FeiHou_MiniMax-H3_Remix_v0.6_int8_convrot_v2.safetensors
更新2026-09-14 01:22发布时间2026-09-01 17:31
33
0
0
FeiHou_MiniMax-H3_Remix_v0.6_int8_convrot_v2.safetensors - 1
头像
Lorena
类型
UNet
基础模型
MiniMax H3
发布时间
2026-09-01 17:31
文件签名
0382a35e7c20c8afeaaacb3a7f6d47a1a05045179804e57508f5d068152b200a
FeiHou_MiniMax-H3_Remix_v0.6_int8_convrot_v2.safetensors
31.67 GB

base_model:

  • MiniMax-H3 tags:
  • minimax-h3
  • text-to-video
  • image-to-video
  • ref2va
  • comfyui
  • video-generation library_name: comfyui license: other

MiniMax-H3 Remix

MiniMax-H3 Remix is a custom dual-model H3 remix checkpoint built for ComfyUI video generation.

It uses a hierarchical recombination workflow instead of naive weight averaging: the two H3 branches are reorganized at targeted response blocks, with key modulation layers such as AdaLN handled separately. The goal is to retain stable generation behavior while introducing stronger visual detail, dynamic expression, and refined stylistic response.

LoRA contributions used during remix creation are statically fused into the checkpoint. No corresponding external style LoRAs are required for normal inference.

Available variants

VariantFilePurpose
BF16FeiHou_MiniMax-H3_Remix_v0.6.safetensorsFull-precision, quality-first release
INT8 ConvRotFeiHou_MiniMax-H3_Remix_v0.6_int8_convrot_v2.safetensorsComfyUI-native low-VRAM deployment

The INT8 ConvRot version is a quantized deployment variant of the same Remix design. For maximum fidelity, use BF16 when hardware allows.

Installation

  1. Place the .safetensors file in:

    ComfyUI/models/diffusion_models/
    
  2. Load it with ComfyUI-FeiHou-Easy-H3.

  3. Use a MiniMax H3-compatible workflow.

Recommended starting settings

  • Sampler: res_multistep
  • Scheduler: simple
  • Steps: 8-12
  • Optional Turbo LoRA strength: 0.75

These are starting points rather than fixed requirements. Adjust according to resolution, motion intensity, reference-image strength, and your intended visual style.

Notes

  • This is a remixed checkpoint, not an officially trained MiniMax H3 base release.
  • Avoid stacking the same LoRAs that were already baked into this model; excessive strength may cause oversaturation, unstable motion, or distorted details.
  • INT8 reduces memory use, but BF16 remains the preferred version for demanding quality evaluation.
  • Please follow the original MiniMax H3 license and all applicable local laws when using or redistributing this model.

Credits

  • Remix design and release: FX-FeiHou
  • Built for the MiniMax H3 / ComfyUI ecosystem
  • Recommended runtime nodes: ComfyUI-FeiHou-Easy-H3

Links

作品

暂无作品