WIP - Play with it and share the results would like to see what people come up with.
Updates to come. Send me some buzz if you feel inclined.
This LoRA is designed to generate a realistic rendition of Ginny, a character inspired by Ishuzoku Reviewers, featuring cow horns and huge breasts. It’s compatible with both HunyuanVideo Video-to-Video and Image-to-Video workflows. This model was trained on 1092 generated realistic images at 1024 x 1440 of Ginny using PonyXL and a Ginny Pony Lora. It's undertrained but I've posted this due to request as work in progress.
Works with Bouncing Breasts HunyuanVideo lora. But expect a pretty big image quality drop and since my Lora is under trained it can get overpowered it seems.
""\_(:3)_/""
Kling not required.
How to Use:
Load the LoRA in your preferred Hunyuan workflow (V2V or I2V). ( Some of my earlier posts contain workflows)
Load any short clip of a woman with large breasts. Preferably with lots of bounce. Resolution I've been resizing to is 800 x 1120.
Use the trigger prompt:
"Ginny, a woman with cow horns and huge breasts".Add your scene description and adjust CFG settings as needed.
Example prompt style:
"Ginny, a woman with cow horns and huge breasts, leaning forward as her breasts sway gently the cinematic lighting highlights the motion of her breasts" Or something like that...keep it under 248 tokens.
Most of the generation are 97 frames at 800 x 1120 CFG 1 25 steps.. With a 24gb card you can push it to 129 frames but you'll need to swap all the blocks Double Block swap 20 and Single Block 40 to save memory. Generations take 45 min each.
Quick runs, Steps 10-15, CFG 1, 49 frames. takes about 7 minutes. The higher the inference steps the more the model focuses visual features of the video defined by how words interact within the prompt, influencing semantic comprehension. Which I suspect leads to better quality and more movement but at the cost of memory.
CFG Settings:
- CFG Scale: 1
- CFG values in this range significantly improve animation quality and coherence. Higher values require significantly more VRam and tend to over-constrain the model, leading to stiff or unnatural motion. If generating smaller resolutions like 544 x 720 try CFG 2 and 49 frames.
Negative Prompt:
low quality, bad hands, bad teeth, bad eyes, bad limbs, distortion, talking, speaking
Training Insights
Dataset Size & Quality
Small Datasets (<100 images):
Results were okay but lacked consistency in breast physics and overall realism. Prompt styles varied, but the model struggled to generalize.
Large Dataset (~1000 images):
Best results so far came from a large, varied dataset with interleaved detailed prompts.
The diversity in the larger dataset seemed to help the model learn better motion dynamics and anatomical details.
Getting the breast bounce right, especially with larger breasts, is tough. The model often exaggerates or underplays the motion.
A possible solution is to pretrain on videos I'm experimenting with pretraining on videos featuring large bouncing breasts to teach the model realistic physics.
Small Dataset Integration: After pretraining, I’ll fine-tune with a smaller, curated dataset of Ginny-specific images.
Pending Results: Depending on how this goes, I may expand the dataset further.
Next Steps
Pretraining on Videos: Focus on breast physics and natural motion.
Fine-Tuning: Incorporate a smaller, high-quality dataset of Ginny-specific images.
This is a work in progress, and I’d love your input! Please test the LoRA and share your results. What works? What doesn’t? If you got a better approach for breast physics let me know.
Updates to come.