BFS H3 Duet (pinned panel, all in one)

The idea in one line

You have a clip of someone doing something. You want a different person doing exactly that thing - same gestures, same cuts, same camera, no reshoot. The normal answer is a body-swap LoRA trained on your model. This node is the answer for people who don't have one: it pins the source clip beside the video as a panel, tells MiniMax H3 to generate the new performance in sync with it, and cuts the panel off before you ever see it. No training, no LoRA, no dataset.

That's the trick. H3 treats the pinned strip like its own conditioning rows - clean, at the condition timestep - so it copies the motion frame for frame instead of being loosely "guided" by it.

Where this sits in the H3 world

MiniMax H3 is the 33B omni-modal video model that opened its weights in August 2026 with native stereo audio, and it quickly became the reference open model for video editing. Two things matter here: it has real reference conditioning built into ComfyUI's native H3 nodes, and it's big. Note the pack author's own credit - the latent-pin duet idea came from TSC, who pinned the source with a noise mask first; the BFS contribution is the shifted RoPE layout, dynamic references and this all-in-one node.

How it works, mechanically

Inside, BFS H3 Duet is a whole workflow: it calls H3's Reference to Video to build positive conditioning and the AV latent, optionally adds an aligned guide, pins the panel strip into the latent with a noise mask, samples with Guider_Basic (which means CFG 1 - there is no negative prompt here, and no CFG knob), crops the strip out of the latent, then decodes picture and audio. So it both builds conditioning and runs the sampler. Two layout modes: canvas makes panel and video share one wide RoPE grid (TSC's approach), and shifted keeps the video on exactly the RoPE positions it would have had with no panel at all, putting the panel past its edge - that's the layout the author says is new. If you'd rather use your own sampler, use BFS H3 Duet Conditioning instead; same panel, stops before sampling.

The inputs you'll actually touch

• model, clip, vae - the H3 model, the H3 Qwen3-VL text encoder and the video VAE. clip is where the panel content gets seen; the panel itself never goes to the text encoder.
• task + instruction - with an empty prompt box, these write a draft prompt; instruction is "what changes, in a few seen words". Be honest about the drafts: the node can't see your clip, so they're generic, and the ones that worked best in testing were written for the clip.
• prompt - your own REF2VA prompt and always the winner over task. The panel has no tag: it is not <Picture n> and not <Video n>, so name it by position ("the LEFT half is the kept footage") or write {layout} and let the node insert that sentence for the current side and size. And don't describe the panel's person, clothes or room even to contrast it - what you leave undescribed is what gets copied.
• width / height / length - output size and frames at 24 fps, snapped to 17k+5. length of 0 means "as long as the panel clip or guide".
• position / size - which side the strip goes on, and how big it is relative to the video (1.0 = two equal halves).
• panel_noise - 0 pins the panel exactly, which is what a character swap or a lighting change wants. Push to 0.1–0.2 when you want room to invent (style transfers, turning someone into a creature).
• rope_mode / rope_gap - leave canvas unless you're experimenting with shifted, in which case rope_gap is a trap: keep it small (0–2 at low resolution), because a gap close to the video's own width makes the model draw its own split screen.
• Optional panel - the clip or still you're pinning. Leave it empty and this is just a plain guided H3 render. guide is the aligned latent guide the body-swap LoRAs want. ref_images / ref_videos / ref_audios grow as you connect them, like the native Reference to Video node.

Outputs: images (the generated video, panel already cropped), audio, canvas (the full split screen - worth a look, it's how you check the two halves are actually in sync), layout_text (the {layout} sentence), latent, and prompt (what was actually sent, draft included).

Install

Search ComfyUI-BFSNodes in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/alisson-anjos/ComfyUI-BFSNodes
cd ComfyUI-BFSNodes && pip install -r requirements.txt

Restart ComfyUI. The pack is large - one clone gets you the whole BFS toolbox (LTXV Edit Anything, shot loop, LoRA surgery), and requirements.txt pulls insightface==0.7.3, onnxruntime, scenedetect, librosa and opencv-python along with it. That insightface pin is the one most likely to fight something else in your environment.

Where people get burned

If the node doesn't appear in the Add Node menu but the Side Panel nodes do, your ComfyUI build is too old - BFS H3 Duet is registered through the newer comfy_api.latest interface and silently skips registration when that import fails. You also need a build with native H3 support; BFS H3 Side Panel throws an explicit "expects a MiniMax H3 AV latent" error (and one about sizes needing to be multiples of 32) rather than doing something surprising, which is the good kind of failure.

And the boundary is worth stating plainly: this copies motion and timing, it is not an identity LoRA. The author's own tests found the pinned panel pulls the background toward the reference's backdrop when used with a guide, so pick one mechanism until a LoRA is trained for both.