Minimax H3 LONG VIDEO (FL2VA, REF2VA)

MiniMax H3 Long Video Workflows — V4.0V4.0 is a complete redesign of the MiniMax H3 long-video workflows. It is smaller, easier to configure, and supports from 1 to 12 sequential sections.The workflow transfers protected video and audio latent context between sections, allowing MiniMax H3 to generate a longer continuous video instead of unrelated clips.Included workflows<ul><li>FL2VA — supports T2V, I2V, last-frame guidance, and first/last-frame generation.</li><li>REF2VA — supports image, video, video-soundtrack, and standalone audio references.</li></ul>This release contains two archives:<ul><li>Workflows + Instruction Prompts — both V4.0 workflows and bilingual Russian/English prompt guides.</li><li>h3_direct_latent_head — the shared custom node pack required by both workflows.</li></ul>Download and install both files.What changed since V3.0<ul><li>The large fixed graph was replaced with compact dynamic generation.</li><li>The number of sections can be selected from 1 to 12.</li><li>You mainly configure only the MASTER and H3 MODEL STACK nodes.</li><li>The learned latent upscaler and HQ refinement were removed.</li><li>Every section is generated directly at the selected native resolution.</li><li>Only one selected model branch and one attention backend are used.</li><li>Turbo 4-step, Turbo 8-step, Native 20-step, and Native 25-step profiles are available.</li><li>Seed increment per section is enabled by default to reduce repetitive motion.</li><li>FL2VA and REF2VA use separate validated prompt formats.</li><li>REF2VA now supports image, video, video-audio, and standalone audio references.</li><li>REF2VA includes a safer reference policy for long-video continuation.</li><li>Final audio is decoded once from the complete cumulative AV timeline.</li><li>Model Preview Override is included and enabled.</li></ul>V4.0 prioritizes native generation, continuity, and a simpler workflow over experimental upscaling and refinement.How the long-video system worksEach section is a generation window. A protected part of the previous section’s video and audio latent is transferred into the next section.The recommended overlap is 22 frames. It helps preserve:<ul><li>the current pose and movement;</li><li>character appearance;</li><li>camera position and direction;</li><li>lighting and scene geometry;</li><li>continuous sound and ambience.</li></ul>With the default settings:<ul><li>6 sections;</li><li>141 frames per section;</li><li>22 protected context frames;</li><li>24 FPS;</li></ul>the final video contains 736 frames, or approximately 30.67 seconds.The first section contributes all 141 frames. Each following section contributes 119 new frames because 22 frames are reused as continuity context.Installation<ol><li>Stop ComfyUI completely.</li><li>Delete the previous folder:<code>ComfyUI/custom_nodes/h3_direct_latent_head</code></li><li>Extract the new <code>h3_direct_latent_head</code> folder into:<code>ComfyUI/custom_nodes/</code></li><li>Do not merge files from different node versions.</li><li>Install the required external custom nodes.</li><li>Place the MiniMax H3 models, text encoder, video VAE, audio VAE, and optional LoRAs in the standard ComfyUI model folders.</li><li>Restart ComfyUI.</li><li>Open the required FL2VA or REF2VA V4.0 workflow.</li></ol><code>H3 MODEL STACK</code> uses portable <code>AUTO</code> selections by default. If several matching files are found, select the correct model or LoRA manually.FL2VA modesControl the generation mode using the START and END image inputs:<ul><li>T2V: START and END are bypassed.</li><li>I2V: enable START only.</li><li>Last-frame guidance: enable END only.</li><li>FL2V: enable both START and END.</li></ul>Each active section requires a separate prompt with three blocks:<ul><li><code>integrated_multimodal_description</code></li><li><code>overall_soundscape</code></li><li><code>non_diegetic_music</code></li></ul>S1 establishes the subject, scene, camera, movement, and sound. S2 and later prompts should describe the next stage of the same continuous action.Do not repeat the same action as a new command in every section, because the model may replay it repeatedly.A complete Russian/English FL2VA prompt guide is included.REF2VA referencesREF2VA supports:<ul><li>up to 4 image references;</li><li>up to 3 video references;</li><li>soundtracks from reference videos;</li><li>up to 3 standalone audio references.</li></ul>Enable inputs consecutively inside each type: input 1 before input 2, and input 2 before input 3.Reference videos are loaded at 24 FPS and limited to the first 360 frames / 15 seconds.For the first test, use:<ul><li>one reference;</li><li>0.31 MP resolution;</li><li>22F context;</li><li>0F audio history;</li><li><code>reference_policy = S1 ONLY</code>.</li></ul>S1 ONLY — recommendedReal image, video, and audio references are presented to the model only during S1. S2–S12 continue through the protected AV latent.This reduces the risk of later sections suddenly reproducing:<ul><li>the original reference image;</li><li>the original pose or composition;</li><li>the reference background;</li><li>the original video movement;</li><li>the reference audio as a restarted sound.</li></ul>Use <code>&lt;Picture N&gt;</code>, <code>&lt;Video N&gt;</code>, and <code>&lt;Audio N&gt;</code> tags only in the S1 prompt. Do not use reference tags in S2–S12.EVERY SECTIONThis mode presents all references again during every section.It can provide stronger reference adherence, but it may also reproduce the original pose, scene, motion, background, or audio as a separate insert. It also requires more processing time and memory.Each active REF2VA section requires a separate prompt with six blocks:<ul><li><code>subject_definitions</code></li><li><code>summary</code></li><li><code>retention_analysis</code></li><li><code>detailed_description</code></li><li><code>overall_soundscape</code></li><li><code>non_diegetic_music</code></li></ul>A complete Russian/English REF2VA prompt guide with S1 and continuation templates is included.Recommended starting settings<ul><li>Resolution: 0.31 MP</li><li>Section length: 141F</li><li>Context: 22F balanced</li><li>Audio history: 0F off</li><li>Seed strategy: increment per section</li><li>Reference image size: MATCH</li><li>REF2VA reference policy: S1 ONLY</li><li>Video shift: 12</li><li>Audio shift: 3</li></ul>Increase the resolution and number of references only after a successful test.Sampling and LoRAsAvailable sampling profiles:<ul><li>Turbo 4 steps</li><li>Turbo 8 steps</li><li>Native 20 steps</li><li>Native 25 steps</li></ul>The sampling profile must match the selected LoRA. For example, use a 4-step Turbo LoRA with the 4-step profile and an 8-step Turbo LoRA with the 8-step profile.FL2VA and REF2VA require LoRAs intended for the corresponding model type.Attention backendsThe workflow uses exactly one selected attention backend:<ul><li>COMFY KITCHEN</li><li>SAGE</li><li>PYTORCH</li></ul>Comfy Kitchen is selected in the supplied workflows. If it is unavailable in your ComfyUI build, select SAGE or PYTORCH.SAGE requires ComfyUI-KJNodes.Required modelsMiniMax H3 models:<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main">Comfy-Org/MiniMax-H3</a>Compatible 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>Optional 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 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>Required custom nodes<ul><li><a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">ComfyUI-KJNodes</a> — required only for SAGE Attention.</li><li><a target="_blank" rel="ugc" href="https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite">ComfyUI-VideoHelperSuite</a> — required for REF2VA video references.</li><li>h3_direct_latent_head — included with this release and required by both workflows.</li></ul>No latent-upscaler node is required.Model Preview Override and stabilityModel Preview Override is included and enabled so intermediate sampling previews can be monitored during generation.On some Windows and PyTorch configurations, preview decoding may occasionally cause:<code>Fatal Python error: Aborted</code>If this happens:<ol><li>Refresh the ComfyUI browser page.</li><li>Bypass the Model Preview Override node.</li><li>Queue the workflow again.</li></ol>Bypassing Model Preview Override disables only the intermediate preview. It does not change the final generation, prompts, video continuity, audio timeline, or output quality.<code>Comfy model compiler graph breaks</code> is diagnostic information and does not mean that generation has failed.The node pack and workflow contracts were tested against ComfyUI 0.34.0.Optional final post-processingIf additional upscaling or frame interpolation is required, process the completed video afterward with:<a target="_blank" rel="ugc" href="https://github.com/Merserk/dlss5-visual-enhancer">DLSS 5 Visual Enhancer</a>It supports RTX Video Super Resolution, DLSS Neural Rendering, and optional frame generation. It is a separate application and is not required for these ComfyUI workflows.<a target="_blank" rel="ugc" href="https://github.com/Merserk/dlss5-visual-enhancer/releases">Download DLSS 5 Visual Enhancer</a>

Files with SHA256: 49512112e533f67ff8db9db702603b957650b58d442ea9663513d708e7dcb6f0