What is Anima?
Anima is the first ComfyUI sponsored model! Please read the official announcement from ComfyUI here.
Please check the official repo for important details about how to use the new base model: https://huggingface.co/circlestone-labs/Anima
Partnership
A partnership with ComfyUI is very cool! I'm not part of this effort, but it would have been a real honor to be a Lora creator that champions the base model. I guess… I'll just champion it unofficially!
(Dear CircleStone Labs, Please collaborate with me! 🙇♂️ I'm already advertising your model for free with my Loras. 🤣 I don't want money just mention somewhere that I'm one of your Lora Champions!💪)
Training Notes
I made 10 Loras last night, and here are my findings.
1. You can train with Kohya-ss/sd-scripts!
1a. As of PR #2260 – https://github.com/kohya-ss/sd-scripts/pull/2260 – training is supported and easy!
1b. Huge shout out to https://civitai.com/user/duongve13112002 for the code to make Lora training happen!
2. Easy Training – For low VRAM settings, it looks like I am able to train Loras on 6GB VRAM. I'm not sure if it ever spiked up to my max VRAM capacity, but every time I checked nvidia-smi I saw my VRAM usage at ~6GB. I'm training with fairly low VRAM settings (resolution 768 instead of 1024), so I guess normal training might cost more VRAM. (Note: over 300 of my Illustrious Loras were trained at 768 resolution. Its not me being cheap is an intentional training strategy.)
3. Training on Tags – I trained all my ANIMA Loras using the exact same dataset as my Illustrious Loras without any changes, so training on tags looks good.
Output Notes
4. Illustrious × Flux – Full Flexibility – I'm still experimenting, but my initial impression is that Anima is the worthy successor to Illustrious and brings the power of FLUX to anime.
Flat color – https://civitai.com/images/121213475 and https://civitai.com/images/121213444
Text generation - https://civitai.com/images/121213486 and https://civitai.com/images/121213500
Vivid colors - https://civitai.com/images/121213493
"High Art" fancy abstract things - https://civitai.com/images/121213494 and https://civitai.com/images/121213473
High Detail Scifi & Fantasy - https://civitai.com/images/121213504
And of course plenty of sexy and anime
5. Prompting – I had no issues reusing tag-based prompts from Illustrious generations, and it handles natural language prompting beautifully.
Legal?
Quick AI analysis of the license agreement for Anima implies that we are free to share derivative works as long as it’s non-commercial. That works for me: I almost never do things commercially anyway.
ComfyUI themselves sponsored this model on the premise that it is “open” and they are quoted as saying: “Grant recipients retain full control over their model and license (as long as it remains open).”
Sounds like we can make and share Loras/models but I'm no lawyer, so DYOR 🤷♂️
Final Thoughts
We live in an era where there are lots of models that come and go: Chroma, ZImageTurbo, Neta Lumina, Pony v7, but this is the first base model since Illustrious that feels like it’s actually good for anime. I'm not saying that Illustrious is dead, but so far this is the only base model that has impressed me since Illustrious.
I'm still testing Anima so I'm not drinking the kool-aid yet, but it looks promising!
I'm looking forward to seeing if the community adopt it. CivitAI has not implemented support for this base model so we are still very early!
I'd ask you all to test out my new Checkpoint and let me know what you think but its not enabled for onsite generation. So instead of testing it... just look at it? https://civitai.com/models/2394819