<h1 id="minimax-h3-long-video-workflows-v3.0">MiniMax H3 Long Video Workflows — V3.0</h1>V3.0 includes two six-section MiniMax H3 long-video workflows and the <code>h3_direct_latent_head</code> custom node.<ul><li>FL2VA — supports T2V, I2V, last-frame guidance, and first/last-frame generation.</li><li>REF2VA — supports up to four optional image references across S1–S6 and HQ refinement.</li></ul><h2 id="main-features">Main features</h2><ul><li>Direct transfer of video and audio latents between sections.</li><li>Exact phase-aligned 22-frame protected AV context.</li><li>Accurate 24 FPS video / 40 Hz audio timeline.</li><li>Optional audio history: 0F, 24F, 48F, or 72F. Recommended: 48F / 2 seconds.</li><li>Final audio is assembled cumulatively, decoded once, and trimmed to the exact video duration.</li><li>Automatic section-length and final-duration calculation.</li><li>H3 denoise-mask compatibility fix based on <a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/ComfyUI/pull/15988">ComfyUI PR #15988</a>.</li><li>MASTER node for resolution, aspect ratio, quality, continuity, section length, audio history, and S1–S6 prompts.</li><li>Individual section checkpoints and final assembled video.</li><li>Separate FL2VA and REF2VA prompt-writing guides.</li></ul><h2 id="quality-modes">Quality modes</h2><ul><li>NATIVE — generation without latent upscaling or HQ refinement.</li><li>DIRECT — learned latent upscale without diffusion redraw. Recommended for maximum stability.</li><li>HQ REFINE — learned latent upscale followed by three-step high-resolution refinement. This can improve detail but may slightly redraw some frames.</li></ul>Upscaling and HQ refinement do not regenerate the audio.Turbo LoRA affects S1–S6 and HQ refinement. The secondary motion/style LoRA affects only S1–S6.<h2 id="video-length">Video length</h2>Valid section lengths follow this grid:<code>section_frames = 5 + 17 × N</code>Minimum value: <code>39</code>.For six sections:<code>final_frames = 6 × section_frames − 110</code>The first section keeps its full length. Each following section contributes <code>section_frames − 22</code> new frames.<h2 id="models">Models</h2>MiniMax H3 models:<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main">Comfy-Org/MiniMax-H3</a>Heretic Text Encoder — choose one:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot/tree/main">INT8 ConvRot</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4/tree/main">NVFP4</a></li></ul>LoRA sources:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/DarkRomeo88/MiniMax-H3-turbo-lora-comfyui/tree/main">DarkRomeo88 MiniMax H3 Turbo LoRA</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/rzgar/minimax-h3_fl2v_8Step_motion_enhancer/tree/main">FL2VA 8-Step Motion Enhancer</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras">Kijai MiniMax H3 LoRAs</a></li></ul>Learned latent upscaler:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/tree/main">Upscaler model</a></li><li><a target="_blank" rel="ugc" href="https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler">ComfyUI upscaler node</a></li></ul><h2 id="required-custom-nodes">Required custom nodes</h2><ul><li><a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">ComfyUI-KJNodes</a></li><li><a target="_blank" rel="ugc" href="https://github.com/Saganaki22/ComfyUI-sol-attn">ComfyUI-sol-attn</a></li><li><a target="_blank" rel="ugc" href="https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite">ComfyUI-VideoHelperSuite</a></li><li><a target="_blank" rel="ugc" href="https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler">ComfyUI MiniMax H3 Latent Upscaler</a></li><li>h3_direct_latent_head — included in the V3.0 archive.</li></ul>The external latent-upscaler node must be installed for the workflows to load, even when using NATIVE mode.<h2 id="installation">Installation</h2><ol><li>Stop ComfyUI.</li><li>Delete the previous <code>ComfyUI/custom_nodes/h3_direct_latent_head</code> folder.</li><li>Copy <code>h3_direct_latent_head</code> from the V3.0 archive into <code>ComfyUI/custom_nodes/</code>.</li><li>Install the other required custom nodes and models.</li><li>Restart ComfyUI and open the FL2VA or REF2VA workflow.</li></ol>Do not merge files from different versions of the custom node.Model paths stored in the workflows are examples. Select the corresponding files installed on your system if the paths or filenames differ.<h2 id="usage">Usage</h2>Configure the workflow through the MASTER node:<ul><li>quality mode and resolution;</li><li>aspect ratio;</li><li>continuity profile;</li><li><code>section_frames</code>;</li><li>audio-history length;</li><li>prompts for S1–S6.</li></ul>FL2VA modes:<ul><li>T2V: keep START FRAME and END FRAME bypassed.</li><li>I2V: enable START FRAME only.</li><li>First/last-frame generation: enable both START FRAME and END FRAME.</li></ul>In REF2VA, all four references are bypassed by default. Enable them consecutively: REF 1, REF 2, REF 3, and REF 4. Use only the corresponding <code>&lt;Picture 1&gt;</code>–<code>&lt;Picture 4&gt;</code> tags in the prompts.For the first REF2VA test, use REF 1 with the DIRECT quality mode.<h2 id="recommended-final-post-processing">Recommended final post-processing</h2>After assembling the final video, I recommend processing it with <a target="_blank" rel="ugc" href="https://github.com/Merserk/dlss5-visual-enhancer">DLSS 5 Visual Enhancer</a>.It can improve perceived detail, upscale the completed video, and optionally increase the frame rate using DLSS Frame Generation.DLSS 5 Visual Enhancer is optional and runs separately from ComfyUI.<a target="_blank" rel="ugc" href="https://github.com/Merserk/dlss5-visual-enhancer/releases">Download DLSS 5 Visual Enhancer</a>My Telegram channel:<a target="_blank" rel="ugc" href="https://t.me/aibobpublic">https://t.me/aibobpublic</a> MiniMax H3 Long Video Workflows — V2.0<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/7684df1c-485c-42a5-af2b-3d56cbee097c/original=true/7684df1c-485c-42a5-af2b-3d56cbee097c.jpeg" />V2.0 includes two six-section long-video workflows:<ul><li>FL2VA — supports T2V, I2V, last-frame guidance, and first/last-frame generation.</li><li>REF2VA — supports up to four optional image references connected across all six sections and HQ refinement.</li></ul><h2 id="main-features">Main features</h2><ul><li>Direct transfer of video and audio latents between sections.</li><li>22-frame AV context for seamless continuation.</li><li>Accurate cumulative audio timeline.</li><li>Automatic section length and final video duration calculation.</li><li>MASTER node for aspect ratio, resolution, quality preset, section length, and S1–S6 prompts.</li><li>Native generation or optional learned latent upscale with 3-step HQ refinement.</li><li>Individual section checkpoints and final assembled video.</li><li>Compatible with Sage Attention, Fused Modulation, and FFN chunking.</li><li>Turbo LoRA affects S1–S6 and HQ refinement; the secondary motion/style LoRA affects only S1–S6.</li><li>Separate FL2VA and REF2VA prompt-writing guides are included.</li></ul><h2 id="models">Models</h2>MiniMax H3 models:<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main">Comfy-Org/MiniMax-H3</a>Heretic Text Encoder — choose one:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot/tree/main">INT8 ConvRot</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4/tree/main">NVFP4</a></li></ul>Turbo LoRA options:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/DarkRomeo88/MiniMax-H3-turbo-lora-comfyui/tree/main">DarkRomeo88 — ComfyUI version</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora/tree/main">LarryVRH version</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras">Kijai experimental LoRAs</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors">LightX2V REF2VA Turbo</a></li></ul>Optional latent upscaler:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/tree/main">Upscaler model</a></li><li><a target="_blank" rel="ugc" href="https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler">ComfyUI upscaler node</a></li></ul><h2 id="required-custom-nodes">Required custom nodes</h2><ul><li><a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">ComfyUI-KJNodes</a></li><li><a target="_blank" rel="ugc" href="https://github.com/Saganaki22/ComfyUI-sol-attn">ComfyUI-sol-attn</a></li><li>h3_direct_latent_head — included in the V2.0 archive.</li></ul>Copy <code>h3_direct_latent_head</code> into:<code>ComfyUI/custom_nodes/</code>Then restart ComfyUI and open the required FL2VA or REF2VA JSON workflow.Model and LoRA paths stored in the workflow are examples. Select the corresponding files installed on your system if your folder structure or filenames differ.<h2 id="usage-notes">Usage notes</h2>Edit the MASTER node to select the quality preset, aspect ratio, <code>section_frames</code>, and prompts for S1–S6.In REF2VA, all four references are bypassed by default. Enable references consecutively starting from REF 1. The included example prompts assume that REF 1 is active.For REF2VA prompts:<ul><li>S1: define references and subjects, identity retention, action, camera, and sound.</li><li>S2–S6: repeat subject definitions and retention rules, then continue from the transferred latent context without resetting the pose or scene.</li></ul>See the included FL2VA and REF2VA prompt guides for complete templates and examples.My Telegram channel:<a target="_blank" rel="ugc" href="https://t.me/aibobpublic">https://t.me/aibobpublic</a> V1.0 All Minimax H3 models download from:<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main">https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main</a>Heretic Text Encoder int8:<a target="_blank" rel="ugc" href="https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot/tree/main">https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot/tree/main</a>Heretic Text Encoder nvfp4:<a target="_blank" rel="ugc" href="https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4/tree/main">https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4/tree/main</a>Turbo loras:1) <a target="_blank" rel="ugc" href="https://huggingface.co/DarkRomeo88/MiniMax-H3-turbo-lora-comfyui/tree/main">https://huggingface.co/DarkRomeo88/MiniMax-H3-turbo-lora-comfyui/tree/main</a> (without larryvrh nodes) and <a target="_blank" rel="ugc" href="https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora/tree/main">https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora/tree/main</a> (with larryvrh nodes).2) Also recommend turbo loras:<a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras">https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras</a>3) And:<a target="_blank" rel="ugc" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors">https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors</a>Required custom nodes:KJNodes:<a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">https://github.com/kijai/ComfyUI-KJNodes</a>ComfyUI-sol-attn:<a target="_blank" rel="ugc" href="https://github.com/Saganaki22/ComfyUI-sol-attn">https://github.com/Saganaki22/ComfyUI-sol-attn</a>I also developed a custom node for transferring encoded latents between sections. This allows the encoded video and audio latents from the previous section to be passed directly to the next section, helping to avoid overcooking during continuation.For the workflow to work, download the custom node archive and place the folder from the archive into the ComfyUI/custom_nodes/ folder.You can download it from the drop-down list on this page. There is a separate archive for the workflows and a separate archive for the custom node. That is, REF2VA formula for all long-videoS1:REFERENCE BINDING + IDENTITY LOCK + STYLE + ACTION + CAMERAS2–S6:CONTINUE FROM LATENT CONTEXT + IDENTITY LOCK FROM OPENING FRAMES + ENVIRONMENT CONTINUITY + ACTION + CAMERA My TG Channel:<a target="_blank" rel="ugc" href="https://t.me/aibobpublic">https://t.me/aibobpublic</a>

Minimax H3 LONG VIDEO (FL2VA, REF2VA)

MiniMax H3 Long Video Workflows — V2.0V2.0 includes two six-section long-video workflows:<ul><li>FL2VA — supports T2V, I2V, last-frame guidance, and first/last-frame generation.</li><li>REF2VA — supports up to four optional image references connected across all six sections and HQ refinement.</li></ul>Main features<ul><li>Direct transfer of video and audio latents between sections.</li><li>22-frame AV context for seamless continuation.</li><li>Accurate cumulative audio timeline.</li><li>Automatic section length and final video duration calculation.</li><li>MASTER node for aspect ratio, resolution, quality preset, section length, and S1–S6 prompts.</li><li>Native generation or optional learned latent upscale with 3-step HQ refinement.</li><li>Individual section checkpoints and final assembled video.</li><li>Compatible with Sage Attention, Fused Modulation, and FFN chunking.</li><li>Turbo LoRA affects S1–S6 and HQ refinement; the secondary motion/style LoRA affects only S1–S6.</li><li>Separate FL2VA and REF2VA prompt-writing guides are included.</li></ul>ModelsMiniMax H3 models:<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main">https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main</a>Heretic Text Encoder — choose one:<ul><li>[INT8 ConvRot <a target="_blank" rel="ugc" href="https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot/tree/main">https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot/tree/main</a></li><li>NVFP4 <a target="_blank" rel="ugc" href="https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4/tree/main">https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4/tree/main</a></li></ul>Turbo LoRA options:<ul><li>DarkRomeo88 — ComfyUI version <a target="_blank" rel="ugc" href="https://huggingface.co/DarkRomeo88/MiniMax-H3-turbo-lora-comfyui/tree/main">https://huggingface.co/DarkRomeo88/MiniMax-H3-turbo-lora-comfyui/tree/main</a></li><li>LarryVRH version <a target="_blank" rel="ugc" href="https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora/tree/main">https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora/tree/main</a></li><li>Kijai experimental LoRAs <a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras">https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras</a></li><li>LightX2V REF2VA Turbo <a target="_blank" rel="ugc" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors">https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors</a></li></ul>Optional latent upscaler:<ul><li>Upscaler model <a target="_blank" rel="ugc" href="https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/tree/main">https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/tree/main</a></li><li>ComfyUI upscaler node <a target="_blank" rel="ugc" href="https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler">https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler</a></li></ul>Required custom nodes<ul><li>ComfyUI-KJNodes <a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">https://github.com/kijai/ComfyUI-KJNodes</a></li><li>ComfyUI-sol-attn <a target="_blank" rel="ugc" href="https://github.com/Saganaki22/ComfyUI-sol-attn">https://github.com/Saganaki22/ComfyUI-sol-attn</a></li><li>h3_direct_latent_head — included in the V2.0 archive.</li></ul>Copy <code>h3_direct_latent_head</code> into:<code>ComfyUI/custom_nodes/</code>Then restart ComfyUI and open the required FL2VA or REF2VA JSON workflow.Model and LoRA paths stored in the workflow are examples. Select the corresponding files installed on your system if your folder structure or filenames differ.Usage notesEdit the MASTER node to select the quality preset, aspect ratio, <code>section_frames</code>, and prompts for S1–S6.In REF2VA, all four references are bypassed by default. Enable references consecutively starting from REF 1. The included example prompts assume that REF 1 is active.For REF2VA prompts:<ul><li>S1: define references and subjects, identity retention, action, camera, and sound.</li><li>S2–S6: repeat subject definitions and retention rules, then continue from the transferred latent context without resetting the pose or scene.</li></ul>See the included FL2VA and REF2VA prompt guides for complete templates and examples.My Telegram channel:<a target="_blank" rel="ugc" href="https://t.me/aibobpublic">https://t.me/aibobpublic</a>

Files with SHA256: 7fe12eac431865aeb5de420ff86614f807f9331312ab7c1359717912843a3152