RDBT | Anima2.9B
About Anima 2.9B: https://huggingface.co/Gazingstars123/Anima-2.9B
Temporary page:
More about the RDBT dataset, usages etc. see this page: https://civarchive.com/models/2356447/rdbt-or-anima
This is a test run, not a full run yet. Stopped at ~30% steps. This LoRA is expected to have weaker effects compared to the one for original anima base ckpt.
Versions:
finetuned v2:
LoRA finetuned with 10k images. Improve quality, stability.
distillation addon v2:
This file let you use CFG 1 and 16 steps. You need to load both this add-on LoRA and the "finetuned v2" LoRA to get best quality.
Sharing merges using this model is not allowed. This is a free model and it will always free. This "restriction" only aimed at those fake base model creators who steal others' models to sell.
FYI: all my models have special trigger words and latent watermark.
Description
FAQ
Comments (12)
Could you make a int8 convrot version of the base model?
Big boss just casually trained his LoRA even before the creator of the base model. Orz
"To-do: Lora Training will be supported via my Anima Standalone Trainer in a few days."
And a turbo lora... Orz
Hmm~ I wonder how many people will retrain all their loras for this new model 😅
models, there is 2.9B and 3.8B - i dont know yet how well it works but there are some bridge nodes that make loras work for both new models (i dont know how well they work as i not tested them thoroughly yet)
the 3.8 node loader works flawlessly and transfers style as clearly as day, can't tell anima base from anima 3.8b apart when looking for diffrence in style
Well done!
Are you planning to train a model based on this? I added a style lora on top of these two loras and it broke the image, so I hope you would train a model based on this one.
LoRA for original Anima won't work on 2.9b.
https://huggingface.co/Gazingstars123/Anima-2.9B/discussions/2
@reakaakasky But it does! https://civitai.red/posts/30387688
At least with latest commit of Forge Neo.
3.8B model cries from some RDBT lora much harder, the 2.9B one can use regular anima RDBT turbo lora with no problems.
But question is if it even is worth the effort. 3.9B is 90% slower than regular anima, and from what i could seen, it's not really better than 2.9B anima (40% slower than regular anima) so yeah.
and at the end neither of them has illustrious fidelity, nor either of the allows for hires fix.
Anima is frustrating to work with if you love high quality above everything else. There is just no good answers, only trial and error. :/
Kinda always like this, always there is something. Illustrious is great, has every toolest under the sun but text encoder is just absurdly outdated, and even if illu is best fidelity oriented checkpoint, it takes lot of processing time to get that fidelity (and some robust workflow)
Anima is light, small, compact, fited with tags and styles, has decent encoder, but fidelity is bad, images often grainy, you zoom in you get hearth attack how bad it looks, you can't do inpainting, you cant do hires fix.
Krea2 resolutions up to 4k no problem, amazing text encoder, but censored on multiple levels, dont know styles, dont know art, dont know tags, and tainted with realism so hard you need peculiar finetuned mix of loras to get some decent anime looks, and f-ing 10 nsfw loras 4 text refusion reductions and whatnot to just get nudity, or at times, just poses to work because censorship. Like for example if you tell character to lift leg in a knee, that already triggers refusal on official krea2 checkpoint version.
Chroma is just to slow, same with Flux
SDXL and SD1.5 are out of the grace, they are history.
I'm ranting hahaha sorry ╮ (. ❛ ᴗ ❛.) ╭
i'm just waiting for this perfect checkpoint to drop.
That will have razor sharp quality, will be malleable to work with inpainting and etc., good encoder that understands both natural language, styles and tags & will not be censored in any way shape or form.
and hopeful wont be some massive overkill requiring super computer to run ;) but one can only dream.
Should be happy with what we got, all this stuff, is so amazing, would never thought in 2021 that in 2026 i will just have all that on my pc and generate stuff all i want on a whim.
Only thin i miss to my perfection is ability to make my very own loras, i need to sit up on that in free time and get it done >:)
also there is whole AI music world for me yet to conquer (˶ˆᗜˆ˵) I Love Technology
my opinion on 2.9/3.8b is they are completely unnecessary. The original 2b architecture is capable to handle tens of billions images. You don't need to expand +100% parameters to learn just 0.5 millions new anime images. Those new layers are still very close to identical mapping, and are wasting compute. ( that's why LoRA from 2b arch still works)
I'm not spending resources on those ckpts because I'm questioning the skill level of the creators, are they really know what they are doing? don't think so. or they are just vibe coding? Very likely.