BFS Shot H3 Conditioning (Advanced)

What it replaces

If you were building the shot loop by hand, this node is the pile of work you'd otherwise do once per shot: call MiniMax H3's Reference to Video with the shot's prompt, references and length, optionally bolt an aligned guide on with Add Guide, optionally pin the shot's own clip as a duet panel, and push the model through a RoPE patch if you're using the shifted layout. You feed it a shot and it hands back positive, latent, model and prompt.

Then it runs once per shot, because the planner's shots list made it a list node. You sample the latent, decode, and drop the frames into BFS Shot Join.

The setting that decides everything

guide_mode is the one to think about:

• aligned guide (Add Guide) - the shot's frames sit on the generated frames, frame by frame. This is the mode for body-swap LoRAs, which were trained on same-resolution aligned guides.
• native reference video (Video 1) - the shot enters as a reference video, the way H3's Reference to Video node takes a clip, with its soundtrack.
• both - the two at once, which is heavier and smarter than it sounds.
• none - conditioning from prompt and references only.

first_frame is a different mechanism from continuity, and it's easy to confuse the two: shot's first frame anchors the shot's own frame from the source video at frame 0, which is what you want for a cut-free continuation of the real footage. Continuity in the planner pulls the previous generated result, which is what you want across cuts. Both at once is usually wrong.

duet turns the shot's own clip into a pinned side panel - off, canvas, or shifted RoPE. For shifted RoPE you must route the model through this node and use its model output, since the shift is applied to the model. The nice detail: BFS Shot Join cuts the panel off by itself, so the sampler and decode half of your workflow doesn't change at all when you turn the duet on. panel_position, panel_size, panel_noise and rope_gap only matter when it's on, and panel_noise 0 means "pin it exactly" while 0.05–0.2 loosens it for bigger changes. Same warning as everywhere else with the shifted layout: keep rope_gap small or H3 draws its own split screen.

The prompt writer, and setting_ref

By default task = planner prompt, which uses each shot's prompt from the planner as-is. Pick any other task (character swap, style, setting, appearance, lighting/weather, custom) and the node writes each shot's prompt in the duet format instead - with a Qwen3-VL on vlm it looks at the shot and its references and writes it, otherwise it falls back to the task's template. The finished text comes out on prompt so you can read it, which is how you decide whether the VLM is earning its VRAM. instruction is the few words about what changes; for a pure character swap you can leave it empty because the person comes from the references.

setting_ref is worth knowing because it's a trick rather than a feature, credited in the docs to TSC: with it on, the node adds one more reference - the shot's middle frame with the person masked out in TV static - so the model sees the place in full detail when the panel or guide is too small to carry it. It becomes the last <Picture n> in the prompt, and setting_mask lets you supply the person's mask yourself; otherwise it falls back to the shot's SAM 3 crop mask, then to SAM 3 "person" on that frame. on (source size, slower) uses the video's own resolution up to a 2048 short edge - sharper, and it costs you.

use_ref_2 and ref_image_size are small but real: the second reference is on by default, and ref_image_size decides whether references come back at match (the generation size) or max.

Install

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

Restart, and reload the tab. You need a ComfyUI build with the native H3 nodes and an H3 model, VAE and text encoder. Give it an audio_vae when with_audio is on, or the shot's soundtrack won't ride along with the guide video.

Watch for

The latent output is the H3 AV latent, so the sampler's latent_image is this one - not something you built separately. And if you turn the duet on and the result is a split screen, you almost certainly didn't connect the planner's shots output through the whole chain, so the join never got the panel geometry it needs to crop.