<a target="_blank" rel="ugc" href="https://www.youtube.com/watch?v=n4yO0oqBWdM">▶ Video guide: how I use it in ComfyUI</a>: <a target="_blank" rel="ugc" href="https://www.youtube.com/watch?v=n4yO0oqBWdM">https://www.youtube.com/watch?v=n4yO0oqBWdM</a>Put the person from one photo into your video clip. Everything else in the clip stays: the place, the moves, the camera and the sound.<a target="_blank" rel="ugc" href="https://discord.gg/RwVxY9Fuaj">Discord</a> · <a target="_blank" rel="ugc" href="https://www.youtube.com/@ausbossdev">YouTube @ausbossdev</a><h3 id="how-to-use-it">How to use it</h3><ol><li>Load your clip in Video - Your clip and trim it to one shot, up to 10 seconds.</li><li>Load the new person's photo in Picture 1.</li><li>Say who goes, then describe the new person's hair and face on a second line: <code>Replace the man in the red shirt with @1.</code> <code>@1 looks like a man with short curly brown hair and glasses.</code></li><li>Pick High VRAM for 24 GB or more, Low VRAM under that, and press Run.</li></ol>Turn on Describe @1 for me and the workflow writes the second line from the photo. Turn on SeedVR2 under Finish the clip for a sharper, bigger clip.<h3 id="what-you-need">What you need</h3>Update ComfyUI (0.38.0 or newer) and <a target="_blank" rel="ugc" href="https://github.com/ausboss/ComfyUI-AusBoss">AusBoss nodes</a> (2.21.2) first.<ul><li><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>models/diffusion_models</code> · 21.0 GB</li><li><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>models/text_encoders</code> · 15.7 GB</li><li><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>models/vae</code> · 5.2 GB</li><li><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>models/vae</code> · 0.6 GB</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/pika_h3_flex_ref2va_v1_bf16.safetensors">pika_h3_flex_ref2va_v1_bf16.safetensors</a> → <code>models/loras</code> · 2.0 GB</li><li>Low VRAM · <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_w6a8.safetensors">minimax_h3_ref2va_pruned_w6a8.safetensors</a> → <code>models/diffusion_models</code> · 16.0 GB, in place of the first file</li><li>Low VRAM · <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_int8_convrot.safetensors">minimax_h3_video_vae_int8_convrot.safetensors</a> → <code>models/vae</code> · 2.8 GB, in place of the video VAE</li><li>Describe @1 for me · <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Qwen3-VL/resolve/main/text_encoders/qwen3vl_8b_fp8_scaled.safetensors">qwen3vl_8b_fp8_scaled.safetensors</a> → <code>models/text_encoders</code> · 10.6 GB</li><li>SeedVR2 · <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/SeedVR2/resolve/main/diffusion_models/seedvr2_3b_int8_convrot.safetensors">seedvr2_3b_int8_convrot.safetensors</a> → <code>models/diffusion_models</code> · 3.46 GB</li><li>SeedVR2 · <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/SeedVR2/resolve/main/vae/seedvr2_ema_vae_fp16.safetensors">seedvr2_ema_vae_fp16.safetensors</a> → <code>models/vae</code> · 501 MB</li></ul>A 9 second clip took about 3.5 minutes on 24 GB, and about the same on 16 and 12 GB with Low VRAM. With SeedVR2 on, it took 5.5 minutes on 24 GB. On 8 GB, use Low VRAM with Fast preview on; the Settings + troubleshooting note in the workflow has the launch flags.<h3 id="tips">Tips</h3><ul><li>Only two lines reach H3 in a swap: who goes, and what @1 looks like. A separate <code>@1 wears ...</code> line is left out, so put clothes in the looks-like line: <code>@1 looks like a young woman with a short black bob, wearing a red raincoat.</code> Clothes that differ a lot from the clip's can make the video open on the photo itself. If that happens, take them out of the line.</li><li>Pick a clip where the face is big enough and turns toward the camera at some point. When it is small or turned away, H3 can keep the clip's own face and change only the hair and clothes; words about skin or eye colour did not change that in our tests.</li><li>Without the second line, the swap can start late or not happen at all.</li><li>Turn Steer the mouths on only when people talk in the clip. With music, the new person sings along.</li><li>Describe @1 for me loads one more model, and a run used about 60 GB of system RAM. With 64 GB or less, type the line yourself.</li></ul>Powered by MiniMax H3 under the <a target="_blank" rel="ugc" href="https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE">MiniMax H3 Community License</a>. The Pika H3 Flex LoRA is by Pika; the ComfyUI files are Comfy-Org's. Optional upscaling uses <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/SeedVR2">SeedVR2</a> under Apache 2.0.v1.0 (2026-11-07) · First release.<a target="_blank" rel="ugc" href="https://github.com/ausboss/ComfyUI-AusBoss/issues">Report an issue</a> · <a target="_blank" rel="ugc" href="https://discord.gg/RwVxY9Fuaj">Discord</a> · <a target="_blank" rel="ugc" href="https://www.youtube.com/@ausbossdev">YouTube @ausbossdev</a> · <a target="_blank" rel="ugc" href="https://x.com/Zanzibased">X @Zanzibased</a> · <a target="_blank" rel="ugc" href="https://github.com/ausboss">GitHub</a>

MiniMax H3 Pika Person Swap | High + Low VRAM, 8 GB+

v1.1: runs on 8 GB cards, and FlashVSR finishes the clip.
Needs AusBoss nodes 2.25.2 or newer.
<ul>
<li>Finish the clip now offers FlashVSR in place of SeedVR2: sharper and 2x larger, one model file (3.5 GB).</li>
<li>Low VRAM is for 8 to 24 GB. On 8 GB it runs at full size with clips up to 8 seconds and 32 GB of system RAM, with no launch flags (ComfyUI 0.45). A new Low VRAM note in the workflow lists its two files and our 8 GB test.</li>
<li>Describe @1 for me fit in 32 GB of system RAM in our test.</li>
<li>The Steer the mouths switch is gone: it never changed anything. Keep the clip's sound moves the new person's mouth, as before.</li>
<li>Trimming the clip: the IN handle now drags on any clip.</li>
<li>A fix for the <code>'latent'</code> error from another H3 pack is in Settings + troubleshooting.</li>
</ul>

Files with SHA256: 1e084d5a2097dadedeffb8a443b77c0baf71b4208eac89a19ff1a2b37a9f4bc3