{"id":12700,"date":"2026-11-17T11:01:31","date_gmt":"2026-11-17T02:01:31","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=12700"},"modified":"2026-11-17T16:11:38","modified_gmt":"2026-11-17T07:11:38","slug":"fasth3-omniref-comfy-released","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/fasth3-omniref-comfy-released\/","title":{"rendered":"FastH3-OmniRef-Comfy: File List"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Repository<\/td>\n<td><a href=\"https:\/\/huggingface.co\/FastVideo\/FastH3-OmniRef-Comfy\">FastVideo\/FastH3-OmniRef-Comfy<\/a><\/td>\n<\/tr>\n<tr>\n<td>Family guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-minimaxai-minimax-h3-en\/\">MiniMax-H3 guide (14 articles)<\/a><\/td>\n<\/tr>\n<tr>\n<td>Publisher guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-minimax-en\/\">MiniMax: models and licenses<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-20<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>minimax-h3-community-license-agreement<\/td>\n<\/tr>\n<tr>\n<td>Formats<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">safetensors<\/a><\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>FastVideo has released <code>FastVideo\/FastH3-OmniRef-Comfy<\/code>, a native ComfyUI package for &#8220;FastH3 OmniRef&#8221; that generates synchronized video and audio from text prompts and multimodal reference materials (images, videos, and audio).<\/p>\n<p>This model is a private release candidate derived from the official checkpoint of the FastH3 OmniRef student model, which operates in 8 forwards (8 steps). It adopts V2-format ComfyUI packing technology and is provided as a single diffusion model file. It is configured to be used in combination with a shared Qwen3-VL text encoder and respective video and audio VAEs.<\/p>\n<h2>Specifications<\/h2>\n<p>Specifications confirmed from published materials and information on the original model (MiniMax-H3) are as follows:<\/p>\n<ul>\n<li><strong>Input Conditions:<\/strong> Text prompts, along with reference materials including images (up to 9), videos (up to 3, 2\u201315 seconds each), and audio (up to 3, 2\u201315 seconds each). Supports a total of up to 12 reference files. Designated tags are used within the prompt to identify references.<\/li>\n<li><strong>Architecture:<\/strong> Retains all 50 transformer blocks, 2 token refiner blocks, and 32 PDD heads. Employs a rank-8, 1025-point AdaLN timestep curve.<\/li>\n<li><strong>Inference Settings:<\/strong> 8 forwards (8 steps), guidance 1, video shift 12, audio shift 3<\/li>\n<li><strong>Related Components:<\/strong><\/li>\n<li>Text Encoder: Based on Qwen3-VL 32B (includes vision tower for reference conditioning) &#8211; Video VAE: 16x spatial and 4x temporal compression, 24 latent channels. Includes reference encoder and 36-layer decoder &#8211; Audio VAE: Supports 32 kHz stereo (compresses left and right channels independently into 40 Hz latent tokens)<\/li>\n<li><strong>Original Model (MiniMax-H3) Output Specifications:<\/strong><\/li>\n<li>Output Content: Synchronized video and 32 kHz stereo audio &#8211; Resolution: Default 768p short side (supports 2K generation via API-based H3-Regenerate-2K, etc.) &#8211; Output Duration: 4\u201315 seconds &#8211; Frame Rate: 24 FPS &#8211; Supported Aspect Ratios: Supports a wide variety of aspect ratios including 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16 &#8211; Supported Languages: Stable support for 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish<\/li>\n<li><strong>License:<\/strong> The model weights are subject to the &#8220;MiniMax-H3 Community License&#8221;. Note that the included ComfyUI-derived conversion code and upstream tools are under the GPL-3.0 license<\/li>\n<\/ul>\n<h2>Performance and Quality<\/h2>\n<p>Regarding the ComfyUI version of the FastH3 OmniRef checkpoint provided in this repository, the model card explicitly states that inference behavior on GPUs, quantization kernel equivalence, and end-to-end final output quality are <strong>unverified<\/strong>.<\/p>\n<p>CPU regression tests to confirm the accuracy of the conversion process (6 items including strict model loading, QKV order, SwiGLU output, and PDD reconstruction) have passed. For the newly converted BF16 curve fit, it is reported that the relative residual is 0.0001013, with a maximum relative L2 error of 0.0003073 across 129 held-out timesteps and sampling per modulation projection. However, these are numerical modulation verification results and do not represent actual video quality scores.<\/p>\n<p>Additionally, the execution environment in Native ComfyUI has several behavioral differences from the original FastVideo inference protocol (such as unit grid consolidation in PDD and using 64-token video cubes for 128-token tiling). Therefore, users are cautioned against applying the original FastVideo benchmark figures directly to this package.<\/p>\n<p>Meanwhile, the official card for the original model, MiniMax-H3, provides reproduction scripts and generated video files for text-to-video with audio (T2VA), first\/last frame specified generation (FL2VA), and multimodal reference generation (Ref2VA) cases, showing that high-quality synchronized audio-visual content can be generated under diverse input conditions.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>This model is suitable for applications where multiple media types\u2014such as images, videos, and audio\u2014are combined as reference materials alongside text prompts to generate synchronized, high-quality video and audio.<\/p>\n<p>Specifically, users can feed up to 12 reference files freely combining up to 9 images, up to 3 videos (2\u201315 seconds each), and up to 3 audio clips (2\u201315 seconds each). By writing identification tags such as <code>&lt;image_1&gt;<\/code>, <code>&lt;video_1&gt;<\/code>, and <code>&lt;audio_1&gt;<\/code> within the text prompt, it is possible to generate videos that flexibly control specific image or video characters, backgrounds, compositions, and audio elements.<\/p>\n<p>Furthermore, backed by the excellent multimodal comprehension and synchronized audio generation capabilities of the original model MiniMax-H3, it also supports creating videos with stable conversational audio in 11 languages, including Japanese. Because it operates in a mere 8 steps of inference (8 forwards), it is suited for creators and researchers looking to experiment with advanced reference specifications while reducing inference time. Additionally, since it supports the Native ComfyUI <code>MiniMaxH3ReferenceToVideo<\/code> node, it can be easily integrated into visual workflows on ComfyUI.<\/p>\n<h2>How It Differs from Similar Models<\/h2>\n<p>The <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/10\/06\/fasth3-v2-fp8-nvfp4-released\/\">FP8 and NVFP4 versions of the audio-visual video generation model &#8220;FastH3 V2&#8221;<\/a> previously introduced on our site were models that emphasized lightweight quantization (NVFP4\/FP8) and high-speed single-GPU inference mainly for text-to-video generation. In contrast, this &#8220;FastH3 OmniRef&#8221; differs decisively in that it specializes in advanced control functions (Reference-to-Video) utilizing multimodal reference materials including images, videos, and audio alongside text.<\/p>\n<p>Also, in the <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/10\/06\/fasth3-trim-8-step\/\">8-step version of the audio-visual video generation model &#8220;FastH3 Trim&#8221; with reduced transformer blocks<\/a>, it was explicitly stated that a trade-off with quality occurs. On the other hand, the FastH3 OmniRef provided in this package is applied to the ComfyUI environment while maintaining all 50 transformer blocks and 32 PDD heads, featuring a design that minimizes any loss of the original model&#8217;s expressive power.<\/p>\n<p><!-- lmw:files --><\/p>\n<h2>Distributed Files<\/h2>\n<p><em>Weight files published in <a href=\"https:\/\/huggingface.co\/FastVideo\/FastH3-OmniRef-Comfy\/tree\/main\">FastVideo\/FastH3-OmniRef-Comfy<\/a>, listed by this site from the Hugging Face API. Sizes are the actual file sizes.<\/em><\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>File<\/th>\n<th>Size<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>diffusion_models\/fasth3_omniref_8step_pruned_bf16.safetensors<\/code><\/td>\n<td>44.24GB<\/td>\n<\/tr>\n<tr>\n<td><code>diffusion_models\/fasth3_omniref_8step_pruned_int8_convrot.safetensors<\/code><\/td>\n<td>22.21GB<\/td>\n<\/tr>\n<tr>\n<td><code>text_encoders\/qwen3vl_32b_minimax_h3_bf16.safetensors<\/code><\/td>\n<td>51.51GB<\/td>\n<\/tr>\n<tr>\n<td><code>text_encoders\/qwen3vl_32b_minimax_h3_int8_convrot.safetensors<\/code><\/td>\n<td>27.14GB<\/td>\n<\/tr>\n<tr>\n<td><code>text_encoders\/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors<\/code><\/td>\n<td>15.69GB<\/td>\n<\/tr>\n<tr>\n<td><code>vae\/minimax_h3_audio_vae_fp32.safetensors<\/code><\/td>\n<td>605MB<\/td>\n<\/tr>\n<tr>\n<td><code>vae\/minimax_h3_video_vae_fp16.safetensors<\/code><\/td>\n<td>5.21GB<\/td>\n<\/tr>\n<tr>\n<td><code>vae\/minimax_h3_video_vae_int8_convrot.safetensors<\/code><\/td>\n<td>2.81GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><!-- \/lmw:files --><\/p>\n<h2>How to Get It<\/h2>\n<p>The model checkpoint and related files are distributed in <code>safetensors<\/code> format in the Hugging Face <code>FastVideo\/FastH3-OmniRef-Comfy<\/code> repository.<\/p>\n<p>Acquired files are placed and used in the following ComfyUI directories:<\/p>\n<ul>\n<li>Diffusion Models (<code>ComfyUI\/models\/diffusion_models\/<\/code>):<\/li>\n<li><code>fasth3_omniref_8step_pruned_bf16.safetensors<\/code> (BF16 version) &#8211; <code>fasth3_omniref_8step_pruned_int8_convrot.safetensors<\/code> (INT8 ConvRot version)<\/li>\n<li>Text Encoders (<code>ComfyUI\/models\/text_encoders\/<\/code>):<\/li>\n<li><code>qwen3vl_32b_minimax_h3_bf16.safetensors<\/code> &#8211; <code>qwen3vl_32b_minimax_h3_int8_convrot.safetensors<\/code> &#8211; <code>qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors<\/code><\/li>\n<li>VAE (<code>ComfyUI\/models\/vae\/<\/code>):<\/li>\n<li><code>minimax_h3_video_vae_fp16.safetensors<\/code> (or <code>int8_convrot<\/code> version) &#8211; <code>minimax_h3_audio_vae_fp32.safetensors<\/code> (for audio)<\/li>\n<\/ul>\n<p>When running, select one each from the diffusion model, text encoder, and video VAE, and use them in combination with the audio VAE. Connect the <code>MiniMaxH3ReferenceToVideo<\/code> conditioning node standardly equipped in Native ComfyUI, specify reference tags within the prompt, and perform inference.<\/p>\n<p>Note that diffusion model checkpoints in FP8 and NVFP4-QAD formats are not included in this package at this time due to a lack of runtime support on the Native ComfyUI side to preserve numerical behavior.<\/p>\n<p><!-- lmw:runnability --><\/p>\n<h2>Can You Run It Locally?<\/h2>\n<p>The publisher distributes this model as safetensors.<\/p>\n<p><strong>License \u2014 <code>minimax-h3-community-license-agreement<\/code> (Commercial use prohibited):<\/strong> MiniMax&#8217;s MiniMax H3 Community License. Use of the weights is limited to the EU, the UK, South Korea and the US (Sections II and V.4); elsewhere, including Japan, a separate license is required.<\/p>\n<p><em>Compiled by this site&#8217;s code from the published formats and the license field. License summaries are not legal advice \u2014 check the publisher&#8217;s original terms before relying on them.<\/em><\/p>\n<p><!-- \/lmw:runnability --><\/p>\n<p><!-- lmw:lab --><\/p>\n<h2>Our Own Measurements<\/h2>\n<h3>Measurements We Did Not Take<\/h3>\n<p>This model&#8217;s license (<code>minimax-h3-community-license-agreement<\/code>) does not permit commercial use. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).<\/p>\n<p><!-- \/lmw:lab --><\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/fasth3-omni-comfy-released\/\">FastH3-Omni-Comfy: 80GB+ VRAM, File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/fasth3-omni-fp8-comfy-released-2\/\">FastH3-Omni-FP8-Comfy Audio-Synced Video Generation Model: 32GB+ VRAM<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/fasth3-omni-nvfp4-qad-comfy-released\/\">FastH3-Omni-NVFP4-QAD-Comfy: 16GB+ VRAM, File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/fastvideo-fasth3-trim-comfy-released\/\">FastVideo-FastH3-Trim-Comfy Text-to-Video and Audio Model: File List<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>Explore the same model family<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-minimaxai-minimax-h3-en\/\">MiniMax-H3 family overview (14 articles, 4 converted builds)<\/a><\/li>\n<li><strong>Formats this model is available in<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">Safetensors format guide and models<\/a><\/li>\n<li><strong>Learn about the publisher<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-minimax-en\/\">MiniMax: models, licenses and articles<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/FastVideo\/FastH3-OmniRef-Comfy\">https:\/\/huggingface.co\/FastVideo\/FastH3-OmniRef-Comfy<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/MiniMaxAI\/MiniMax-H3\">https:\/\/huggingface.co\/MiniMaxAI\/MiniMax-H3<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>FastVideo has released FastH3-OmniRef-Comfy, a native ComfyUI package for generating synchronized video and audio from text and multimodal references.<\/p>\n","protected":false},"author":1,"featured_media":12699,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[203],"tags":[677,3804,3031,730,1547,959,213],"class_list":["post-12700","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-image-video-and-audio","tag-comfyui-en","tag-fasth3-omniref-en","tag-fastvideo-en","tag-minimax-h3-en","tag-verified","tag--en"],"lang":"en","translations":{"en":12700,"ja":12698},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12700","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=12700"}],"version-history":[{"count":3,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12700\/revisions"}],"predecessor-version":[{"id":18247,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12700\/revisions\/18247"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/12699"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=12700"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=12700"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=12700"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}