BFS H3 Duet Conditioning (prompt writer)

What this one exists for

BFS H3 Duet is the friendly version: it builds the conditioning, pins the panel, samples and decodes, all behind one node. That's great until you want something the author didn't anticipate - a different guider, a LoRA you load yourself, an inpainting pass, your own noise. So there's this: the same pinned-panel conditioning, handed to you as conditioning to wire into whatever sampler you like.

Use it when you care about the graph. Use the all-in-one when you just want the result.

What it hands you

The node pins a panel (your source clip, or a still) beside the video inside the same latent, then outputs positive, latent, model, panel_info, prompt and canvas_preview.

That model output is the part that's easy to miss. In rope_mode = shifted the RoPE shift is applied to the model, so you must connect the model through this node and use the output - feed the original model to your sampler and the conditioning and the model will disagree about where things are. In the default canvas mode the model passes through untouched. The chain people run:

BFS H3 Duet Conditioning ─ positive ─► BasicGuider / SamplerCustomAdvanced
                          ─ latent  ─► (same sampler)
                          ─ panel_info ─► BFS H3 Side Panel Crop (feed it the sampled latent)
                          ─ model ─► sampler (shifted mode)

BFS H3 Side Panel Crop takes panel_info plus either the sampled latent or decoded images, so crop the latent before VAE Decode - same picture, less work.

The prompt, and the VLM

You get three tiers, in order of who wins: your own text in prompt; otherwise a connected VLM; otherwise a draft.

vlm is a normal CLIPLoader running a Qwen3-VL text encoder (the same encoder family H3 uses). Connect it and the node asks it to write a full duet prompt for task + instruction while looking at the clip and the references - genuinely useful, because the draft path can't see anything and therefore produces generic text. The prompt it writes is a six-section REF2VA prompt, and the result comes out on the prompt output so you can copy the good ones into your own prompt box and stop paying for the VLM.

task picks the shape of the request (character swap, style, setting, appearance, lighting/weather), and instruction is the one useful sentence about what changes - "a 1990s anime cel style", "an elderly woman with grey hair". For a straight character swap you can leave instruction empty, since the person comes from the references.

The prompt-writing rules are the same as the all-in-one node and they're the difference between a good result and a mushy one: the panel has no tag, it is invisible to the text encoder, so refer to it by position ("the LEFT half is the kept footage") or drop {layout} in and let the node fill it for the current side and size. Never describe the panel's performer, wardrobe or room - not even to say what it isn't. Anything you leave unspecified is copied from the panel, which is the whole point.

Inputs worth knowing about

• clip, vae, panel are required; panel is the clip you pin (one image per frame for a clip, or a single still for a static panel).
• width / height / length - output geometry; length 0 follows the panel clip, and frames snap to 17k+5.
• panel_noise - 0 copies the panel exactly. 0.1–0.2 gives the model room for big changes.
• position, size, fit, gap, hold - where the strip goes, how big, how the clip is fitted into it (contain keeps the whole frame with grey around it; cover crops), the grey separator width, and whether the panel is held for the whole clip or just its first latent frame.
• rope_mode / rope_gap - canvas or shifted. In shifted mode keep rope_gap tiny; a large gap against the video's width makes H3 draw its own split-screen.
• keep_mask + keep_video - the inpainting door. With a mask, the video area starts from keep_video (or the panel clip) and only the masked region is regenerated, so background and lighting stay pixel-exact. This is how you swap a person without the room drifting.
• Optional guide is an aligned latent guide in the video area, and ref_images grows as you connect reference pictures - they become the <Picture n> slots your prompt refers to.## Install

Manager → search ComfyUI-BFSNodes, or:

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

Then restart. You need an H3 build of ComfyUI and the H3 model + VAE; the pack's other dependencies (insightface==0.7.3, onnxruntime, scenedetect, librosa, opencv-python) come along whether or not you use them.

If something's off

If BFS H3 Duet Conditioning isn't in your node list at all while BFS H3 Side Panel is, your ComfyUI predates the comfy_api.latest node interface - the pack registers these two through it and skips them quietly when the import fails. If you're in shifted mode and the frame looks like a genuine split screen with a seam, you almost certainly didn't route the model output into your sampler, or rope_gap is too large. And if the VLM writes something bland, that's usually the encoder: the pack's own docs warn that some smaller or modified Qwen3-VL builds return empty answers to certain wordings.