BFS Shot Noise

What it's for

The first few sampling steps are where a video decides its composition and who the subject ends up being. That decision is made almost entirely by the noise you start from - which is why re-rolling a seed changes the layout so violently, and why a shot loop with the same seed everywhere produces shots that all secretly agree on the same coarse structure.

BFS Shot Noise is the starting noise for SamplerCustomAdvanced, with optional structure on top: you can share the coarse pattern across every shot, freeze part of it across time, or lean the noise toward your guide and reference. Leave every amount at 0 and it is exactly ComfyUI's RandomNoise for the same seed - that's the design constraint the author built around, and it's a good one. This is not a node that "improves" your baseline run.

How the shaping works

Everything here keeps the noise at unit variance, so the sampler still receives the amount of noise it expects - that's what stops these knobs from being an accidental strength slider.

• shared_base replaces the coarse block averages of the noise with averages drawn from base_seed instead of the shot's seed. Fine detail stays per shot. The grid is defined in blocks (base_grid, default 6 along the short side) rather than latent cells, so the same base seed gives you the same coarse pattern at any resolution, and time_block (default 2 latent frames) makes it a long tape that each shot reads at its own timeline position. Give a successful shot's seed to base_seed and you keep its layout while the detail re-rolls - that's the feature, and it's why the input exists at all.
• static_share mixes in a component that's identical in every frame. Temporally correlated noise, basically: it steadies things up, and the tooltip is honest that too much of it freezes your motion.
• guide_structure encodes the guide with the vae and mixes it into the noise outside the mask, so the background starts closer to the guide. Without a mask it applies everywhere, which also pulls toward the original person.
• ref_tint leans the noise inside the mask toward the reference image's average latent - its colours. Very small values only; 0.05–0.15.
• distribution - gaussian (what the model was trained on), laplacian, student-t (heavier-tailed, so sparser and harder kicks) or uniform. spectrum_tilt and pyramid are two ways to push energy into the large scales: tilt is a 1/f^α spectrum, the pink/fractal noise the tooltip compares to AdvancedNoise's fractal alpha, and pyramid adds progressively coarser layers. The pitch: the coarse layout is where a swap is decided, so give it a harder push.

Inputs and outputs

Required: seed, shot_seed (same / increment / random per shot - only active when a shot is connected), and the four amounts plus distribution, spectrum_tilt and pyramid. The optional side is where the interesting bits live: shot brings the per-shot seeds and the timeline position, base_seed overrides which seed owns the coarse pattern (-1 means "use the one above"), overrides takes shot:seed pairs, base_grid and time_block tune the coarse grid, and vae + guide_frames + ref_image + mask are the inputs guide_structure and ref_tint need.

Two outputs: noise (NOISE) goes to SamplerCustomAdvanced - this is not a seed you can paste into a plain KSampler, which takes a seed and makes its own noise - and seed (INT), which is the seed that shot actually used. Wire that second one to your filename or a note node; it's the only way to reproduce a take.

Install

ComfyUI Manager: search ComfyUI-BFSNodes. Manual:

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

Restart ComfyUI. No extra models - vae is the video VAE you're already loading.

When to bother, and the ways it bites

The author's own description of this node says it best: the model was trained on plain noise, so use small amounts and compare against 0. That's not boilerplate. All of this is off-distribution input, and a swap that follows the pose beautifully at guide_structure 0.1 can look like it's fighting the source person at 0.4. Suggested starting ranges, from the tooltips: shared_base 0.5–1, static_share 0.1–0.3, guide_structure 0.05–0.2, ref_tint 0.05–0.15, spectrum_tilt 0.2–0.6, pyramid 0.3–0.6.

Two concrete failure modes worth knowing before you blame the LoRA:

• A red error from the node means you asked for a mix without the ingredients: guide_structure needs both vae and guide_frames, ref_tint needs vae and ref_image. The error text says so.
• It only touches the video latent. If the latent isn't the joint video+audio structure, the node hands back untouched noise, and the audio half always stays gaussian regardless - don't expect these knobs to change the soundtrack.
• shot_seed does nothing without a shot. No shot, no index to iterate: you get exactly the seed you typed. Intended behaviour, and the first thing to check when the node seems inert.