Drop in one image, describe the scene and the sound, and get a short video with audio generated by H3 itself: voice, ambience, music.What makes this workflow different:<ul><li>🔁 Last frame saved automatically. Every run saves the final frame as an image. Load it as the next start frame and your clips continue seamlessly. It’s the easiest way to build longer sequences.</li><li>🎬 Two videos per run. A raw 24 fps version, plus a smoother 48 fps, 1.5x upscaled version. Both keep the original audio in sync.</li><li>🎯 Start frame only. No end frame forcing the motion, so H3 moves more freely and naturally.</li><li>⚡ Turbo switch. Fast 8-step drafts with one click; turn it off for full-quality renders.</li><li>⏱ Duration in seconds. Just type how long you want; the workflow handles H3’s frame math.</li></ul>Clean layout, notes inside the workflow, and only three common custom node packs. How it works<ul><li>Base: official Comfy-Org MiniMax H3 I2V template (<code>MiniMaxH3ImageToVideo</code>, fl2va model), unpacked from the subgraph so every setting is visible.</li><li>Conditioning: <code>first_frame</code> only (<code>last_frame</code> intentionally unconnected). Resolution comes from the Resolution Selector (default 0.7 MP / 1152×640, 16:9; 0.98 MP = official 768p).</li><li>Length: Math Expression converts seconds to frames, rounding up to H3’s 17k+5 grid at 24 fps (5 s → 124 frames).</li><li>Sampling: <code>SamplerCustomAdvanced</code> + <code>res_multistep</code>. A boolean switches between base (20 steps) and the fl2v turbo LoRA (8 steps, on by default). A Power LoRA Loader after the switch accepts extra H3 LoRAs; Wan LoRAs are not compatible.</li><li>Output chain:<ol><li>Video and audio VAE decode → <code>CreateVideo</code> → <code>GetVideoComponents</code>, which splits frames and audio.</li><li>Last frame: <code>GetImageSize</code> (batch count) → <code>a − 1</code> → <code>ImageFromBatch</code> → <code>SaveImage</code>. It works with any duration.</li><li>Raw: VHS Video Combine, 24 fps, h264, CRF 19, original audio.</li><li>Upscaled: RIFE VFI 2x (rife49) at native resolution → lanczos 1.5x → VHS Video Combine, 48 fps, CRF 17, same audio. Interpolating before the upscale saves VRAM.</li></ol></li></ul>Requirements<ul><li>Latest ComfyUI. It uses recent core nodes: MiniMax H3, Math Expression, Get Video Components, Switch and Resolution Selector.</li><li>Custom nodes: ComfyUI-VideoHelperSuite, ComfyUI-Frame-Interpolation, rgthree-comfy.</li><li>Models:<ul><li><code>minimax_h3_fl2va_pruned_int8_convrot</code></li><li><code>qwen3vl_32b_minimax_h3_nvfp4_awq</code></li><li>H3 video VAE (fp16) and audio VAE (fp32)</li><li><code>minimax_h3_fl2v_turbo_8step</code></li></ul> Download links are in the “Model Links” note.</li></ul>Tips &amp; limits<ul><li>Out of VRAM in RIFE: lower <code>clear_cache_after_n_frames</code>. Out of RAM: lower <code>scale_by</code> or bypass the upscale branch.</li><li>Match the Resolution Selector aspect to your start frame to avoid stretching.</li><li>H3 is heavy; a high-VRAM GPU or a cloud instance is recommended. I run this workflow on Nordy (cloud ComfyUI).</li></ul>

MiniMax H3 Image-to-Video — with auto last frame, smooth 48 fps and native audio

First public release.<ul><li>Start frame only (no end frame) for freer motion</li><li>Native H3 audio kept in all outputs</li><li>Last frame auto-saved for chaining clips</li><li>Raw 24 fps + RIFE 48 fps with 1.5x upscale</li><li>Turbo switch (8-step LoRA, on by default) + Power LoRA slot</li><li>Duration in seconds, auto-snapped to H3’s 17k+5 frame grid</li></ul>

Files with SHA256: 9fdb4f5fc3145aef9981db4fe5037fe937b62f76dc12dfc45e0d9f9019c60b0e