<p>Hand it <strong>one photo of a person</strong>. It gives you back a completely different shot, in a place you never photographed, with that same person in it. Your photo never shows up in the result. It is read, and then set aside.</p><p>That is not the same thing as image to video, even though almost everyone starts by assuming it is. Image to video takes your picture and makes it move. This reads your picture, puts it down, and writes a new scene from scratch with that face in it.</p><div data-youtube-video><iframe width="640" height="480" allowfullscreen="true" autoplay="false" disablekbcontrols="false" enableiframeapi="false" endtime="0" ivloadpolicy="0" loop="false" modestbranding="false" origin playlist src="https://www.youtube.com/embed/_jRKXMaRYgM" start="0"></iframe></div><p>🎬 <strong>Watch it built step by step:</strong> <a target="_blank" rel="ugc" href="https://www.youtube.com/watch?v=_jRKXMaRYgM">the full walkthrough on YouTube</a>. Every setting, in order, on a normal home machine.</p>

<h3 id="two-files-in-this-post">Two files in this post</h3><ul><li><p><strong>Standard build</strong> - 20 steps. The one to start with if you want the safest result.</p></li><li><p><strong>Turbo build</strong> - 8 steps, and one extra file to download. On my machine a 10 second clip took <strong>153 seconds</strong> here against <strong>349 seconds</strong> on the standard build. Same person, same picture quality. I ran both on the same seed before putting this up.</p></li></ul><h3 id="what-you-need">What you need</h3><ul><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors">minimax_h3_ref2va_pruned_int8_convrot.safetensors</a> → <code>ComfyUI/models/diffusion_models/</code></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors">qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors</a> → <code>ComfyUI/models/text_encoders/</code></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors">minimax_h3_video_vae_fp16.safetensors</a> → <code>ComfyUI/models/vae/</code></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors">minimax_h3_audio_vae_fp32.safetensors</a> → <code>ComfyUI/models/vae/</code></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors">minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors</a> → <code>ComfyUI/models/loras/</code> <em>(turbo build only)</em></p></li></ul><p>The folder is the part people get wrong. A correct file in the wrong folder looks exactly like a missing model, and ComfyUI will not tell you which.</p><h3 id="two-things-worth-knowing">Two things worth knowing</h3><ul><li><p>The prompt does far more work here than in the other H3 workflows. You are describing a whole scene, not a movement, because nothing about the scene comes from your photo.</p></li><li><p>Both decoders are needed. One is for the picture and one is for the sound, and leaving either out fails in a way that does not obviously say so.</p></li></ul><p>I wrote up every setting, what each one does and the mistakes that cost me time, in <a target="_blank" rel="ugc" href="https://forgebun.com/go/civit_h3_turbo_ref">the full written guide</a>. It is free and there is no card.</p><p>There is also a paid course that goes further with this model if you want it. Either way the workflows here are the whole thing, not a trimmed version.</p>

Files with SHA256: 1f1c7108ddfab43e8d73b13ff370573c00151e5425a7aa9f1190e7a250ff6be8