BFS Sampler (crop · mask · loop)

The BFS Head Swap Sampler (crop · mask · loop) is the node half of the BFS "Best Face Swap" LTX workflow, and the name undersells it. It's a whole sampling pipeline in one node: guide video and identity image in, and it crops, masks, samples in temporal chunks, and feathers the result back - all while the head-swap LoRA is never asked to understand a mask. This is the node the LoRA was trained around: guide video on source 1 (body, motion, camera), identity image on source 2, 73 frames per chunk.

LTX head swaps aren't free. LTX 2.3 gets you a fast, genuinely good video model, but a convincing swap still needs a LoRA trained on this reference-conditioning recipe (the pack README calls it video_to_video_ref_adaln). What this node adds on top is the boring production plumbing that makes swaps survive real footage: a stable crop when the face is small in frame, per-frame mask inpainting so only the head changes, and chunked sampling so long clips don't melt your VRAM.

How it works

Under the hood it's a sampler-shaped node - that's why it takes noise, sampler, sigmas, and guider inputs straight out of a normal LTX sampling chain. The guide and identity frames get VAE-encoded into reference latents and injected into the model through transformer_options patches (via the pack's LTXMultipleControls), then each chunk is denoised through your guider. Three extras sit on top:

• Mask inpainting rides ComfyUI's native path. The mask becomes a latent noise_mask, so outside the mask nothing changes and the original face stays visible to the model. No LoRA training needed for this.
• Cropping uses a planner vendored from drozbay's MaskVidExperiments (GPL-3.0, same as the pack). Its boxes hold still through mask noise and occlusion - a jittering per-frame crop reads to a video model as camera motion, and it can't carry identity.
• Paste-back is edge-aware. The feather only fades inward; sides of a crop sitting on the image edge aren't feathered, because there's nothing outside to blend into.

Why crop at all? At 512×288 a person filling a fifth of the frame leaves a face ~25 px tall. No LoRA recovers identity from that. Sampling the head region full-frame gives it 200–300 px, then pastes the result back.

The inputs that matter

• guide_video and identity_image - the two things that make it a swap. Crop the identity to the head; the guide carries the whole source performance.
• subject_mask - per-frame mask of the region to edit (head, with a margin). Leave it unconnected for a plain swap. Include the hair if you want it retained; people trip on this constantly.
• crop_mode - off / combined / tracked / zoomed. Off is fine when the face is big in frame; the others sample inside a box around the subject instead. Start with combined, one static box for the clip.
• latent - connect it. Feed an empty latent from EmptyLTXVLatentVideo sized to the crop. LTX-2.5 latents are AV (video+audio); without one the node builds a plain video latent and the model may ignore the guide entirely.
• temporal_tile_size - 0 samples the whole clip in one pass. For clips longer than 73 frames (the training length of the BFS head-swap recipe), set it to 73. One catch: chunked sampling with an AV nested latent throws an error - set this back to 0 or feed a video-only latent.

Outputs worth knowing: images (final frames), latent (the crop's latent when cropping is on - this is what you upscale in a second pass), crop_bboxes (feed to the BFS Head Swap Paste Back node), plus inspection views cropped_guide, mask_over_source, and latent_mask that show you exactly what the model was fed.

Install and what it needs

Install once, like any pack:

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

Restart ComfyUI, then add a BFS head-swap LoRA to ComfyUI/models/loras/. You also need an LTX model and VAE. One dependency gotcha: the crop planner imports scipy, which is listed in the pack's pyproject.toml but not in requirements.txt. If a crop mode dies on import, pip install scipy.

Where people get burned

• Identity drift or a face that ignores the reference is almost always the face being tiny in frame (turn cropping on) or a reference that isn't a tight frontal head crop. The reference image matters more than any knob here.
• A soft seam at the head's edge means softness reached the denoise mask. Keep mask_hard_for_inpaint on (default): blur belongs to the paste-back, not the sampling, or the original identity bleeds back in at the boundary.
• Second-pass decode hangs on a big upscaled latent - switch decode to tiled.

The sampler rewards the "connect only what you need" philosophy: guide + identity is a plain swap, add a mask to restrict it, add a crop when the face is small. Start there and you'll be surprised how often the defaults just work.