REanimaTE - [BDD] 3D/CG Edition
I love seeing your sample images, it's really encouraging when you post your generated images in the gallery! Thank you <3
This is an alternative fine-tune from BDD that shifts the focus toward a sharp, high-cohesion 3D/CG aesthetic while retaining the core flexibility and stability from the base Anima model. By dialing back real-world images in the datasets, this version delivers cleaner lines, enhanced crispness, and a distinctly polished rendering style.
What’s New:
Enhanced 3D Depth: Sharper, more defined cinematic and CG rendering.
Retained Flexibility: Preserves the strong prompt adherence and structural cohesion of base Anima.
Usage & Recommendations:
Dataset Focus: Trained with a strong emphasis on high-quality NSFW aesthetics.
Core Prompting: Optimized and tested for 1girl subjects. While standard anime variety is fully supported, composition variety may be slightly tighter due to the focused training set.
Description
More realism, more contrast, more brightness, more flexibility
FAQ
Comments (23)
出图牛子太大了，全是巨屌。。
该模型基于“[BDD] Big Dick Diffusion”理论，因此预计会出现身材魁梧的男性
@CallMeMaybe V2好用 V3牛子太大，直接占了 下体 3分之2 🙂
@mekeme126com537 谢谢，我会着手处理的——也许很快就会推出V3.1版本
@mekeme126com537 你能将一些包含提示词或工作流的图片上传到图库吗？我想在训练过程中用它们进行验证。
It feels like something is broken compared to v2.
Can you explain! I'd really appreciate feedback on what to address
If you look at Lizardon1025's example images in the gallery also, I don't see an issue - do you have a workflow, or can you post images to the gallery, please?
v2 was an absolutely GOATed model.
v3 is more powerful but harder to use, needs more style tags, etc. the default style isn't nearly as nice
Thanks, can you please try to explain or show some examples? The testing that I did showed a significant improvement in the soft 3D aesthetic, while being more consistent with the original anima composition. I need to know how other people are using it though!
A few people have commented about issues with V3.0 - please post images in the gallery with attached workflow and I can use that to try to identify and fix the issues!
@CallMeMaybe I posted two images
@FallingAngel Thank you! Can you explain what difference between your expected output and what was generated? I want to clearly understand the issue as you are seeing it. The issue I'm working on for v3.1 is the added 'brightness' in my dataset ended up messing up skin tones and making them very bright and/or saturated - so super pale, or super dark skin tones. If you increase the CFG to 5.0 it's better without affecting the composition
@CallMeMaybe I added two extra images generated with version 2.0 to that upload, using the same prompts.
The issues I mentioned earlier regarding pale skin and character fidelity actually seem to stem from two different aspects.
The skin tone is likely a matter of style; the difference is so significant now that they don't just feel like two versions of the same model, but rather like two entirely different models.
As for character fidelity, the sense that the character "doesn't look right" likely arises because the stylistic shift makes the image look much more photorealistic—closer to a real person.
If judged by that standard of photorealism, however, there shouldn't actually be an issue with fidelity.
@FallingAngel Yeah, I increased the brightness and contrast in the dataset which shifted the output too far, I'm working on a 3.1 to bring it more in-line with v1.0, but maintaining the flexibility of 3.0
@CallMeMaybe Wow~ Can't wait!
Sir, is this usable as a base model for training character LoRAs?
It will work, but I wouldn't recommend it -- character lora trained on base anime should work perfectly fine with this model, and will be more flexible to work with other models as well. I don't see any benefit to training directly on reanimate - but you can if you want to. I worked hard to keep this as compatible as possible with base anima.
@CallMeMaybe Thanks for the info! Btw, I'm planning to train a LoRA for my favorite anime character in a 3D realistic style. Any base models you'd recommend?
@wwwmanmoon103 It depends on your dataset images, but I would train on anima base, then you can try using it with various finetunes (like mine) and/or combine with style loras to achieve your desired look
@CallMeMaybe Thanks a lot for your detailed advice! I will use Anima Base as the base model to train my 3D realistic character LoRA as you suggested. May I ask if you have any tips about dataset arrangement or training hyperparameters for this style?
@wwwmanmoon103 sometimes it feels more like an art than a science. Personally I have not been happy with the anima character loras that I have trained, so I would strongly suggest you experiment and test your results - then you can find out what settings make you happy.
@CallMeMaybe I’m looking for a base model optimized for high-precision UNT / 3D modeling aesthetic. Do you have any relevant recommendations, especially regarding Anima?