RDBT [Anima]
This is a general finetuned + distilled model.
Better quality, better prompt adherence.
Dataset contains ~10k handpicked images with accurate NL captions from LLM, body/hands anatomy. Does not contain any shiny plastic glossy AI image.
Some cover images look ridiculous; they're just for demonstration purposes for corner cases.
I use this model as a clean starting point to stack more LoRAs.
See this page for update log and version info.
For advanced users: RDBT model is trained as LoRA natively. See this page for original LoRA.
For those who don't want to stack style LoRAs and is looking for a out-of-the-box ckpt: RDBT | Anime, which is based on RDBT finetuned base, but optimized for high quality digital art.
Links to base model:
prefix with ym: AnimaYume (hf link) (civitai link).
prefix with b (base), p (preview): Anima pretrained (hf link)
Usage:
Settings:
CFG: 1~3. This model has been distilled. You can disable CFG (CFG 1) and run the model 2x faster. Cover images are without CFG for demonstration. "RenormCFG" node is highly recommended if CFG is enabled (CFG > 1), set "renorm_cfg" value to 1.1.
Steps: 16+
Sampler: Euler (best diversity), Euler a/er_sde etc. (better stability)
Res: 1MP
Prompt:
Always specify style in prompt, or use a style LoRA. Otherwise, you will get random/mixed style. This is a feature, not a bug. This model does NOT have overfitted default style (which ignores prompt and is always active).
Quality tags:
Omit ALL quality tags. You don't need those. The fine-tuning dataset has higher quality than "masterpiece". Thus quality tags don't have effects. Omitting those redundant tokens allows LLM to pay more attention on other words.
Sharing merges using this model is not allowed.
This mode is a free and it will always be free. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
Known model thieves: NukeA.I (selling this model behind paywall on tensorart).
I wrote a story about it. Also contains a guide for trainers about "how to bake special trigger word into your model".
Description
FAQ
Comments (15)
Thank you, I was waiting for this update.
One thing I've noticed with Anima Base, in this and other models, is that the number of available characters within the model has decreased significantly.
In Usage, what does "Shift 3 or 1." mean?
Comfyui "ModelSamplingAuraFlow".
Doesn't required anymore in v0.35. Page updated.
peak
Hey boss, the v0.35 does a good job of stabilizing the base model, the anatomy is much better mostly. Though there seems to be some downgrades, even though you skipped step distillation, I couldn't find much considerable difference in prompt adherence or creativity from v0.34b. Your previous model made the images much shaper compared to the p3, especially for 3d styles. The v0.35 model looses the sharpening and clarity compared to the base v1 model.
I was using your previous models with CFG 2 with 24 steps, even though they were distilled and It made 3d images perfect with a loss of some prompt adherence, this model too from how much I have tested, works best with CFG 2 and 24 steps. With these settings, its the best anima checkpoint for 3d currently. Don't know if it means anything but I just thought, I'd share.
Great work though boss ♥, your models are still very good compared to all available on Civitai.
Also it has a bias of female pov
yeap. v0.35 I was trying to avoid the bias of always containing human characters in the middle of the image, so I added some pov images and weighted it to ~3% of the training dataset. Seems to be too high, and becoming a new bias. lmao
although in my prompt adding something like "male pov" can fix the issue.
From my initial testings it still seems to strongly favor some particular compositions, with squiggly, thick strokes and a certain style of soft shading with lines, no matter the style prompt. The raw Anima base does know and adapts to the prompted styles decently well.
prompt adherence is worst. i even add the weight to 1.5, e.g (lying, on side:1.5), the subject did not lying. while the base it self can do without need to add weight
Try natural language "subjectname is lying on their side on the couch with their head resting on their hand propped up with their elbow" for example. Quick edit to say do also describe your subject before that example, just use that as the "action" for the image, describe the rest of the scene details after if you don't like what its giving you
Since the model is based on Anima 1.0, weights can go up to 5. I usually set like (fishnets:2) and got better prompt adherence. Alternatively, using natural language and explain can increase prompt adherence.
v0.35 + turbo lora (I prefer 0.8 weight) + 16 steps seems to work really well for me. If you don't want to use turbo then raising CFG to 2-3 seems alright too but it is slower of course. CFG 1 no lora works but it makes the model not adhere to styles and character designs all that well.
I actually found this model to be BETTER at prompt understanding, especially natural language, than base 1.0. Some compositions I prompted it got right over 80% of the time while base only succeeded about 10~20% of the time.
The only odd part is sometimes the image will come out black and white or monochrome and changing the seed does not help even though nothing in the prompt suggests the image should be black and white (only has artists and characters that use color by default). But fixing it is pretty easy, just describe some stuff like character hair color and eye color and it should color in the rest of the image too.
file not found trying to download
It's incredible how you managed to create a high-quality model with generation times that are absolutely insane! 🤯⚡ Simply wowww... 🤩 I'm really looking forward to seeing more of your work. 👀🔥 It works wonderfully with my LoRAs! ✨ It's just a shame that on-site generation isn't enabled for it on Civitai yet 😔... they have no idea what they're missing out on hehe. 😏🚀