BlackHat404

/

DefacationIL

like 17

I need to do a bit of research into Anima first but currently it looks like it it's too different from IL in training. I was waiting until things matured a bit but maybe now is the time to start poking Anima and seeing how it behaves.

Update: I was able to train a Anima Defecation LoRA but it almost feels like I went back to Pony since the anatomy is wonky and going above 1024x1024 total resolution images(My IL images were 1280x1280 total) come out deformed just like old Pony Gens were. Can you link what checkPoint and settings you are using that are so much better than IL because right now my Anima gens(Even without my LoRA) are kinda bad. Prompt adherence might be really good like you said but quality and anatomy is not looking great.

For clarity I'm not trying to insult you or anything, just trying to figure out if I'm missing something.

I'm using Forge Neo
masterpiece, best_quality, absurdres, amazing_quality, dark_skin, blush, looking_at_viewer, flower_field, riverbank, river, wooden_bridge, tank_top, huge_breasts, sandals, facing_viewer, sitting, Negative prompt: bad_quality, worst_quality, worst_detail, sketch, censored, watermark, signature, artist_name, patreon_username, patreon_logo, worst_aesthetic, low_quality, lowres Steps: 25, Sampler: DPM++ 2M, Schedule type: Beta, CFG scale: 4, Seed: 3834460609, Size: 768x1344, Model hash: bb00b76213, Model: animaRealAnimeCATAraCAT_v10, Clip skip: 2, RNG: CPU, Beta schedule alpha: 0.6, Beta schedule beta: 0.6, Version: neo, Module 1: qwen_image_vae, Module 2: anima_qwen_3_06b_base

I just use the default ComfyUI template. The anatomy is definitely slightly worse, but I wouldn't say it's horrible. The cause is likely that the current preview version for some weird ungodly reason, was trained at 512×512 with only a small layer of 1024×1024 training at the end, plus base Cosmos knowledge. The full release should be fully 1024×1024, I'm pretty sure.

That said, I'd still like to see LoRA support even if it isn't perfect.

As for training, I was able to use this setup since my GPU isn't great, and it produced this model. However, I'm nervous about getting my Google Account nuked. They don't allow vanilla NSFW content, technically.

Script: https://github.com/gabengaGamer/kohya_ss_anima
Result: https://civitai.com/models/2466489/jacinthe-character

Checkpoints I use: https://civitai.com/models/2385278?modelVersionId=2798563

There's also this checkpoint made specifically for training, though it could be snake oil: https://civitai.com/models/2454865/any-anima-preview-for-lora-training?modelVersionId=2785294

Other benefits beyond just prompt understanding include art style knowledge and character recognition. I was able to get it to understand artists with around 200 posts even same with characters almost. The database can be found here. It definitely improves quality a bit to use artist tags e.g. @dikko: https://thetacursed.github.io/Anima-Style-Explorer/

Style LoRAs can help as well, since they're likely trained at higher resolutions: https://civitai.com/models/976750/merrytail-or-style

Finally, as for your prompt. I wouldn't use underscores outside of like one part of the quality prompt and I also wouldn't use such a long one it likely doesn't even understand anyway. Here is what I use

Positive: masterpiece, best quality, score_9, nsfw, @artist

negative: worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, sepia

It also is smart enough to understand natural language a bit, I usually try to stick to tag words but expand on them a bit with natural language if needed. E.g. Necklace might be a tag but rainbow heart necklace might not be but understands anyway

I'll be more interested in Anima once they release the next version of the model. They have stated that lora trained on preview should be considered "throwaway" models and would rather spend my time on other stuff(training uses my main PC so training a LoRA means no PC gaming until it's done) while things are still being finalized for Anima.

Also this is mostly a me thing but I don't like score tags or direct living artist emulation(i.e the style of an anime is fine) so I really have no motivation to use Anima until it's by and large better than Illustrious.

I do thank you for the detailed breakdown though, it's just that I don't personally gain much from the advancements Anima offers at this point in time. The goodish news is that based on the preview I can pretty much just retrain all my stuff with slightly different settings for Anima so making the jump shouldn't be too hard when the time comes.

I think Anima probably excluded scat from its training. (It seems to interpret scat as "cat," so increasing the prompt weight results in a cat :D)
Illustrious also gave poor results, but its interpretation of scat was not this weak. Models like Noob, which were additionally trained on e621 data, produce better results, but idk whether they'll(Noob team) do the same work for Anima.

https://civitai.red/models/2544636/wai-anima?modelVersionId=2859702 is considerable upgrade still not final model but Preview 3 and this model specifically have been big quality jump starting to maybe get closer to SDXL based models

https://civitai.red/models/2544636/wai-anima?modelVersionId=2859702 is considerable upgrade still not final model but Preview 3 and this model specifically have been big quality jump starting to maybe get closer to SDXL based models

I'll take a look and see how I like the newer version. Hopefully it's faster than V2 because that was a major annoyance for me.

Odd, about the same speed as SDXL based models on my crappy RX 6800

Might be a strange optimization thing, I use forge neo and Anima takes about 40 extra seconds to generate 4 images compared to Illustrious. It's really consistent too lol.
Some good news though, I think I figured out how to train Anima and I'm currently trying a bunch of stuff.

Might be a strange optimization thing, I use forge neo and Anima takes about 40 extra seconds to generate 4 images compared to Illustrious. It's really consistent too lol.
Some good news though, I think I figured out how to train Anima and I'm currently trying a bunch of stuff.

Odd, works fine on ComfyUI

Some good news though, I think I figured out how to train Anima and I'm currently trying a bunch of stuff.

Excited to see what you come up with, Anima is rather lacking in this area!

Just as an update, I'm still trying to train Anima for this kind of stuff but the model seems kinda resistant for some reason. I can get it to work but it's not as reliable as I would like so I'm trying different things to figure out how to fix that.

Odd, I was able to fine tune on character really well with just https://github.com/citronlegacy/citron-colab-anima-lora-trainer without any changes and some config on my local machine with Gemini modified version of https://github.com/kohya-ss/sd-scripts that was able to run on my GPU albeit unreasonably slow

Here was the result of that, didn't know the character before but reproduced it fairly well after. Likely would be a lot better if I didn't have to stop the training run on my phone due to irl things
https://civitai.red/models/2466489/jacinthe-character

Character and style training has gone really well actually, teaching new concepts is what I'm having issues with. I did a model for darker skin and a few futa model without too much trouble but reliable defecation eludes me for now.

Good news, I have a model with consistent results for standard defecation. Currently testing to make sure it doesn't have any weird issues before I call it a success.

Interesting, what did you do different exactly?

Played with the learning rate a bit based on some tips I found. Hopefully this goes well and I'll have something to post in the next couple of days.