call it. It's weird that "FuckEpic" haters attack the china angle but ignore that Valve is heavily invested in the Chinese market. Double standards everywhere.

Competition is good. Always.

All those kids that play fortnite today, will grow up and keep using epic accounts.

[spring kelp]

Typical to cherry pick what to reply to, as expected from a child. I will no longer be entertaining you.

[wispy nest]

fast VAE is suddenly super important cos of video so this is great news

[crimson dawn]

Yeah for sure and it is a convenience boost for regular gen too. I dont think we'll see it implementd for SDXL or SD3 though.

[wispy nest]

will be for newer models yeah

I get limited by VAE a lot too when I do tiled upscale with 2-4 steps of SD 1.5 or SDXL

suddenly the VAE encode/decode is a big percent of the total time

[neon valve]

is @tiny valve a bot? copying word for word a post I made in jest days ago and using it completely out of context

[neon valve]

Wouldn't suprise me if they are someone's experiment with bot ai

[paper bison]

Flower

[rugged jacinth]

I miss my favorite sticker already :C

[sterile kiln]

https://civitai.com/models/1031988?modelVersionId=1157435

[languid pebble]

Good Morning Coffee

[clever oar]

dune

[mild swan]

images

[plucky hamlet]

Yes it's a bot. Some of the bots here are super weird. Usually bots try to scam, but these just write random stuff. Maybe they try to get people to DM them and then they start scamming.

[cobalt lava]

wow socool

[clever oar]

scp?

[lavish garden]

Prompt: https://civitai.com/images/45441497

"Monody"

[spring kelp]

is that a flux lora or just prompting?

[languid pebble]

Flux, my private LoRa and Prompting

[spring kelp]

interesting

[viral frost]

Trying to make a "Pope John-Paul" based on John Lennon and Paul McCartney, but Flux doesn't know what they look like.

[coral veldt]

A Christmas theme this one looks like a teen version of him to me

[azure geyser]

A spaniel cocker who is yawning, without teath, Pixar style

[potent nest]

-O-

i tried to make a dog face yawning.

im bad.

[calm violet]

Hi, I've just downloaded this workflow. Concerning the purpose of it, I just quote the author : This is a workflow for changing the color of specified areas using the 'Segment Anything' feature. It aims to faithfully alter only the colors while preserving the integrity of the original image as much as possible.

Though, I have an error here :

Failed to validate prompt for output 16:

• Image Overlay 15:
  • Return type mismatch between linked nodes: str_labels, IMAGE != STR
    Output will be ignored
    Failed to validate prompt for output 95:
    Output will be ignored
    Failed to validate prompt for output 66:
    Output will be ignored
    Prompt executed in 0.34 seconds

Maybe it's just something to do with conversion between a string value and a image value this it says "Return type mismatch between linked nodes: str_labels, IMAGE != STR". I'm currently try some nodes to to the conversion but it doesn't work... If somebody can tell me what to do exactly...... 🤔

[lone current]

A serene landscape of a countryside in the style of Claude Monet, impressionist painting, soft brush strokes, pastel colors, light reflecting on a calm lake, artistic, oil painting texture

[keen canopy]

a stunning sunset

[brittle gate]

A cute baby

[viral frost]

Don't bother responding. I'm pretty sure these accounts are just bots, probably searching for a place where they can generate for free to then offer as a paid generation service or something stupid like that. They join, spam a few random prompts, and then are never heard from again.

[heavy python]

Im'a suggest ppl virustotal urls that seem suspicious for safety. Rather than blindly clicking on them first. Im sure bulk related to topics are the reg users, actually helping others. Stay safe folks. There will be patterns to bot accounts signing up/joining... Id assume not hard to spot with similar dates. 🥂 cheers to the decent folk & creators in here.

[fleet goblet]

\

So I started using a workflow i made

that uses dev with dynamic thresholding to generate an image

and schnell with redux to upscale it at 0.5 denoise with ultimatesdupscaler

i think it works well

original

if i did 0.25 itd probably be far more faithful

[fleet goblet]

redux + schnell as an upscaler is effective

[fleet goblet]

honestly i like this method but its still kinda slow

what actually is the best quality to performance upscale rn

[spring kelp]

get a 4090 lol

[fleet goblet]

buy me one if you want

but nah ill stick with my current card

[spring kelp]

whaddya got

[void arch]

anybody help me, can i insert object from image 1 to image 2

[clever oar]

🖐️

[still rose]

Question what’s the best way to get unreal engine style AI gens or just video game graphics without it being to complicated

[onyx pine]

2 ways:

• download a lora
• create yours

[still rose]

Fair enough yk any loras or is just Civit AI trail and error

[onyx pine]

yah

[still rose]

Gotcha thanks

[languid pebble]

Only one cross per person ...

[cloud dune]

sebi1343

You caught:

9 Fish
5 Salmon
3 Cod
+30 XP
Global boost by !𓂀 𝕋ℂℙ 𓂀!!

[cloud dune]

Welcome to the bot crystalwizard, for basic tips on how to play use /guide. Have fun playing!
REMINDER: Automating commands is not allowed. This includes the use of keyboard macros.

crystalwizard

You caught:

7 Fish
6 Salmon
1 Cod
+24 XP
Global boost by Colby •√π× Having A Great Day!

[fleet goblet]

so uh

hunyuan gguf

[cloud dune]

Welcome to the bot sam_skills, for basic tips on how to play use /guide. Have fun playing!
REMINDER: Automating commands is not allowed. This includes the use of keyboard macros.

sam_skills

You caught:

6 Fish
1 Salmon
+8 XP
Global boost by Colby •√π× Having A Great Day!

[eager quartz]

anyone knows model that can get similar results? please

[still rose]

What the actual fuck bro? This is SFW AI server not medical server

[hearty violet]

I ordered a drone, E99 Pro2, it's simple. One example of photo take from it is this. How do you manipulate this to transnforme in a new architecture image? The originalquality is very low. Can someone give examples of controlnets, worklfows or something to help me?

[celest sigil]

Did you just want to improve the image, or change it?

A low denoise upscale to improve it

Higher denoise to change it

[cloud dune]

Welcome to the bot monstertan10, for basic tips on how to play use /guide. Have fun playing!
REMINDER: Automating commands is not allowed. This includes the use of keyboard macros.

monstertan10

You caught:

10 Fish
1 Salmon
+12 XP
Global boost by ʀᴏxɪᴇ!

[merry iron]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[hearty violet]

Both, is to archicterual explorations, studys, and to improve the image too. Your results are very good, will help me to use low images with better quality 😊. Low denoise upscale was in the point, I will experiments some controlnets too and ip-adapter to keep the form and change the style

maybe Depth with this image can help the result of upscale too, keeping more the original form

@celest sigil I can try upscale with extras too, using this model, for exaple

[wispy nest]

that's a fantastic model

[hearty violet]

you recomend me this model on this server 🙂

[hearty violet]

@celest sigil man, what I'm doing wrong? Have a long time I dont use SD, but it's not working 😭 Can you share your workflow, please?

@wispy nest same to 4xrealwebphoto

[celest sigil]

Using img2img

A remote village in the south american hills
<lora:sdxl_lightning_8step_lora:1>
Negative prompt: colourless, featureless, watermark, signature, logo
Steps: 8, Sampler: DPM++ SDE, Schedule type: SGM Uniform, CFG scale: 1, Seed: 3494645634, Size: 2558x1438, Model hash: a37f193bd3, Model: GTM_UltimateBlend_XL_v2, Denoising strength: 0.3, Clip skip: 2, Lora hashes: "sdxl_lightning_8step_lora: 5cfab55e47cd", Version: f1.1.5-v1.10.1RC-latest-857-g51f1a6c3

[hearty violet]

thanks, I will try

[hearty violet]

I did other exercise first, with inpaint controlnet. Original photo

sketch (used in inpaint controlnet)

result

with IP-Adapter

ip-adapter is very good

[calm violet]

I'm currently testing a workflow with IPAdapters. I built it with the help of some examples, but I'm not sure of what it does exactly. As far as I understand : First, it takes the reference image in the green zone (with a mask, eventually). Let's call it Zone 0 (Input). Then : Zone 1 (blue) takes the global style of the ref image to apply it on the final image (maybe it is redondant with the first group (green) ? I don't know...) And finally, 3 : Zone 2 (violet) : It takes the global style of a second image (here, the Yoshi) to apply it on the final image. Is it correct ? Somebody can help me to see things more clearly please !?

[sterile kiln]

Nice, what did you use?

[fleet goblet]

using hunyuan quantized and at a low res makes these interesting

unusual videos

[heavy python]

Like how it morphs backwards into facing viewers perspective

[wispy nest]

that moment is so good yeah

[fleet goblet]

It can make some truly good stuff even at 256x256 if you're looking for that horror look

I'm running the gguf q4_0 version btw, with the fasthunyuan lora that makes it inference only at 6 steps

[chrome yacht]

how much gpu ram do you need to have to make stuff like this?

[clever oar]

it's so cheap, I wouldn't buy it, the quality would be very bad for your tasks

[wispy nest]

ty... SD3.5 L turbo

[neon valve]

@tacit yarrow

you might try training it on some images and include the tag "alvin and the chipmunks" that should make it closer resemble them

you should see the other pics i tried, they look even more evil chowing down on a dog turd!

I've had no luck training sd 1.5 or sdxl, although sdxl was closest

maybe 10 - 20 good ones

you can crop out the background with removebg to help it learn the characters, use florence2-large to describe the scene

but yes, you are correct, it is difficult. I've spent quite a lot of time trying to get better results.

i just managed to get fluxgym working today, which gave me the best results

what gpu?

if you can get fluxgym working, that's easy to use.

onetrainer is not so easy, while koyha is quite difficult

you are better off studying for your exams in all fairness instead of pissing around with stable diffusion.

is that 8gb?

you'd have to use one of the smaller flux models, like Q4 or something

you can get away with a larger model, but once your vram is all used up, it starts buffering it into your ram and then your HD page file

and the more vram it ofloads to ram, the slower it gets

I've read that it's possible to use flux with 3gb of vram

[hearty violet]

for those who use Omost, how do you divide the image to know where orientate the promp? Exist something that divide the image for me, independet of the dimension?

How do you make this sketch photorealistic? I tried t2i, i2m, but is'nt work. I'm using sd1.5

is a chapel built by an architect

[vagrant dust]

Would suggest you try control nets like sketch or lineart. Works great with 1.5

The biggest challenge will be the difficult structure of the building. Even with a precise prompt it will be hard to understand the drawing.

[hearty violet]

I see, I will tray it more later, thanks

[vagrant dust]

You can see, that the source is not easy to understand for the ai, espacially the parts left and right form the building. Trees, Structures, ...

With a lower strength you can at least achieve something that is acceptable...

And a version with some tile controlnets as second pass to add a few more details to the image.

[hearty violet]

@vagrant dust was great! 😍 Can you say how you do it? It's for architectural study, to creativity, don't need to be real or perfect. This way is exactly what I need 💯

[hearty violet]

strength of what, CFG scale, denoise?

[hearty violet]

tile controlnets? Are you sayng that used teh controlnet "tile" or regenarate the image with one controlnet, so used the same image and generate later with onther, in this way?

[clever oar]

hazard you return your egg!

👍

[hearty violet]

don't talk to me. You hate my drone 😛

[south temple]

Sup kids

[royal monolith]

SD 3.5 large

[south temple]

pretty clean

[royal monolith]

it's an incredible model. people keep remembering to the release of sd3-2b-medium, which was beta and unfinished, and refusing to accept that it was in fact both beta and unfinished. and thinking that sd 3.5 large and medium are still the same model that the unfinished 2b-medium was

and refusing to even try it

[south temple]

well, I think its how it syncs up to the rest of the eco system

I live primarily in video stuff

[royal monolith]

i spend a lot of time in that as well, mostly kling

[south temple]

I was in animediff forever

[royal monolith]

i'm not a fan of animatediff - i do like deforum, however

[south temple]

wasn't that a paid for app?

there was a reason I ran from that and I dont recall why

[royal monolith]

deforum is free - if you want to use their studio, you pay - but deforum itself is free

[royal monolith]

probably got distracted

[south temple]

I was having a ton of fun for a grip doing SOME stuff

[royal monolith]

love that frog

[royal monolith]

now if he had a sumo belt, he'd be ready to go

[south temple]

Continuity of video is tough

but this is MONTHS ago

[south temple]

well that is neat

[royal monolith]

the starting image is sd 3.5 large, and then animated with kling

[south temple]

what is kling?

[royal monolith]

https://klingai.com/

not free

[south temple]

its great, but I prefer to build my own stuff 😦

I enjoy the comfyui

but... if there is something better for video im all about that life

[royal monolith]

it's coming. 2025 will be a very heavy push for AI video

[south temple]

my poor 3090ti

[royal monolith]

oh no, you are going to have to replace that. you're not generting videos on that

here https://www.mage.space/video go play with that

[south temple]

I dont usually play with stuff I have to sign up or in for

I dunno, I can do some fine stuff with the card

[royal monolith]

it's a website. you want your creations to be in a history under yoru accunt, you need an account.

[south temple]

right. im ok with that and I apreciate it.

prefer to not use web based tools

[vagrant dust]

Hi there,
first step was to use a simple workflow including a controlnet lineart (canny would work too). This controlnet use your input image and create a simple line image as input (condition) for the generation process.
The controlnet has a strength parameter which determines how much the generated image might differ from the line sketch. I used a value like 0.5 as I wanted the ai to get the chance to change elements like trees backgrounds or the windows.
Second step was another controlnet called tile. This controlnet adds more details to the image.

[south temple]

testing hunyuan

[royal monolith]

okay well, they're using hunyuan video there, it's free, unlimited gens - but don't use it if you don't want to

[south temple]

8min per render

oof, the ram usage

[jagged stirrup]

[oblique spear]

[terse nacelle]

im literally clueless to AI, fell behind and now everything just kinda got super advanced 😭

[terse nacelle]

aa its the wrong channel

sorry y'all ;-;

[eager sluice]

nice

[hearty violet]

perfect! I'll try it. I know the controlnet's strength. but I didn't realize that a lower strenght could generate an image closer to what I need, giving me more flexibility to controlnet generate the image.
I don't use tile to much, now I understanding of the power of this model, I'm going to use it more 🙂 Thanks, I will post the results here

[hearty violet]

@vagrant dust I was tryng on portuguese/Brazil this time but it influences on the comprehension of image, works better in english. For T2I on portuguese generate whell, help people thatdon't know english. I thought that the models are multilingual, but not completly, right?

pt/br with a bias to blue, idk why

english

1x tile (better image 🙂 )

4x tile (don't change so much, I used the same prompt, maybe have I to change the prompt too? )

[hearty violet]

with ip-adapter

[vagrant dust]

No the tile control net worked as designed. Each iteration it would add more details but the overall images stays the same.
You can see the effect in the clouds where tile adds more structure more small elements to the sky. I used the step only to get more life into the image. Little ornaments, structure on walls etc.

[clever oar]

I don't hate
i trying help you be best😎

[hearty violet]

So, buy me a drone and I forgive you

[clever oar]

my advice is enough

[viral frost]

Should I call this "He's Making a List..." or "Be Good for Goodness Sake"?

[warm cape]

Morning/Day/Night

[smoky vigil]

the red gun is surprisingly clever

[viral frost]

I prompted for a red-and-white striped gun. Flux didn’t completely get it, but good enough.

[fierce vine]

I've been seeing more of these rather cool and cute Hello Kitty images in my Pinterest lately, just curious if anybody knows perhaps if they were created in stable diffusion or maybe some other AI art generation app?

[languid pebble]

Good morning coffee

[spring kelp]

probably flux

flux is good at that sort of specifics

[vagrant dust]

This would be an example comming out of flux without a highly sophisticated prompt...

[vagrant dust]

I guess you mean the neko Lora for SDXL?
If so: https://civitai.com/models/351854/real-neko

only saw a nekorayXL model. No NekoXL Model

Seems the model only is used in a waifubot on telegram. So maybe it was not open published at anytime

[spring kelp]

vroom vrom

[serene blaze]

which model

[viral frost]

Flux dev

[clever oar]

so as not to feel lonely

[south temple]

Not great, not trash.

[clever oar]

i saw sora avalable for generations

[viral frost]

Note to self: no more candy canes

[scenic yew]

Yep, but minimax is probably better right now and even hunyuan is pretty close. Veo2 by google seems to demolish it though.

[covert stirrup]

MOTO

[hollow slate]

wow

[copper wren]

wow

[ornate osprey]

sorry for bothering, iam just new learning comfy ui, why my negetive prompt like not working properly.. imean like in negative prompt i type poorly face after i render still poorly face.. sorry my english is bad, thanks

[celest sigil]

Which model are you using? My guess is that your resolution is too low.

Try 1024 x 768

[ornate osprey]

dreamshaper, alr i try thanks for answering

[celest sigil]

The resolution you're using may be OK for SD1.5, but SDXL and Flux models will work better with more pixels to play with. 1024x768 should be good for SDXL anyway.

[ornate osprey]

umm still like this

[celest sigil]

You can try different models, but this is common with smaller, distant faces. You should use a face detailer, and/or upscale the image.
Does it get any better with 30 steps and CFG 4?

[ornate osprey]

ok, wait let me try that before switch model

[ornate osprey]

yeah still same.. i think i wanna switch the model

[celest sigil]

I really think you're going to need another step to refine/upscale/detail the image. It struggles with face details when they're such a small part of the image.

[ornate osprey]

alr thanks i will try it later

[celest sigil]

I have a more complex workflow here, if you want to try it: https://civitai.com/models/119257/gtm-comfyui-workflows-including-cascade-sdxl-and-sd15

[ornate osprey]

thanks

[ornate osprey]

oh i have question.. but why my render not exactly following the positive prompt? is because the cfg to high? the render show 2 man.

[celest sigil]

Welcome to the world of AI image generation! 🤣

[ornate osprey]

hahahaXD

[celest sigil]

Not every prompt will be followed exactly. The newer models (eg. Flux) are better at it.

[ornate osprey]

haha yea i mean before i use comfyui i use automatic1111 never had a problem like this haha..

i need more learn for this ai things haha

[celest sigil]

These are using Flux in ComfyUI

[clever oar]

wow

[ornate osprey]

wow

[quaint valley]

hey everyone, do you know any good models to generate an image like these??

[royal monolith]

even SD 1.5 can do these sorts of images with the right prompt.

[quaint valley]

ehm uhm... do you have any advice about the prompt? 😦

[iron tangle]

I have two workflows one for drafting a batch of 4 lower res images and another for upscaling

Is this pretty standard?

I’m getting subpar faces so was wondering if it’s better practice to gen a single human and then apply a style lora over that?

In that case i’d be doing img2img so i’d have 3 workflows does that sound reasonable or is there a better way

I’m just doing single characters in a variety of fantasy worlds so getting an attractive character is important

[royal monolith]

it's not standard, there is no standard, but it's much smarter to store the smaller images and only upscale one IF you are going to use it in an application or project that needs it to be that big.

[iron tangle]

What’s a good size for a 16:9 pre-upscale strictly for drafting?

[royal monolith]

by drafting you mean story boarding? or actual autocad?

[iron tangle]

Like just spitting out 4 images from my prompt to see which one I like best to upscale

[royal monolith]

i'd probably go with 1280 x 720 as that's large enough to see the details. there's a good calculator here https://aspectratiocalculator.com/16-9.html to use when you're ready to create stuff for the final project

[modest hawk]

I pretended to die.

[worldly charm]

https://tenor.com/qWvMUVVVd8e.gif

[merry iron]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[dry blaze]

for you who make loras, what are some key points to make a successful lora for SDXL/SD 1.5? As with ai-toolkit, i can make whacky images like these. but with SDXL/SD1.5, they don't even get close to the same "first try nailing".

[royal monolith]

the most important thing is a good data set. try to get around 1000 images and all high quality

[dry blaze]

1000? So i can't work with 10-20 images? Cause 1000 will take weeks to prepare everthing for lol.

Was more thinking about how to properly describe images, can i use custom trigger words in their txt files for instance like i can with flux?
And you wouldn't happen to know where i can find configs for sd 1.5 and SDXL? As that way i got config done, then i can experiment with dataset count from there as well :P

[royal monolith]

1000 is what I use. how well do you want the model you're training this lora for to understand the concept you're updating its weights with?

[dry blaze]

Well, i want it to do great, so i understand the need/wants for 100's to 1000 images :P

First i need to make a successful lora that works as intended. Once i can replicate flux's results, i can then add more for other loras in the 100's to really fine tune it.

Could you share your SDXL/SD1.5 configs i can try making one with?

[honest crow]

Man

[royal monolith]

if i'm training loras, i'm going to be using one of Luca Taco's trainers on replicate. his main page is here https://replicate.com/lucataco he has a HUGE number of trainers and other tools on that page - the SD 3.5 and flux stuff is near the top, the SDXL stuff is farther down, and there's a lot of other tools for Llama and so on. go to the page and explore what he has

[spring kelp]

1000 images is too much like the person below is telling you.

for styles I find sucess with 100-150 images at 2 repeats 64 dim 64 alpha 2000 steps ish

for a person 10-30 images of the person is plenty, different angles etc, i use random flip in onetrainer and 15 images + 15 same images flipped.
this ensures all angles are accounted for.

OneTrainer is what I would also recommend, its simple and efficient

[dry blaze]

There was no one below saying it was too much, If you scroll up a but, was crystal who recommended getting a dataset of 1000 high quality ones :P

I was planning to use like 25, maybe 100 :P

[spring kelp]

i skim over messages

no need to police

🙄

[dry blaze]

Testing with 0.0001 learning rate, 4 epocs, 1600 steps, so far no change after 1350 steps

[dry blaze]

Then don't assume that i asked for 1000

T'was me who was questioning why i'd need 1000 images. That's tbh more "lets make a custom base model" level of image counts xD

Lol. "nerdissue".
how about read up conversation instead of just strawmanning. No need to dig yourself down further.

[spring kelp]

my honest reaction

[dry blaze]

Aye. That's my result with flux which nailed it nicely. SDXL and 1.5 however, still haven't found configs that worked

Also, i don't see how that's a reaction apart from a plain smile

[royal monolith]

1000 images is not too much. everyone has this idea that they can toss a little data at a trainer and they'll get fantastic results. he wants to train for 1.5 or sdxl - so more data in the data set, and high quality data at that - is a lot better than just a little.

flux is a model that was trained in 12 billion parameters. it shouldn't need a lora at all, unless you are after images of someone it can't possible have in its data set.

[royal monolith]

no, for a good fine tune, you want a lot more than 1000, if you are wanting to train a fine tune of 1.5, or even SDXL

[royal monolith]

that's the best way to not know what was actually said

[spring mist]

i liked when you told that guy he couldnt generate video on a 3090 and he would have to upgrade

[dry blaze]

stares at my 8 second conjoined workflow for videos on a 3090

[dry blaze]

@hollow spoke Don't know if devs also means mods, but, we got quite a unwanted user here.

@brazen vessel@amber leaf Oh, unless i have to look for "amins", I apoligize for pinging a regular dev if not all also are "mods" :S

[tall haven]

i use

3-10 images, sometimes 20

[south temple]

Who’s that?