minimax_h3_ref2va_XUELUO_int8_convrot

minimax_h3_ref2va_XUELUO_int8_convrot is a video generation model deeply optimized with Layer Merging, built upon the INT8 version of the MiniMax-H3 model library.

Credits: Special thanks to @雪落眉间路已远 for providing the core technical approach. He is currently using this H3 model to create AI short dramas on Bilibili. Check out his work here: Bilibili Space.

This model addresses the key pain points of original video generation models regarding image quality and stability. By combining the visual detail advantages of fl2va with the structural stability of ref2va, it delivers an enhanced Reference-to-Video experience.

🚀 Workflow (Civitai):

• Model-Configured Dual-Sampling Workflow: Minimax_H3_ref2va_HQ+Dual-Sampling+Workflow+(New) - v1.0 | Hailuo H3 by MiniMax Workflows | Civitai

🔥 Core Improvements & Features

• Cross-Model Layer Merging: Extensive comparisons showed that fl2va significantly outperforms ref2va in visual detail. This model precisely extracts the core feature layers (Layers 25–49) from fl2va and seamlessly merges them into the ref2va architecture.

• Dual Breakthrough in Quality & Stability: Combines the high resolution and delicate visuals of fl2va with the temporal consistency and subject stability of ref2va, drastically reducing artifacting and distortion.

• Low-Step (Fast LoRA) Compatibility: Demonstrates strong robustness, maintaining extremely stable video output even when paired with a 4-Step acceleration LoRA for high-speed rendering.

• INT8 Quantization & Lightweight Execution: Built on INT8 quantization, it significantly reduces VRAM requirements while preserving expressive capability, making it well-suited for consumer GPUs.

💡 Recommended Settings

ParameterSettingBase ModelMiniMax-H3 INT8Generation ModeRef2VA / Reference-to-VideoSampling Steps4 – 20 steps (Supports 4-step LoRA ultra-fast rendering)Best Used ForPairing with high-detail reference images/videos for high-quality, dynamically stable video generation

🙏 Acknowledgements & References

Special thanks to the original model creators and the ComfyUI community for their foundational work:

• MiniMax-H3 Base Model: MiniMaxAI/MiniMax-H3

• Comfy Implementation: Comfy-Org/MiniMax-H3