RDBT [Anima]
This page contains the "original LoRA files" for advanced users.
See the main page for more info. (ckpt version, usage, faq, int8 ckpt...)
See this page for update log and version info.
Sharing merges using this model is not allowed. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
Description
512px + cfg1 = 10x faster than 1024px. You can test your prompts/ideas in 3 seconds.
FAQ
Comments (20)
Will you be retaining these great LORAs with anima preview 2. Also how does your LORA work as it drastically improves the quality, stability, anatomy and overall prompted Adherence and image.
A good dataset. There is no magic. AI is "garbage in garbage out".
Captions are from LLM. Concise, efficient, and error-free.
Images in dataset are mainly picked by my artist friends, from artist prospective. I also rob my friends for their unfinished works (high quality sketches), which are not common in public. So you might have noticed my model is somehow good at sketch style...
Maybe it's just my imagination. Preview2 feels slightly worse than 1.(?)
@reakaakasky trust me, I've been testing it out quite a lot today and the prompt adherence is definitely better with preview2. It's still pretty wild and creative as the author of the model said on higginface but it's a pretty decent base to fine-tune cuz I think it has more knowledge and stuff like that. I said definitely feels a bit more fuller.
I would definitely try preview 2 as they said this this that's this in official higginface page. "A significant part of the training is redone with different hyperparameters and techniques, designed to help make the model more robust to finetuning."
I know that it's a new model, but if you have any ideas I would highly recommend that to you. Have a look at the community tab of anima higginface page the as there are some pretty cool people on there that have done some cool stuff like one guy did a ComfyUI custom node to support the Qwen 3.5 4B hybrid (Mamba2 + Attention) text encoder And other stuff And the main guy tdrussell seems to be quite active there as well. Like I think he's experimenting on a bigger text encoder but he said that he may or may not do it because base he has to relearn everything.
I'm just really excited about this model as very long time I've always liked Illustrious and NoobAI and back in the day. Good old-fashioned human hand drawn hentai back when I was a young adult. The only problem with Illustrious is it's based on outdated architecture now and the clip is definitely a drawback unless you have special ones like Rouwei-Gemma. Even ELLA for SD1.5.
I just love the open source community and especially creators like you Crody and the legendary WAI0731 (his checkpoint was the first Illustrious fine-tune I used) as you all are really amazing and you help make models much better too especially with your stability LORAs.
I really hope that anima takes off as I know the licence has made it pretty unpopular but even if you periodically ask for donations to carry on developing checkpoints and LORAs personally, I would donate to you to keep doing it as I think that this model really has improved with your stability LORAs and is a big reason why I keep coming back to it as the training is still quite new and it still has quite a few ways to go before it's fully done out of preview I think.
Thank you as always for your hard work and sorry it's a long message
I don't believe the license for anima will hinder its popularity.
The license doesn't restrict donations, so for model creators looking to monetize their work, the license shouldn't be an issue. nvm, idk.
Fine-tuning a model like this one RDBT doesn't require substantial computational power—it costs just a few cups of coffee. That’s actually why I’ve never felt the need to accept donations. The reason I’ve decided to stop training this model is simply that I feel uncomfortable with the license terms. It’s 50% due to personal experience, I don't trust any company, and 50% because of the datasets used. It's entirely personal reasons.
But I do believe the license indeed prevents large-scale fine-tuning, like NoobAI. In my view, 80% of illustrious's popularity is owed to the continued training provided by NoobAI. (FYI, all models currently labeled as 'illustrious' are actually NoobAI-based). But that’s a problem for the future.
Any chance you're gonna make this work with the new anima preview 2? The LoRA seems to only work with first preview unfortunately.
I've made one, but something is off. See https://civitai.com/models/2385278/animayume?modelVersionId=2753918&dialog=commentThread&commentId=1135805
@reakaakasky does the issue persist without your distillation?
@m4rbleye yes
Note: No preview2 version for now. I've finetuned one, but the quality is not good. So I'm sticking to p1.
p2 somehow can't generalize the style. Maybe I need to adjust something, but I'm too lazy.
3/28: p2's text encoder seems to be more stable, and will pay more attention to other words/styles within the prompt. Long prompt is not ideal.
And that commitment to the styles is why you're my favourite creator here. I really don't care about shiny gloss, I want the characters depicted faithfully to the source material, and the combination of NoobAI and your single lora handles almost everything.
Do you think you can work around the sepia thing on Anima?
I respect that and to be honest it most definitely takes a long time to train. And they could release more previews so it's probably best to stick with the first one. Hopefully the full release comes out this year too. Great work as always tho.
@AnimaXx "takes longer to fine-tuning" means the base model is overfitted.
I see some people also said the same thing when training LoRA on p2. It's not a good sign for a "pretrained" model.
@m4rbleye v0.19 seems fixed the sepia bias?
@reakaakasky unfortunately it still seems to avoid colours like the plague, at least with short prompts without colour instructions. whether it's tag, natural language or mixed doesn't matter.
As an aside, your lora dramatically improves this model's capabilities when it comes to photorealism.
Edit: It seems the quality tags influence the colours. I was not using them, as per your instructions, but when I added them for testing they do tend to give some colour back.
Edit 2: The Mission: It might be a case of "suffering from success". When the artist has a lot of B&W works, the lora seems to veer the image to that direction. Your NoobAI lora doesn't have this issue for me, and neither does the raw model of Anima or the AnimaYume finetune.
Edit 3: Tokyo Drift: A glance to my standard negative prompt on the raw model testings reminded me that I had "monochrome" in there (not possible to use with your lora). Removing that makes the raw model also lean to B&W when the artist has them in their repertoire. Adding quality tags give some colours back, as mentioned.
I would blame the base Cosmos Predict for lots of problems. Colors, contrast, aesthetics...
@m4rbleye You can use cfg >1. It won't burn the image like other distilled models.
I use cfg 3 all the time. It greatly improves image quality even with empty negative prompt.
yeah, cfg is recommended. I'll update the page.
my distillation is cheap. It works but it's limited. e.g. no artist tag (10k artists are too much for me), no text rendering etc... in target
v0.19: v0.19 is still based on preview1, not preview2. It won't work on p2.
When preparing cover images, I copy-pasted many prompts from other people. Thank you guys. And forgive me if I did not reference them. Editing civitai post sucks.
I have tried it with P2 and it works fine as it helps with the prompt adherence and stability. The art style and anatomy is a little bit of a wild but that is due to the P2 not being as stable as P1.
It might have effects. But LoRA needs the current base model. Otherwise the patched weights are in the middle of nowhere. Using it on p2 is not recommended.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.