An attempt to create a realistic ANIMA model.
Version 2.0
Version 2.0 was primarily tuned for the "er_sde" sampler, steps 12-16.
Run: Euler/simple (Beta), CFG 1, Steps 8-16
Text encoder: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders
VAE: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae
Description
Another variation of a realistic model, take a look, try it out, maybe someone will like it. It seems to me that the realism has improved, and the picture has become clearer.
Run: euler/beta, step:12-16, CFG:1
FAQ
Comments (15)
I tested Turbo V2.3 and i'm not sure because of better to version 2.2.5 with Turbo LoRA but it is other. As you wrote, one person will like it, and the other won't. I use dpmpp_2m_sde/beta for me it looks a bit better then euler/beta but i'm not sure :)
I haven't yet found a setting for the second KSampler in my two-KSampler setup. No matter which sampler or scheduler I use, the image always ends up far too noisy. Thanks for the model :)
V2.3 has higher contrast which can be to much but overall it seems better than 2.2
Again proving that you make the best Anima models. Thank you for your work.
This model is awesome, thank you for your work. I made a img2img style conversion workflow using it and the results were very good: https://civitai.com/articles/31980/sam-anima-anime-to-realistic-image-transformer-with-auto-captioning-and-pure-high-res-fix
For me, version 2.0 understands the prompt better than version 2.2.5. Has anyone else noticed this?
The problem is that the more the model strives for realism, the less knowledge it has left, because the training dataset of those people who teach the model realism is weak.
@toya_san you should have a dataset with a very detailed description, except "realistic" keywords. They should not exist in description, so it will become a default. It's like a sorting system - it's enough to just sort main things at beginning of training to reach smaller loss, then it's more and more difficult. When model don't know what to do to decrease loss, it starts "breaking" by force - functional weights may be replaced with constant values which decrease diversity. Each finetune breaks model. Your goal is to minimize amount of unsorted data and leave on floor only what you want to stay.
I updated my img2img anime to realistic workflow to make it totally vanilla, only using standard ComfyUI nodes, and tried using V2.3 of the model, it works very well, thanks for your work!
This is still the best realistic anima checkpoint every other checkpoint i've tried just slops the image and fails to follow the prompts.
This model works.
Made a lora, and ran it through all the other realistic checkpoints. testing different samplers, steps, cfg. other models would break easily, or be too overcooked. but no matter if i used 8 or 30 steps, 1 or 5 cfg, added a turbo lora or used one of the fancy keywords other models seem to require, this model just kept going. not always providing the best output with the shenanigans I put it through, but always at least decent.
I started my gooning journey with 1.5, did sdxl, did illustrious, tried flux klein and zit, but with anima and this finetune, I feel like my journey finally has come to an end.
@Bkarloff i feel the same like your gooning journey :)
it is fucking crazy that i went from taking up to 2 minutes to generate decent images at 1080p in sdxl, to 26 seconds for 1920x1536 in Samanima v2.3 turbo and i don't even need detailer passes for the face and hands/feet. I'm so happy we made it here so soon after illustrious became the standard lol.
does this work on forge neo? and what type of model should choose for it, I tried everything and not working
This was done only for comfyui
man seriously out of all the models and versions for Anima i still come back to v2.0 all the time people need to get on board its poppin :D