Hello everyone!
This LoRA brings the distinctive style of Rude Frog together with more realistic skin, with a strong focus on very curvy girls and nude results that stay close to the original artist’s look.
All source images were carefully reworked from pure 3D aesthetics toward realism, so the model keeps the characteristic proportions and vibe of Rude Frog while delivering more natural skin.
There’s still room for improvement, but the current version already works well for both clothed and nude generations.
If you like this LoRA and would like to see future versions or ports to other models, leave a like and a comment with your suggestions.
Description
Hello everyone!
This LoRA focuses on very thick girls and delivers solid nude results that stay close to the original style of Rude Frog.
I’d like to present a LoRA that combines Rude Frog’s distinctive look with more realistic skin. Since Flux already handles realism well, I carefully modified every image with edit generations to avoid breaking that strength with pure 3D aesthetics and watermarks.
The LoRA was trained in a single stage on more than 1,700 images by the 3D artist Rude Frog. All of his works were reworked in an attempt to move them from 3D toward realism. Training settings:
10,000 steps
Resolution: 512–1024
Batch size 2, gradient accumulation 2
Rank 32
I believe that with longer and more careful training it would be possible to go up to rank 64 or even 96 with the current dataset size. A second stage at higher resolution for better detail is planned and I intend to work on it in the coming days.
Training was done with SimpleTuner.
There’s still a lot of room for improvement — both in anatomy and character tagging. Some image-level issues remain as well. Unlike the previous Anima version, this one uses natural language captions. Checking captions across 1,700 files takes a lot of time, so there’s clear potential for further refinement here too.
Please note that all the results shown above were generated using the Turba LoRA, Euler sampler, 12 steps, and CFG 1.8. Your results may differ from mine.
If you like this LoRA and would like to see the next version (on Flux or any other model), leave a like and a comment telling me which base model you’d prefer.
FAQ
Comments (3)
In the next version, can you add the trigger/tags you included in the training data?
Thanks this looks promising.
Thanks for the feedback. I’m thinking of adding the character names into the captions, but because of how many there are I don’t think a small LoRA model will learn them all — there are more than 100. Also, the Anima captions were in SD tags format with the trigger @rudefrog, while the Flux ones were in natural English and I didn’t use any special tags, I can add some so that certain concepts from the dataset become easier to use.
Can you add your prompts to the example images? That would help with prompting for this Lora. Thanks.