<h1 id="fasth3-omniref-8-step-gguf-and-convrot-quantizations">FastH3 OmniRef 8-Step — GGUF &amp; ConvRot Quantizations</h1>Unpruned, full-weight quantizations of FastVideo's FastH3 Omni 8-Step V1 reference-to-video model, prepared by <a target="_blank" rel="ugc" href="https://huggingface.co/realrebelai">Rebel AI</a> for ComfyUI experimentation.<a target="_blank" rel="ugc" href="https://huggingface.co/FastVideo/FastVideo-FastH3-Omni-8-Step-V1">Original FastVideo model</a> · <a target="_blank" rel="ugc" href="https://huggingface.co/FastVideo/FastH3-Omni-Comfy">Official ComfyUI model release</a> · <a target="_blank" rel="ugc" href="https://github.com/comfyanonymous/ComfyUI">ComfyUI</a> · <a target="_blank" rel="ugc" href="https://github.com/city96/ComfyUI-GGUF">ComfyUI-GGUF</a><blockquote>Full weights, not the pruned release. These conversions were made from the upstream 14-shard <code>transformer_ref</code> BF16 checkpoint, not the smaller, pruned ComfyUI checkpoint. Quantization reduces numerical precision; it does not remove transformer blocks, prune AdaLN parameters, or change the underlying model into a different FastH3 release.</blockquote><h3 id="ggufs:-https:huggingface.corealrebelaifasth3-omni_ggufs">GGUFS: <a target="_blank" rel="ugc" href="https://huggingface.co/realrebelai/FastH3-Omni_GGUFs">https://huggingface.co/realrebelai/FastH3-Omni_GGUFs</a></h3><h3></h3><h2 id="about-the-model">About the model</h2><a target="_blank" rel="ugc" href="https://huggingface.co/FastVideo/FastVideo-FastH3-Omni-8-Step-V1">FastVideo FastH3 Omni 8-Step V1</a> is an eight-evaluation distilled derivative of MiniMax-H3, designed to generate synchronized video and audio from text and ordered image, video, and audio references. It is a reference-conditioned (OmniRef / Ref2VA) model, not the older FastH3 text-to-video / first-and-last-frame-only checkpoint.<ul><li>Generation: video with synchronized audio</li><li>Conditioning: text plus ordered image/video/audio references</li><li>Sampling: 8 transformer evaluations, guidance 1.0 (per upstream guidance)</li><li>Architecture: 50 main transformer blocks and 2 token-refiner blocks in the converted checkpoint</li><li>Attention: preserves FastH3 gate-compression / sparse-attention weights</li><li>Source precision: BF16</li></ul><h2 id="model-files-and-formats">Model files and formats</h2>Download one diffusion-model format according to your memory budget and the loader available in your ComfyUI installation. The table documents the intended file naming; check the repository's Files tab for which variants have actually been uploaded.FileFormatNotes<code>FastH3-OmniRef-Unpruned-Q2_K.gguf</code>GGUF Q2_KMaximum compression among the planned GGUF choices; greatest risk of quality loss.<code>FastH3-OmniRef-Unpruned-Q3_K_S.gguf</code>GGUF Q3_K_SCompact GGUF option.<code>FastH3-OmniRef-Unpruned-Q4_K_M.gguf</code>GGUF Q4_K_MHigher-precision GGUF option; a practical first comparison.<code>FastH3-OmniRef-Unpruned-INT8-ConvRot.safetensors</code>INT8 ConvRotComfyUI/kitchen quantized tensor format.<code>FastH3-OmniRef-Unpruned-W4A8-ConvRot.safetensors</code>W4A8 ConvRot4-bit codebook weights / INT8 activation-oriented ComfyUI format, with INT8 fallback for ineligible layers.GGUF and ConvRot are different formats. GGUF files need a compatible GGUF diffusion-model loader; INT8/W4A8 safetensors use ComfyUI's supported quantized-weight loading paths. Do not load GGUF using the standard safetensors diffusion loader.Memory note: Smaller checkpoint size does not mean the entire generation fits in that amount of VRAM. The Qwen3-VL text encoder, video VAE, audio VAE, attention workspace, conditioning inputs, and CPU offload can require substantial extra memory.<h2 id="how-these-weights-were-prepared">How these weights were prepared</h2>The original FastVideo 14-shard BF16 <code>transformer_ref</code> checkpoint was converted into a single ComfyUI-style safetensors file using bounded-memory, shard-aware conversion:<ol><li>Read all 14 safetensors shard headers and verified 688 source tensors.</li><li>Fused 52 attention Q/K/V groups, including groups crossing shard boundaries.</li><li>Applied 52 SwiGLU row-half reorderings required by the ComfyUI MiniMax-H3 layout.</li><li>Preserved the reference-conditioning and sparse-attention parameters, including <code>to_gate_compress</code> weights.</li><li>Wrote an unpruned 70.14 GB BF16 intermediate containing 585 tensors.</li><li>Quantized from this BF16 intermediate rather than requantizing an already-compressed model.</li></ol><h3 id="w4a8-conversion-report">W4A8 conversion report</h3>The completed W4A8 ConvRot run reported:MeasurementResultQuantized layers308Output tensors1,717Mean relative weight-reconstruction error6.272%Minimum / maximum reported error0.855% / 7.346%Quantization runtime482.5 secondsThese are weight-space reconstruction measurements, not VBench scores, perceptual quality scores, or proof of successful ComfyUI inference. Visual and audio comparison against BF16 remains necessary.<h2 id="comfyui-setup">ComfyUI setup</h2><h3 id="1.-update-comfyui">1. Update ComfyUI</h3>Use an up-to-date ComfyUI version with native MiniMax-H3 / FastH3 support (FastH3 workflows require ComfyUI 0.36.0+ according to ComfyUI's documentation).<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/FastVideo/FastH3-Omni-Comfy">Official FastH3 Omni ComfyUI model card and setup</a></li><li><a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/docs/blob/main/tutorials/video/minimax/minimax-h3.mdx">ComfyUI MiniMax-H3 documentation</a></li><li><a target="_blank" rel="ugc" href="https://github.com/city96/ComfyUI-GGUF">ComfyUI-GGUF</a> (for GGUF variants; MiniMax-H3 architecture support must be verified)</li></ul><h3 id="2.-install-the-diffusion-checkpoint">2. Install the diffusion checkpoint</h3>For ConvRot safetensors, put the selected diffusion model in:<pre><code>ComfyUI/models/diffusion_models/
</code></pre>For GGUF, install a GGUF loader with MiniMax-H3 (<code>minimax_h3</code>) architecture support and put the <code>.gguf</code> file in the directory your GGUF loader scans (commonly <code>ComfyUI/models/unet/</code>). A generic GGUF loader may reject the architecture; this is a loader-compatibility issue, not evidence that the model must be re-quantized.<h3 id="3.-install-the-other-required-components">3. Install the other required components</h3>These are not included in the diffusion checkpoint. Use matching MiniMax-H3 components from the <a target="_blank" rel="ugc" href="https://huggingface.co/FastVideo/FastH3-Omni-Comfy">official ComfyUI FastH3 Omni release</a>:<pre><code>ComfyUI/models/
├── text_encoders/
│ └── qwen3vl_32b_minimax_h3_int8_convrot.safetensors (example)
└── vae/
 ├── minimax_h3_video_vae_fp16.safetensors
 └── minimax_h3_audio_vae_fp32.safetensors
</code></pre>The text encoder needs its vision functionality for image/video reference conditioning. Alternative official encoder and video-VAE precision variants are available; choose ones compatible with your hardware. The NVFP4 AWQ encoder variant requires appropriate Blackwell hardware and is not a universal low-VRAM option.<h3 id="4.-use-a-reference-to-video-workflow">4. Use a reference-to-video workflow</h3>Start with ComfyUI's MiniMax-H3 Reference to Video / FastH3 Omni workflow, replacing its diffusion-model selection with the appropriate loader for your chosen format.Upstream FastH3 Omni guidance includes:<ul><li>8 steps, guidance 1.0</li><li>Video shift 12.0, audio shift 3.0 in the published ComfyUI workflow</li><li>Reference tags such as <code>&lt;Picture 1&gt;</code>, <code>&lt;Video 1&gt;</code>, and <code>&lt;Audio 1&gt;</code> in the prompt</li><li>A suitable sparse-attention implementation and attention settings matching the official workflow</li></ul>Do not assume that replacing a pruned checkpoint with the unpruned one is automatically plug-and-play: checkpoint parameterization and loader handling should be checked. Native and GGUF loader compatibility have not been established by the conversion logs alone.<h2 id="troubleshooting">Troubleshooting</h2><ul><li><code>Unexpected architecture type ... minimax_h3</code>: your GGUF loader lacks support for MiniMax-H3. Use a compatible version or integration.</li><li>Missing weights / size mismatch: confirm you're using a loader that supports the unpruned form rather than requiring the pruned AdaLN table layout.</li><li>Reference image ignored: verify the Qwen3-VL encoder's vision path and the reference conditioning nodes; do not use a text-only encoder configuration.</li><li>Out of memory: model weight size is only part of peak memory. Try lower resolution/frame count, offloading, suitable encoder/VAE quantization, or a smaller model variant.</li><li>Different quality from upstream: quantization changes numerical precision, and different inference kernels/workflow settings can change the result.</li></ul><h2 id="credits-and-attribution">Credits and attribution</h2><ul><li>Original model and distillation: <a target="_blank" rel="ugc" href="https://huggingface.co/FastVideo">FastVideo</a>, based on <a target="_blank" rel="ugc" href="https://huggingface.co/MiniMaxAI/MiniMax-H3">MiniMax-H3</a>.</li><li>FastH3 Omni collaboration: <a target="_blank" rel="ugc" href="https://nuvalab.ai/">Nuva Lab</a> and FastVideo. Upstream also thanks the NVIDIA FastGen team and Reactor for compute support.</li><li>ComfyUI integration: <a target="_blank" rel="ugc" href="https://github.com/comfyanonymous/ComfyUI">ComfyUI</a> and the FastVideo/ComfyUI maintainers.</li><li>Community conversions and quantization: <a target="_blank" rel="ugc" href="https://huggingface.co/realrebelai">Rebel AI / @realrebelai</a>.</li></ul>This is an unofficial community quantization. It is not an official FastVideo, MiniMax, Nuva Lab, or ComfyUI release.<h2 id="license-read-before-use-or-redistribution">License — read before use or redistribution</h2>The original weights are governed by the <a target="_blank" rel="ugc" href="https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE">MiniMax H3 Community License Agreement</a>, which also governs derivatives and redistribution. This is not an unrestricted open-source license.Important territorial restriction: The published agreement defines its applicable territory as excluding the United States, European Union, United Kingdom, and Republic of Korea. It also includes terms concerning redistribution, notices, commercial usage, and acceptable use. Review the agreement and obtain any necessary permission before using or redistributing these model weights, especially from or into an excluded territory. A precision conversion does not remove the upstream licensing requirements.Distributions under the agreement may require accompanying <code>LICENSE</code> and <code>NOTICE</code> files; see its Section III for the exact requirements. The upstream agreement's required NOTICE wording is available in the license text.<hr />Rebel AI: <a target="_blank" rel="ugc" href="https://huggingface.co/realrebelai">Hugging Face</a> · <a target="_blank" rel="ugc" href="https://github.com/RealRebelAI">GitHub</a> · <a target="_blank" rel="ugc" href="https://www.youtube.com/@RealRebelAI">YouTube</a> · <a target="_blank" rel="ugc" href="https://x.com/realrebelai">X</a>

Files with SHA256: e900b0a031bbd680878a045eeb47f5b14643292ec3ef4e79a137433e4fb5461d