FastH3 OmniRef · 8-step · pruned · W4A8

A 4-bit (W4A8, asymmetric int8 activation) quantization of the FastH3 OmniRef 8-step pruned ref2va checkpoint, for ComfyUI.

What it is

  • Quantized from: the FastH3 OmniRef 8-step pruned bf16 file in FastVideo/FastH3-Omni-Comfy (44,240,780,704 bytes), sha256 c0fac98c0317eb85215bd186b250be9b113675baab1323add40f26262f552ed9.
  • Conversion receipt source: FastVideo/FastVideo-FastH3-Omni-8-Step-V1, revision 88d512c5e7cc38dda499fbb65688ce58611f9c80.
  • Base model: MiniMaxAI/MiniMax-H3 at revision 9bfb6693f2cf6de171db46d1aa586f67d773a1da.
  • This file: fasth3_omniref_8step_pruned_w4a8.safetensors, 16,555,980,952 bytes, sha256 76753f3265c52d5675efc1306278f950cdd3ed2f03b841e6c4de0b221cde6553.

Inference settings

  • Steps: 8. CFG (guidance scale): 1.0 (no CFG pass).
  • Sigma shift: video 12.0, audio 3.0.
  • Conditioning: ref2va (fixed ordered references, target-only flow). Routes to the native reference-to-video graph.
  • No turbo / tier LoRA: the checkpoint carries its own 8-step distillation.

Validation

  • Quantization validator: 423 passed, 0 failed, 1 skipped (validated with comfy-kitchen 0.2.39).

Usage (ComfyUI)

Place the file in ComfyUI/models/diffusion_models/. It uses the AsymW4A8Int8Layout quantization layout of comfy-kitchen; the validation above used comfy-kitchen 0.2.39.

Credit

License

Inherits the MiniMax-H3 Community License Agreement, see the base model page.

Downloads last month
393
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for berryber09/FastH3-OmniRef-8step-pruned-w4a8

Finetuned
(204)
this model

Collection including berryber09/FastH3-OmniRef-8step-pruned-w4a8