🍥｜anime

[keen hollow]

Mornin' 🙂

[keen hollow]

Mornin' again

[gleaming flame]

back to bunny girls

[dusky linden]

What model did you use? Plantmilk?

[gleaming flame]

https://civitai.com/models/1331249/bubbli-cartoon-il

[dusky linden]

Looks great, downloading

[near elbow]

happy neko

[gleaming flame]

Dee Dee

[gleaming flame]

...and that kid

[lyric zealot]

no

[gleaming flame]

making generic stuff again... need to take a break

[near elbow]

a rose is a rose is a ... still SFW image thanks to the cropping 🙂

[keen hollow]

This one is cool.

Seems like a good framerate too

[wooden sail]

16 fps - probably best for animation. Not to high and not too small

[keen hollow]

online or local?

[wooden sail]

@keen hollow
Local Wan 2.2
Sometimes it follows complex outfit details surprisingly well

[keen hollow]

Right on. Thought they were still working on good wan 2.2 workflows.

guess it's been a while 😛

[keen hollow]

90 series?

How long is it taking to run?

Built-in comfy workflow? Or custom?

[wooden sail]

Size dependent. 0.6 megapixels takes about 3 minutes. Standard nodes but with initial resizing. Also this model

[keen hollow]

Intriguing

The models names ... "teasingkiss edition" 😛

"cheekydream" -- good grief

[keen hollow]

ok gonna try to get it up and running

feels like I need a degree in WANology

[keen hollow]

ok tried to do your 'transformation' style

i think this one did ok

[near elbow]

annoyed

[near elbow]

grumpy sketchy cat

even grumpier

[near elbow]

a bit too curvy, but still SFW-ish...

and it's halloween time!

wholesome diversion

[near elbow]

that time of the year again

[keen hollow]

Zombie with drippy bloody finger

[keen hollow]

love that one

[wooden sail]

I think right now the most useful thing is not animation itself but models ability to generate consistent frames. Character posing in other words

[gusty palm]

those are neat

[wooden sail]

"You've dropped it?"

[near elbow]

nice... sweater?

[near elbow]

interesting color combo

[tawny wasp]

Shampoo From (Ranma_1/2) illustrious Apex_Ai
https://civitai.com/models/2052215/shampoo-from-ranma12-illustrious-apexai

[near elbow]

have a buzzcut as a start

[gleaming flame]

since I unsubscribed from leoAI, I might look for other payed service... so trying out wan AI for now... any other suggestions for good quality anime/cartoon motion?

[wooden sail]

Grok

[near elbow]

coffee for you?

[near elbow]

mixing fairies and mermaids 🙂

[gleaming flame]

she can swim, she can fly, but can she walk?

[keen hollow]

Did you pick a new service?

[gleaming flame]

not yet

[gleaming flame]

more dragonlings

[near elbow]

jam

[gleaming flame]

scared crow

[near elbow]

midflight

[wooden sail]

Better than expected

[wooden sail]

Is that a :pony: based model? I recognize this style for faces

[near elbow]

to the contrary - that's seedream, very nice anime vibes (and quite SFW)

anime style waifu flying superhero speech bubble text "Just to prove that this cannot be SD 1.x or SDXL in the slightest", superman below, batman below, wonder woman below, the hulk below, subtitle text "or it would all blend together"

[keen hollow]

Nice. Wan?

[near elbow]

e

[wooden sail]

It's kinda cool how Wan 2.2 keeps style consistent. A better frame generator/image editor than Qwen

But it makes sense: style and characters consistency come from temporal consistency of a video model. I think Nvidia has image editor build on the same idea of termporal changes: ChronoEdit

Ok wan is good but not perfect

• It generates animation therefore smudges
• Resolution is small

If only existed some model/tool that turns frames of animation into crips images by eliminating smudge frames

[dusky linden]

small? i can gen at 720p and then upscale

though the 480p model can do 720p just fine and the 720p model is a pure meme

[wooden sail]

U prob right

[blazing remnant]

Hello. am currently training a chihiro LoRA but it is failing with poses, when i inferenced https://civitai.com/models/71823 LoRA, i find that it is great with poses, how do you guys think they did it

[dusky linden]

and your both using Mistoon_Sapphire?

but looking at the example images i only see basic poses?
this one didnt even prompt for a pose (masterpiece:1.4), (best qualit:1.4), (high resolution:1.4), ogino chihiro, ponytail, yunaifomu, barefoot ,colorful, high contrast lora:chihiroOgino_v11:1

[blazing remnant]

Am using Animagine XL V3.1

[blazing remnant]

Oh yeah so i made an apple to apple comparison and i feel like mine is doing worse, i mean look at her character (civitai lora serious expression), and mine, there is a difference in terms of pose and theirs brings it out perfectly

[dusky linden]

Well training a lora is more then just chucking some images at it.

To be fair, both are kinda bad but i assume thats because its SD 1.5

You could improve the poses by using controlnet

But faces? In sd1.5 good luck with that , i just noticed it wasn't using xl

[blazing remnant]

Actually the image on the right is using xl

[dusky linden]

Hmm well then it's not an apples to apples situation

Probably just the source images

Could try using a illustrious model and a ghibli lora as illustrious is kinda the best in anime style images

Like that's an 2 year old model your using, i guarantee it will be better on illustrious

[blazing remnant]

oh thanks am gona try that

about using control net, wouldnt that mean i would need some kind of poses to guide the lora

[dusky linden]

No? With controlnet you can add a reference image for a pose

But I'm sure illustrious has the character you want built in

And before you start making loras i do recommend understanding the models a bit more

[blazing remnant]

Well would that happen in the training part or at the inferencing stage

[dusky linden]

Differences etc, if you got beefy hardware you can do it locally

Neither, you do it in the generating stage in the UI

[blazing remnant]

Okay been of much help, thanks @dusky linden

[dusky linden]

Np, hope you figure it out, illustrious has tons of loras for styles and built in too so just mess a little with that and see what you can create

Be sure to use the model page for the recommended settings and prompt structuring from the attached images/gallery

[blazing remnant]

Okay🙏

Maybe one last thing, would you advise i use Mistoon_Sapphire for illustrious

[dusky linden]

No absolutely not

Because that's not illustrious

That's sd1.5

Illustrious is a offshoot of sdxl trained only on anime style images

And uses danbooru tags, on civit ai filter for "illustrious"

[blazing remnant]

okay

[near elbow]

"1980s style anime cel shaded neko female giving thumbs up" with illustrious

just "neko female giving thumbs up" same seed

"photograph cinematic neko female giving thumbs up" (definitely not a realistic model but nice semireal touch)

and if you add "explicit" or "nswf" then, well, there goes the neighborhood 😉

not showing that here for obvious reasons, but it's definitely trained on .. spicy and unusual stuff

[near elbow]

bring in the stereotypes

[keen hollow]

grabbing random prompts and models, some fun results

[keen hollow]

ooo I got a heart from Veggie

[cold talon]

[keen hollow]

I struggle with glasses. Always get thin janky ones.

[cold talon]

that's how it is usually

[keen hollow]

"round glasses" works

[near elbow]

"waifu head horn rimmed glasses"

"waifu head nerd glasses"

[keen hollow]

yeah I found a few tags

[near elbow]

"waifu head big thick glasses" for full measure 🙂

[keen hollow]

Ok, I'm done for now, that's enough.

[near elbow]

morning

it's SDXL if your coffee cup contains COFTFETE instead 🙂

can't argue with the waifu results, though

[near elbow]

cheesy wholesome neko mecha interaction with sora 🙂

[near elbow]

and more action

[terse ore]

How can I use qwen edit 2509 to make an explosion at the base of the building? Im using the basic workflow with lightning lora but the output becomes blue and the whole image changes without explosion. Do I need to inpaint?

[near elbow]

seedream with "add an explosion on the ground to this black and white image, so it is clear that the female is trying to escape the catastrophe, keeping the rest of the image intact"
well, the "keep intact" part is a bit relative, she's hovering nicely 🙂

another attempt where seedream didn't get that this is a view from the top 🙂

[terse ore]

Ty but I want the buildings and everything else to be the same too. I made it in 3d to have consistency and ran it through AI to make it manga

[near elbow]

sounds like quite an effort... maybe inpainting with some explosion is easier

[terse ore]

The second one was definitely better...ill try krita inpaint

[near elbow]

always anime models supersizing the poor waifus

whatever he wants

[gleaming flame]

about to crash

[fossil dagger]

https://youtu.be/F66BsX3oVk4

[wooden sail]

Changing environment too

[wooden sail]

Nah

[keen hollow]

Happy Halloween!

[pale mural]

🎃

[near elbow]

animated with grok, on the SFW side

[keen hollow]

Love it! Wan? I like the ghosts 😄

[wooden sail]

Yes, Wan-RadiantCrush

[gleaming flame]

my monkey wolf for song cover image

[near elbow]

leftovers

and some space marines coffee break

[nova remnant]

what model did you use?

[near elbow]

seedream - really excellent, unfortunately too big for local generation

[prime barn]

missed it by a couple days, I only just heard this discord exists lol

just morning stuff

[wooden sail]

To my surprise it works

Girl with white hair in the hood is holding a lantern. Suddenly light in the lantern flickers and lights are off for a moment darkening except girl's eyes that glows red in the dark. Then lantern is lit again and scene reverts to normal. Girl with white hair in the hood is holding a lantern.

[near elbow]

beware of the local tieflings

[daring radish]

Girl with white hair in the hood is holding a lantern. Suddenly light in the lantern flickers and lights are off for a moment darkening except girl's eyes that glows red in the dark. Then lantern is lit again and scene reverts to normal. Girl with white hair in the hood is holding a lantern.

[near elbow]

so much ice cream

[gleaming flame]

that's really not that much

[robust peak]

Good luck

[near elbow]

magic

[keen hollow]

too quiet...

[twilit portal]

I prefer short haired Galbrena

[wooden sail]

"Not so fast"

[near elbow]

"long haired Galbrena, absurdly long hair, hair everywhere"

covering the ground for miles and miles

[twilit portal]

Bruh

[wooden sail]

It's fun and crazy how you can now "look into" any scene with ImagToVideo models

I think this AI image/video creation and manipulation will be much more controllable and powerful in the future. World models that will let you manipulate whole scenes at once, etc.

We're just scratching the surface with this technology

[near elbow]

yeah, and back in 2022 when stable diffusion first came out (and this discord was still the center of the universe and not the ghosts of echoes of the past) this would have been unthinkable

adding an unhappy ending, thanks to grok 🙂

[wispy canopy]

i miss those days, the chat sometimes went so fast xD

[near elbow]

and they deleted all the archive channels 😕

[wispy canopy]

bastards!

[wooden sail]

We're just people who talks in the foyer of an empty corporate building

[near elbow]

hoping that security doesn't throw us out, and that the building does not collapse 🙂

"anime style, dark dystopian, people who talk in the foyer of an empty corporate building, hoping that security doesn't throw them out, and that the building does not collapse"

adding "warhammer 40k themed female space marines neko" because why not 🙂

[keen hollow]

Woop that is fantastic. Wan?

[keen hollow]

Whoa that is one big dino

[wooden sail]

Yep

[keen hollow]

the details in this one came out pretty well... some sort of extra detail lora?

[wispy canopy]

no lora's just promptin.

@keen hollow but... it's a failed test -> it fails some specific aspect of what i actually wanted to achieve

[keen hollow]

nice... I tend to lose some details going up to a resolution that high. the rocks stood out to me with the cracks and stuff still in them.

[wispy canopy]

the ETA (sde randomness) is 5... where the usual is 1. so i'm really just "rerolling" on each step of generation

i noticed the effect of that on my workflow when iw as testing

[keen hollow]

I don't know what that means. 🙂

[wispy canopy]

yeah, i'm pretyt sure i'm very much removed from "normal" people at htis point with my images xD

but. i get nice images. which is the end result, right?

[wispy canopy]

did you know because of the stable diffusion server, there's a couple? 😄

... the people depicted here looknothign like the real life counterparts tho

[dusky linden]

Hmm who did you have in mind, im not up-to-date

Cause if i have the right people in mind, its been over for a while now

[wispy canopy]

My and gf xD we met here

[dusky linden]

Ohh lmao, because same here

Well now ex but yeah lmao

[wispy canopy]

gz and less gz

xD

[dusky linden]

Still, gratz man enjoy

[wispy canopy]

look ma, prompt only

[gleaming flame]

I'm digging through older images / archiving favorite styles... and trying old prompts with new models...

[keen hollow]

is that what you were going for the last time?

those are some large ears 😄

[wispy canopy]

Just trying some other sampler settings there

[granite cipher]

hello

[near elbow]

book lover

fantasy fun

[wispy canopy]

one is a booru type prompt, the other is disciple of slaughter

sometimes life is just not fair...

[near elbow]

"soft studio ghibli anime style, action scene, a female ninja warrior sliding on the ground towards you, katana drawn, dramatic pose, looking at you" with qwen

careful with that foot!

[near elbow]

we fix it!

different style

[near elbow]

stylish?

[granite cipher]

hello

[wispy canopy]

Don't know if it's ok linking my model here? https://civitai.com/models/2114323/hasphoria-llm-prompting-on-illustrious LLM prompting on Illustrious, i believe that's the way forward for anime models!

[pulsar dome]

Chibi girl, 14-year-old police cyber intelligence agent, luminous bright purple hair in high ponytails, cyberpunk city background at night with neon, fluorescent color palette: Electric Purple, Cyan Blue, Lime Green, Black. She wears a sleek black tactical crop jacket with glowing purple circuits, durable cybernetic gloves with holographic HUD panels, matching high-tech combat boots with neon soles. Performing the flowing Tai Chi Chuan pose mid-kick, a small luminous purple dragon orbits her fist like a living energy weapon, neon light dragon trails. kung-fu +

[wooden sail]

Look like Chroma

[keen hollow]

these sites that require linking to your gmail or discord or whatever to "register" make me irate

[wooden sail]

You mean OAuth login? It's either that or "confirm your email"

[keen hollow]

yeah I wanted to download some models from this "seaart" site, but it "doesn't take email sign-ups"

and I'm not connecting my accounts to them

so I guess no downloads for me

[gusty palm]

https://myaccount.google.com/connections

you could always sign in, then go here and disconnect them

it shows aall the places youve signed in via google

[keen hollow]

Yeah I know... and I could create yet another dummy account. I just won't do it out of principle. 😛

[gleaming flame]

meh... checked the site just now, found 2 mildly interesting loras. One wasn't available for download anyway, the second one is on civitai as well

[keen hollow]

There was a specific model I was after that isn't available on civit.

[near elbow]

"anime style but cinematic epic lighting, cute warrior female, bokeh"

"hard contrast low detail grunge anime style, cute warrior female"

"black and white classical manga style, cute warrior female"

[wispy canopy]

no lora's. just a weird workflow that makes artist styles pop 😄

[near elbow]

full transparency

[glossy basin]

cool

[near elbow]

and more fun with qwen image

[wispy canopy]

mlady

[terse ore]

For qwen 2509, can the controlnet union be used without needing canny, depth, etc? Just the image (like with xinsir controlnet union for sdxl)?

[terse ore]

Or is there an image to anime/manga style lora for 2509?

[analog thunder]

Can it be shared?

[sour rivet]

/anim Generate identification photos of people with different hairstyles using a nine grid pattern in the first-line images, with random backgrounds and clothing.

[near elbow]

with qwen running locally

[near elbow]

"random manga cover" already looks quite like manga

[sour rivet]

lion

[terse ore]

I mean for img2img. I want to use blender renders and run them through AI to change the 3d blender style into manga

[terse ore]

I made this in blender by importing a 3d anime character from a character maker, and a 3d building. Then I ran it through a custom sdxl workflow to make it look like manga. When I try with qwen, it turns it black and white but still has the 3d look; it's as though it only went through a cheap b&w filter.

[near elbow]

hm, this is just using your b&w image and saying "transform this into a vibrant anime style"
qwen image edit running locally, no special magic. i wouldn't use SDXL anymore if you have the GPU horsepower, it's just lacking in so many areas compared to more recent models.
using blender as a starting point sounds like a very big initial effort, maybe just use some character shots for consistency and then work prompt-only? even specific poses are not an issue anymore with just prompting.

"transform this into a vibrant anime style, adding demon wings" 🙂

"transform this into a vibrant anime style, and make her balance on one foot, arms extended"

well, you get the idea

[terse ore]

I don't have the render on me right now but it's the basic 3d look and I turned it into manga black and white style. So I wanted to do that with qwen because it should have less artifacts so means less manual work for me when putting panels together. But I wasn't able to do it. Not sure why. I can send the workflow later.

[near elbow]

and of course good text rendering, too (just to feature it explicitly)

[near elbow]

well, just working from your screenshot and doing the anime style prompt should do the job.
are you running qwen image edit locally?
no magic at all, i'm using the normal image edit workflow (enhanced with a "scale proportionally" node and file name output details, no impact on the actual rendering)

[terse ore]

Yes, and I was using the nunchaku one. And i wanted to use blender renders for the character and background consistency, and setting up angles just like how I wanted

[near elbow]

hm, i understand where you're coming from, and with SDXL that's an issue.
personally, i would probably rather train a lora with the characters than start in blender, and use qwen or a more advanced model for the rest, prompt-only. probably 10-100x as fast 🙂
but it has to fit your style and requirements of course

[terse ore]

Yes I was thinking about character lora but there is 3d character maker that let's one make a custom character easily and fast. Let me show you my manga I made with sdxl, 1 moment

[near elbow]

sounds like an interesting use case

btw, "transform this into vibrant anime style, keeping content and composition unchanged"
qwen doesn't get the "looking from above at the building below" perspective in the 3rd panel

but that's just a first try, probably can be fixed with prompting (too lazy to try right now 🙂 )

[terse ore]

I used this workflow

This was the input

This is the output

You can see it doesn't look like manga at all, but just a black and white filter. I also tried using it without lora also

I tried variations but it keeps turning out like that

(I had to drop my brother off at home so I quickly grabbed these on the PC)

[near elbow]

if i do that locally with your prompt "Real life anime, change the image to greyscale manga anime art style, with simplified shading and sharp shadows" i get similar results

but if i just say "transform this to classical cel shaded greyscale manga anime art style" it's much more manga like. you might want to play with that prompt a bit

"transform this to classical flat 2d cel shaded black and white manga anime art style"

yeah, grayscale is probably better, too rasterized otherwise

[terse ore]

Ty, I'll play around with different prompts

[lost creek]

Hello I've been trying to dip into AI gen as of recently, since I finally got my own PC with (weak) graphic cards instead of laptop
My question is when prompting multiple characters, what mitigation/negative prompt you use to avoid repeat characters? I know "cloned face" is one of them, but is there any other prompt that could weight in and reduce probabilities?

[terse ore]

The upper right panel in the manga (same panel you did img2img on) was done with sdxl and it looks way better. I think it's because it takes away a lot of the ugly 3d style and makes it look 2d. But qwen is too good at adhering to the composition, so the ugly composition remains after with qwen

It's also brighter and all the qwen images I try are too dark

[near elbow]

use a modern model like qwen or hidream - flux is ok-ish but doesn't keep characters apart as well.
this is simply "a crowd of diverse varied waifus" with qwen

[desert island]

qwen? I thought qwen was an llm? man I've been out of the visual side for far too long.

holy crap they have so much more then language models now! @near elbow

[near elbow]

yeah, qwen image, qwen image edit etc etc - and there's nice 4 step versions that run rather fast (12s per image on the 4090 which is 2-5x faster than hidream or hunyuan image or wan image)

ah, that wasn't meant as a reply to that

[near elbow]

have you tried adding "high exposure" or "bright image" or something like that? you can change nearly everything with the prompt
you can also change the overall composition with qwen, e.g. "show her from the side" does this to the example image. just tell the AI what you want to see, it's like magic (when it works, which is not guaranteed 🙂 )

[desert island]

😭 so many more models I have to test out 💢 😠 AGHH!!! now I have to look their vision model too! too many things to keep up with.

[near elbow]

personally i'm fine with the fp8 image and image edit models plus 4 step loras resp. 8 step lightning loras

[desert island]

is there an anime specific model or is the base model the only thing I need? https://civitai.com/models/1864281/qwen-image

[near elbow]

nothing anime specific, qwen (and the newer models) know anime/manga very well already

the comfy guides are quite ok https://blog.comfy.org/p/qwen-image-in-comfyui-new-era-of

[desert island]

cool thanks! you're the best 🫂

[near elbow]

and for speed e.g. https://github.com/ModelTC/Qwen-Image-Lightning/

[desert island]

what's your gen time with lightning and what hardware do you use so I can sort of gauge my expectations and resolution you typically render at.

[near elbow]

4090, 12s for 6 step image generation with the 4 step lora and around 30s for image edit with the 8 step lightning lora

all at around 1024x1024 (upscaling afterwards is cheaper than generating at 4k of course, and i rarely upscale anything anyway)

[desert island]

cool thanks I'm on a 4080 so I only have 16G of vram but I think I can make it work

[near elbow]

with the distilled models that should be totally ok

have fun

[desert island]

if it's a 20 gig model will it even fit into my vram?

[near elbow]

use the quantized versions

[desert island]

ughhh damn it woops

[near elbow]

i'm just playing with the 4 step lightning lora for image edit, works nicely as well, rendering times down to 16s from 30

https://huggingface.co/QuantStack/Qwen-Image-Edit-GGUF

for the quants (the 4 step loras are in the lightning link above)

have an image for variety (or we need to rename the channel to "qwen tech support" 🙂 )

[terse ore]

Yep, and I have the new nunchaku update and stuff looks a lot better now. I notice also that a change in composition takes away a lot of the 3d look because qwen makes its own instead of adhering to the messed up 3d look (particularly noticeable with the hands)

[terse ore]

https://huggingface.co/nunchaku-tech/nunchaku-qwen-image-edit-2509/tree/main/lightning-251115 you can also use the nunchaku version, and here are their newest updates with the lightning 4 step and 8 step loras baked into different models. On my 3060, it takes a 1536x1536 image to run through in just half a minute

[bold ether]

that s pretty cool.

[terse ore]

I really wish though, there was a mistoon_anime version for qwen made. It's a really great and popular model that's been made for so many other base models

[near elbow]

update: i managed to get that 4 step image edit to work, for me the trick was to use the wheel installer with version 1.0.1 (https://nunchaku.tech/docs/ComfyUI-nunchaku/workflows/tools.html#install-wheel-json)

down to 12 seconds for a 1k image on the 4090, can't complain 🙂

quality looks very similar to before

[desert island]

I had too many problems installing so I gave up for now. I'll come back to it later when I have more time but I ended up futzing with it for like 3 hours last night basically I ran into an issue where the main comfy loader node said it didn't recognize the model. I grabbed all the necessary components my models are in the diffusion model folder I am using the native workflow in comfyui.

ERROR: Could not detect model type of: /home/nori/comfy/ComfyUI/models/diffusion_models/Qwen-Image-Lightning/Qwen-Image-fp8-e4m3fn-Lightning-4steps-V1.0-bf16.safetensors

I just updated comfy and the same issue... I wonder if I'm missing some required package or something and I'll probably have to go sift through my logs to find out where it complains.

[terse ore]

https://m.youtube.com/watch?v=K7ZQpHiCV50&pp=0gcJCR4Bo7VqN5tD I had install issues, this worked for me

[desert island]

nope I tried nunchaku but the installer node isn't working either. doing a big update to see if that fixes it.

[near elbow]

i had that issues as well IIRC, i switched out the model loader to another node type and that solved it

[desert island]

I'll try that next

the logs aren't any more helpful

got prompt
ERROR UNSUPPORTED DIFFUSION MODEL /home/nori/comfy/ComfyUI/models/diffusion_models/qwenImage_qwenLightning8Step.safetensors
!!! Exception during processing !!! ERROR: Could not detect model type of: /home/nori/comfy/ComfyUI/models/diffusion_models/qwenImage_qwenLightning8Step.safetensors

Traceback (most recent call last):
File "/home/nori/comfy/ComfyUI/execution.py", line 510, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/nori/comfy/ComfyUI/execution.py", line 324, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

I tried with a couple different diffusion model loaders and still the same thing... you'd think that would tell me something 🤔

this is gonna take longer to figure out then I have time for ATM. I gotta get ready to go teach so maybe I'll play around with this more later.

[desert island]

god damn it I'm gonna have to create a totally new virtual environment for this to work. my python version is 3.12 and I need that for my other projects 🤦‍♂️ looks like I need to be in 3.10

[near elbow]

have you tried portable comfy? comes with its own venvs and all, no messing up anything

anime evolution thanks to seedream

[desert island]

I have not tried portable.
I love this workflow didn't you showcase this a while back? I feel like I remember seeing this before. It's still super impressive and I'm actually working on something similar now so I can take my generated sprite sheets and create portraits for the created characters then use a lipsyncing tool to animate their speech and dialogue with index tts... tried to get f5 to work for tts but struggled.

[left wind]

Trying out FLUX 2 seems great

[wooden sail]

Henshin!

[near elbow]

action

[wooden sail]

More action

[wooden sail]

Even MORE action!

[terse ore]

What should I use to inpaint?

[near elbow]

nowadays, an image editing model where you just prompt for the changes you want... flux kontext, qwen image edit, chatGPT, grok, gemini...

[near elbow]

generic cute manga image

[wooden sail]

I like how ZIT can generate pairs like this

[terse ore]

Anyone try qwen edit in krita? It only makes minor modifications to the part of the image I'm trying to change, but doesn't remove it as I'm trying to

[near elbow]

hm, i'm running it locally and all is fine

just a random image to stay in the mood

[near elbow]

don't mess with them

[near elbow]

mecha pilot

[merry tendon]

back in the day i used to post a lot here, now its ded :(

[near elbow]

didn't get the discord logo but neat otherwise

another take with z-image... it's too photo focused

read the small print (lower right) 🙂

[merry tendon]

discord logo?

[glacial coral]

chat what is ur guys "must have" lora for illus or noob based model

[dusky linden]

Theres no must have tbh, just the most compatible for what your goal to make is

Or the 4k/stabilizer lora to spice it up

[glacial coral]

Long time no speak XD

[glacial coral]

Stabiliser is what I have

[dusky linden]

I come and go as the waves, especially as of late as its hella dead on this server

[desert island]

I was like... "I can do better... just need a few more gens"

[desert island]

well I'll post this one anyway 😁

[desert island]

you walked into it.

[desert island]

getup! get her back! SNOWBALL FIGHT!

[desert island]

Okay lunch time 🍕

[near elbow]

"a crowd of comicon attendees varied diverse, standing in front of a painting showing miku rockstar" - amazing 🙂
(z-image)

[near elbow]

wave of zombies for you 🙂

[desert island]

guy on the left looks like "son of thistle"

[near elbow]

overdue for a cleanup

[wooden sail]

https://tenor.com/view/tubbie-fell-down-gif-22660534

[bold sail]

Create polemon

[near elbow]

slight spelling errors but z-image is quite nice

[dusky linden]

Z image is peak honestly considering its the turbo model, imagine the base/anime one

I just hope its not 1girl prompting

[dusky linden]

Turbo doesnt use 1girl prompting

The anime version got the noobai dataset and more so it probably will but i hope it wont

[desert island]

anime with base turbo is pretty good too! 👀 cool model!

[analog thunder]

nice-looking

[oak spoke]

one on the left is literally straight out of PVZ

[oak spoke]

now the unholy angel

[dusky linden]

"eyes up here pal"

ssince when do we watermark ai genned images

[dusky linden]

While i like it im not sure the mods will👀

[near elbow]

that's what i just noticed as well 🙂

ok better 🙂

[dusky linden]

Well tbh, other then the occasional spam bot and the clueless posters in tech support who else is here

[near elbow]

no need to risk anything 😇

[dusky linden]

Fair enough lmao

[stone adder]

OC for an Isekai yandere role play prompt I’m messing with. Name is Sachiko Kuroda.

[wooden junco]

Oh didnt know. Sorry. Should i remove it?

[glacial coral]

U tried ZiT yet

[dusky linden]

Z image turbo? Obviously

[dusky linden]

Here lol

[glacial coral]

U like it?

Mac Ui just got updated to support it 6bit, FP8 and FP16

I use FP16 cuz like im tuff like that

[dusky linden]

Its really good but the creativity is limited

Oh i use bf16

[glacial coral]

I do have BF16 support

[near elbow]

if you mean the "all the images look very similar despite different seeds" thing, change the sampler to res_multistep with sgm_uniform scheduler and 4 steps, that takes care of it

much more variety

[dusky linden]

Sometimes but personally it's also not knowing all the things i want it to know so I'm awaiting the base or anime model

[near elbow]

yeah, it's not trained on booru tags and for anime you need "anime illustration style" or "manwha style" - but then it delivers nicely
(ok this one is very minimalist)

"stereotypical manga magazine cover" 🙂

[dusky linden]

Its nice but not the styles im looking for so yeah

[near elbow]

what else? 🙂

double neko attack 🙂

[near elbow]

... except for the 3rd arm 🙂

[quasi shadow]

It's been a while since I last saw one of you

[bold ether]

dang it did I miss a shiny new scam bot ? a vintage one ?

[dusky linden]

Oh theres many a day, you really could use a few more hands for this fr

Theres one in community rn

[near elbow]

z-image doesn't even know spam 😿

[near elbow]

nekos and yarn balls...

[keen hollow]

these remind me of kpop demon hunters for some reason

[glacial coral]

U lot seen this? https://huggingface.co/NewBie-AI/NewBie-image-Exp0.1

Looks quite good

@dusky linden u might be interested

[dusky linden]

oh yeah ive seen it but its nothhing groundbreaking

especially its based on lumina image 2.0

[glacial coral]

Isn’t it better than SDXL no?

[dusky linden]

hmmm i think illustrious is still superioir

[glacial coral]

I mean it’s 20% in training rn but usable so who knows when it’s completely done 🤷

[dusky linden]

true but im more excited for z-image anime ver

[glacial coral]

Community made or official release?

[dusky linden]

official

[glacial coral]

Never heard of them man doing it

Got a link or sum shi

[dusky linden]

saw it in the swarmUI discord, they got their hands on the noobai dataset

[glacial coral]

@dusky linden also Z-Image runs quite fast on m2 pro chip with 16gb ram which is tuff I think ZiT might be the “2nd SDXL” in some way

[dusky linden]

thats just the tubro ver

idk if the base model will preform as good

in speed

[glacial coral]

Ik the base and edit models should be coming soon

But iirc the base is mainly for training stuff no?

[dusky linden]

or use aswell

theres a de-turbofied model for training already but unsure if it translates over well for normal ver

[glacial coral]

Yh I just don’t want that shi to be like chroma HD where let’s be honest that shi only looks like a training model to me

[glacial coral]

I mean Ui I got for Mac let’s u use a server for AI models and loras so if it’s slow as shit on the Mac I’ll just use that

[dusky linden]

like API or someone elses gpu?

not a big fan of that

[glacial coral]

Basically subscription based it’s the developers own servers running off Linux and people can use and import their own loras if they shall do so everyone gets 10gb of storage for loras

[dusky linden]

ouch i would not be down for that tbh

[glacial coral]

I mean since the Ui is only for Mac and iPhone and iPad they ain’t the most fastest thing for bigger models but if u have a pc u can use ur own as a server on the UI and to rely on them

But most run locally when they can

But laptop hardware moment

[keen hollow]

Nice. Wan? I like the transformation one.

[wooden sail]

Yes, Wan 2.

[near elbow]

mecha vs waifu, round 1

[pale mural]

A lot of great ones since i last checked

[fossil python]

Hi everyone
I'm looking for a list of anime models to use for kids from 3 to 6 years.

I m also interested for see the best ways to have character consistency ( face, hairs) , I m trying lora training but this is somehow a resource consuming process a time wise long

Any resources out there to learn from?

[near elbow]

my personal recommendation would be to use a modern model like z-image, qwen or so (or even google gemini, grok, or chatGPT) and prompt accordingly, no need for a "kids anime" model at all. especially the latter ones are so censored anyway that even if the kids prompt on their own (which i would not recommend) should be safe

[fossil python]

These are good models but generate some generic images I need have results with faces from photos

[near elbow]

qwen image edit and "transform this image into an anime illustration"?

[mint canopy]

@near elbow

[near elbow]

that looks fishy, there is no such team
EVERYONE do not click this link

[glacial coral]

yeees the npc default illus model pick has been updated wai v16

[near elbow]

what's better than in v150?

winter wholesomeness with v150 🙂

[glacial coral]

No clue yet pass me the prompt I’ll try it with v16

[near elbow]

thanks - just checked on civitAI, reviews look positive overall, just downloaded it myself 🙂

one of the arguments was "better fingers" which is always welcome, and cleaner look, and better backgrounds

ok, same prompt and seed with v16 - some more details indeed, nice

[steady grail]

it looks cute

[glacial coral]

Yep till this day after 2-3 yrs with AI I’m still mid at promoting 😂

[near elbow]

generi superheroine

[near elbow]

trippy!

[unique chasm]

There are so many smart people here that I'm just scratching the surface.

[cold talon]

you're posting good stuff

[keen hollow]

I was messing around with z-image.

it does a pretty good job and it is relatively fast

I'm not a great prompter... was mostly just grabbing old prompts and pasting them in to see what z-image would do.

[near elbow]

for anime, you can use "anime illustration" or "cel shaded manga style" etc, it does that rather well

back to illustrious... they had to cut down on fabric costs for the poor mechanics

[outer pagoda]

/Prompt

What are the commands to generate an image?

[dusky linden]

#artisan-faq or by running it locally

[pine sierra]

this is NoobAI+NanoB+Relight+manual drawing and editing

[near elbow]

classical style

[tawny wasp]

Moth Monster Girl illustrious Apex_Ai
https://civitai.com/models/2255424?modelVersionId=2538891

[keen hollow]

she looks grumpy

[gleaming flame]

that's the way I like them

[gleaming flame]

not grumpy

i should work on her fingers, but ehh...

[gleaming flame]

moving away from red background

[near elbow]

grumpy?

more shiny

[gleaming flame]

"sulky" is the word I use usually...

[near elbow]

art nouveau poster (SDXL is so bad with text... but cute)

[near elbow]

tsundere maid

[tardy python]

how can i generate nsfw art?

[near elbow]

google for "uncensored image generation" and try to avoid the scams... or if you have a good graphics card, local generation is still the best

SFW

[tawny wasp]

Kisara Kimi to Koete Koi ni Naru - With You, Our Love Will Make it Through illustrious Apex_Ai https://civitai.com/models/2267992?modelVersionId=2552816

[tawny wasp]

Deme (vtuber) Demenishki - illustrious Apex_Ai
https://civitai.com/models/2268742?modelVersionId=2553710

[near elbow]

generic manga in red

[tawny wasp]

Aquwa https://civitai.com/posts/25565882

[terse ore]

What's the best way workflow for img to img to make consistent manga? I use 3d anime and want to turn it into 3d. Had an old workflow for pony and union controlnet but its old. Was wondering if there's something better now

[near elbow]

no idea TBH - my use case is one-shot images, no need for consistency.
with modern models, you can use input images and get quite good "yes this is the same person" vibes in any results.

e.g. using my cat girl from above as input image and "make her stand and look at the viewer smiling" with qwen image edit gives me this, good enough i think

"place her in a lush anime forest and have her walk there dressed as a witch"

even keeps the blue braid band

[near elbow]

need some bandages?

[gleaming flame]

she be fine, just needs to burn some pewter

[near elbow]

fishy

[dusky linden]

Fish coming from her spine doe

[keen hollow]

Do ya'll keep all your generated images? They are taking up a bunch of space, thinking I might just blow 'em away.

like, do I really need to keep all this crap?

[gleaming flame]

that's really up to you... I don't have the need (yet) to delete anything, but I might do that soon. Maybe except my favorites.

[keen hollow]

just a lot of junk... and it's only gonna get worse... just purged a bunch of old crap

[pale mural]

Put em on your own site.

Like endless scroll feed of all them. Oldest at bottom, newest at top.

[left wind]

If you use Comfy UI id recommend you don't use Save image node use preview image node only save the images your happy with that way you don't end up with clutter and you can save it to a custom folder making it clean and easier to handle.

[keen hollow]

Oh I would heavily cull (with the SD PNG Info app I made 🙂 .. no longer updated though!), they just add up over time. Then quality improves and what was previously considered good or interesting no longer is.

Nowadays I save very little. I basically just do cloud generation now.

[wooden sail]

Run script that convert all .png into .webp. 20 time size reduction. But extracting and saving metadata might be tricker. LLM to the rescue idk

[keen hollow]

the drive space isn't really an issue... just dont like having junk around

[tawny wasp]

My New V2 Lora Kris 🙂

[gleaming flame]

still some surprises hidden within illustrious models... no loras this time

[dusky linden]

basic miku posting

[near elbow]

well, i generate JPG intentionally, that's 500kb per image or so... a LOT of these fit into one gigabyte, so why bother 🙂

and i have the relevant metadata in EXIF and much in the filename, so it's easy to sort and find

if in doubt, let miku do it!

[keen hollow]

Tifa is concerned.

[gleaming flame]

is that cruel of me?

[tawny wasp]

Cecilia Aquarain - There was a Cute Girl in the Hero's Party, so I Tried Confessing to Her - illustrious Apex_Ai
https://civitai.com/models/2290997?modelVersionId=2578333

[north peak]

perfect! 👍

[keen hollow]

oh my gosh, moar

[woven gust]

hey, how do I save a startup for stable diffusion?

I want to have some positive and negative words already added when launching the web ui

I remember there was a way?

but I can't find it

[gleaming flame]

probably depends on what you're using...

[woven gust]

stable diffusion

[woven gust]

web app

[gleaming flame]

no idea what that means, aren't there many web apps forusing stable diffusion?

[woven gust]

what are you using?

[gleaming flame]

at tme moment I work locally with invokeAI only

and there I can use prompt templates from drop-down menu

but every app works differently

[woven gust]

is that better than stable diffusion?

[gleaming flame]

sigh... are you sure you're using a webui that's called "stable diffusion"? SD is a name of a AI model, not the software or webui running it

in invokeAI I use only stable diffusion models

[woven gust]

I have no idea. Hugging face?

[gleaming flame]

you run it online or on your computer?

[woven gust]

computer

[gleaming flame]

ok, so you're probably using Automatic1111 or comfyUI. these are the most popular I think

[woven gust]

in stability matrix I select stable diffusion

[gleaming flame]

either way, i'm not familiar with other programs than invokeAI, so I can't help you

[woven gust]

you say invoke has only drop down menus? Isn't that more for younger people?

[gleaming flame]

not only, but yes, it can be simple

not sure about what age has to do with it

[woven gust]

but it has many options?

like people learning to program, they use drag and drop bricks

[gleaming flame]

you mean nodes? yes, it can work on nodes

[woven gust]

a text window?

[gleaming flame]

but I don't have the need to use these, I'm a simple man

[woven gust]

could you show how your program looks like? I'm curious

it sounds like magic

[gleaming flame]

I don't go into advanced tabs very often. just text2img, img2img and inpanting... it works for me

[woven gust]

that looks cool

might be great for drawing too

[near elbow]

(with z-image running locally on comfyUI - google that 😉 )

[hearty heart]

love it!

[gleaming flame]

ok, it's probably not really an anime style at this point...

[near elbow]

this is not a place for commercial adverts

[wooden sail]

Wan Mocha

[near elbow]

"wholesome manga artwork" with seedream 4.5 - cute 🙂

[wooden sail]

Spoiler just to be safe 🌜

[near elbow]

not realy anything to hide there i'd say 🙂

have a wholesome sketch

[gleaming flame]

same model, same seed... playing with styles

[nova remnant]

@gleaming flame are you using sdxl variant for anime or something else?

[gleaming flame]

I'm using illustrious models for anime, so it's sort of sdxl

[nova remnant]

amazing stuff, didnt think illustrious would produce this level of quality, my guess was z-image

[gleaming flame]

never tried z-image, I guess my gpu might not be good enough for new models, but yea, I don't think there are better models for anime stuff than illustrious ones. New models are more and more refined

[nova remnant]

z-image are not very resource demanding, the models are about the same size as sdxl and they are 6b models which by design can run on vram that you use for sdxl

[gleaming flame]

oh, thats good to know, I was thinking that since they are newer then flux, I might as well ignore them xD

[nova remnant]

flux is an elephant

[gleaming flame]

might tale a look later,

what they're capable of

[nova remnant]

mind telling me how much vram you have atm?

[nova remnant]

for an example here is my blog with z-image mostly https://thedigitalmuse.tumblr.com/

[gleaming flame]

12gb if I remember correctly

[nova remnant]

you are in perfect position to run z-image w/o memory bloat

you can grab the fp8 versions for model and text encoders, civitai community are also releasing fp8 variants with custom tweaks, those file sizes for model is about 6gb

@gleaming flame https://www.stablediffusiontutorials.com/2025/11/z-image-turbo.html

[gleaming flame]

any recommendation what to use to run z-image locally?

[nova remnant]

go for fp8

i use comfyui

[gleaming flame]

i tried comfyUI a few days ago out of curiosity, and installed the thing right away

i don't use nodes and workflows, I want an img2img and txt2img options xD

[nova remnant]

well the challenge with comfyui for beginners is getting used to node based workflow, but once you understand the basics of it you'll realize its the most powerful client for all sorts of generative ai

[gleaming flame]

I might try with new version of invokeAI, since it's the one I use, but I have to migrate a sh#tload of images first, because with invoke clean install is best option... and then configure a sh#tload of models/loras again...

btw nice images on your site... sexy, but not offensive

[nova remnant]

thanks. i avoid the explicit gory stuff

[gleaming flame]

as do I xD

[nova remnant]

i would suggest using comfyui as a secondary installation on your pc to test and learn with and make the shift when you feel comfortable, its a very versatile interface with powerful features and often with 0 day support for industry standard newer ai models coming out all the time

[gleaming flame]

I know it's probably the most popular now, but I don't have any mental powers left to learn all that stuff rn... i need an idiot version of whatever I'm using for AI

and I used invoke since SD 1.4 so..

[nova remnant]

ahh ok, its all good as long as you are comfortable and proficient using a particular interface to express your artistic sides

[gleaming flame]

I think I'm profficient with invoke enough

[nova remnant]

and i can see your anime images are unique and aesthetic

[gleaming flame]

thanks for the info on z-image, will take a look, although for now at least anime models on civitai are few... it's mostly photorealism

[nova remnant]

yes, i enjoy using z-image a lot for its realistic texture

[gleaming flame]

she got wool in her ears... oh well ;p

[dusky linden]

swarmUI, normal ui to gen but also a optional built in comfyUI back end

[gleaming flame]

i'll check it out then

[dusky linden]

Dont do the stableswarnui thats the old one

Just swarmUI on github

[gleaming flame]

there, now she be full battle ram

[keen hollow]

I've been messing with z-image recently. SwarmUI is the answer--they now have built-in z-image support in the installer so it's ready to go. 🙂

Honestly dont know if Invoke has added support yet. I do like Invoke fwiw.

[keen hollow]

haven't seen a tumblr in a while, didn't even know it was still around!

[fresh jasper]

for some reason a giant ant invaded my image lol

[fresh jasper]

another invasion

[near elbow]

i hate comfy for the inconsistent spaghetti stuf as well, used to be a fan of automatic1111 and then forge, but comfy just has all models covered...
as mentioned above, looks like swarmUI might be worth a try as well? no personal experience, but sounds good https://github.com/mcmonkeyprojects/SwarmUI

[near elbow]

anime cel shaded style, a stereotypical manga cover, title text "z-image!", subtitle text "It does anime quite well, too" 🙂

if you try that with SDXL based models, you will start crying... not even to mention the text 😉

[gleaming flame]

If I ever need a text placed like this, I will use gimp or similar

if I try z-image, it will probably be with hope of better understanding complex scenes, but I'll probably finish it up with illu anyway, because I don't care about "generic" anime

I guess I will try it out anyway, not sure if in invoke (when I tidy all the mess with files I have atm) or with swarmui

[nova remnant]

they dont advertise aggressively and they recently went through some policy changes to apply moderation on outright explicit content, despite that tumblr is quite feature rich for multimedia

[keen hollow]

I see I see.

Do you find people view the page, or are people just generating their own stuff?

All z-image.

[nova remnant]

its a public blog, and view count depends on how they expose their links, as for me i'm just dumping images there that i want to keep

[nova remnant]

how did you manage anime this good with zimage? or is that custom tuned?

[nova remnant]

1 3 4 5 are best in this batch

[keen hollow]

Not custom tuned.. just the regular stuff.

meta should be intact

also all z-image

[nova remnant]

there is a custom zit model that focuses on anime illustration, im going to try that out at some point

uploaded at civit

people are also releasing zit models that are tweaked to focus on caucasian women and not jsut asian

[dusky linden]

Any tunes currently are pretty bad/rushed tho

[nova remnant]

community cant wait to tweak the model, and alibaba hasn't even released the base model yet

[terse ore]

What noobai model(s) specialize in 2d, as opposed to 2.5/3d? Ideally, the more detailed the 2d style is, the better

[keen hollow]

well that was weird.. I got 'timed out' for an hour 😐

wonder what I did.. maybe it thought I was spamming images

dont even know who to ask anymore 😛

[dusky linden]

Other then the community guides theres noone to ask anymore lmao

[gleaming flame]

"Don't cry mr knight. I'll fight you once you grow a little. I promise."

[tough valve]

Hi, I am AI agent developer.

[gleaming flame]

I don't think I ever tried making images of Sailor Moon before