ComfyUI-LASED-mask

ComfyUI-LASED-mask

LASED — Learning Ant Swarm Edge Detection as ComfyUI custom nodes: a colour-region-aware, learnable edge/contour extractor (an alternative to Canny/HED), with a masking mode that latches the swarm's answer into a clean, exportable MASK.

You pick a point on the image; a seeded swarm of Snake-style ants spawns there, adopts that pixel's colour as a target, and — confined to that colour region — naturally drifts to and walks along the region's boundary, depositing pheromone. The accumulated edge-pheromone field is the output edge map. Because LASED finds the border of a colour blob rather than intensity gradients, it often gives exactly the contour a user means when they "click a region".

Algorithm contract: LASED-grounding.md (the upstream LASED repository); the masking mode implements the Stage 2 sticky-latch / freeze / cleanup design proven in the LASED desktop app (v0.6.0).

Install

Clone into your ComfyUI custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/MushroomFleet/ComfyUI-LASED-mask

Dependencies: numpy (always present in ComfyUI). scipy is used when available for faster diffusion and component labelling; everything falls back to pure numpy without it.

Nodes (category: LASED)

| Node | Inputs | Outputs | Purpose | |---|---|---|---| | LASED Swarm Edge | IMAGE, seed point, params | IMAGE edge map, MASK | One-shot run: spawn at a point, simulate, return the edge-pheromone field. | | LASED Multi-Seed | IMAGE, seeds ("x,y; x,y; ..."), params | IMAGE, MASK | One swarm per colour region; edge fields merged (max), masks unioned. | | LASED Food Map | IMAGE, food MASK, seed point, params | IMAGE, MASK | The mask is standing food/reward: the swarm is biased toward (and reinforces) your chosen contours — the offline analogue of interactive "success clicks". | | LASED Overlay | IMAGE, edge IMAGE, colour, alpha | IMAGE | Glowing-contour composite for previews. | | LASED Animate | IMAGE, seed point, params, frames | IMAGE batch | The simulation as frames — feed to video-combine / preview-animation nodes. | | LASED Point Picker | IMAGE, point | seed_x INT, seed_y INT, point_str STRING | Visual point-and-click seed picker: the image shows inside the node, click to place the crosshair, wire the outputs into the swarm nodes. |

LASED Point Picker

Picking seed_x/seed_y by hand is fiddly, so the LASED Point Picker node shows the input image inside the node and turns a click into coordinates:

1. Connect the same IMAGE that feeds your swarm node to the picker's image input.
2. Press ↻ refresh preview. If the image traces back to a LoadImage/LoadImageMask, the file is shown instantly with no execution; otherwise the picker queues a partial execution of just its upstream subgraph (needs ComfyUI ≥ 0.3.41). Running the full workflow once also (re)loads the preview.
3. Click (or drag) on the preview to place the crosshair — the live pixel readout is shown bottom-left. The choice is stored in the point widget as image fractions (0.5,0.5), so it stays valid if the image is swapped for a different resolution; you can also type absolute pixels (values > 1) by hand.
4. On the swarm node, convert the seed_x/seed_y widgets to inputs and connect the picker's INT outputs. The point_str output ("x,y") can feed Multi-Seed's seeds field via a string-concat node when collecting several picks.

Classic vs masking mode

Every swarm-running node has a mode switch:

• classic — the reference LASED behaviour. The MASK output is the normalised edge-pheromone field thresholded at mask_threshold. The live field evaporates by design, so the mask reflects the field exactly as the run ended.
• masking (default) — the Stage 2 sticky-latch pipeline. A persistent mask layer latches any cell whose edge pheromone crosses sticky_threshold × the field maximum at deposit time; latched cells never decay. When the run ends (or freeze_after ticks fire first) the simulation freezes and a cleanup pass runs: morphological closing of 1–2 px gaps, then connected-component filtering that keeps the dominant contour(s) and drops stray-ant speckle. The MASK output is that cleaned binary mask — ready for inpainting, compositing, or as a ControlNet scribble/contour hint. no_decay additionally stops edge-pheromone evaporation (explore pheromone still fades, so anti-clumping keeps working).

Parameters worth knowing (grounding §10)

| Param | Default | Effect | |---|---|---| | tolerance | 40 | Colour tolerance τ — the most important knob. Lower = tighter region. Lab distances run smaller; try 10–25 there. | | ants / steps | 250 / 600 | More = denser, more complete contours. | | color_space | rgb | lab traces one continuous boundary where RGB fragments under uneven lighting. | | hand | both | Wall-following handedness; both splits the swarm into CW + CCW cohorts. | | beta / epsilon | 4.0 / 0.05 | Softmax greediness / random-move floor. | | seed | 0 | All randomness is seeded — same seed + params ⇒ identical output, batch item i runs with seed + i. | | max_side | 512 | Simulation grid cap (grounding §3.1); outputs are resampled back to input resolution. 0 = full res. |

Coordinates (seed_x/seed_y, multi-seed points) are in input-image pixels.

Example workflows

Every node has a drop-in example in workflows/ — open one in ComfyUI, point LoadImage at any picture, set the seed coordinates to a pixel inside the colour region you want traced, and queue.

• lased_smoke.json — smoke test: LASED Swarm Edge → LASED Overlay → save (overlay + raw edge map).
• lased_multi_seed.json — LASED Multi-Seed: one swarm per x,y; x,y point, merged edge field overlaid plus the union mask saved via MaskToImage.
• lased_food_map.json — LASED Food Map: a second LoadImageMask provides the food MASK (paint bright where you want the contour rewarded), biasing the swarm toward those edges.
• lased_animate.json — LASED Animate: captures the contour emerging over the run and saves it with SaveAnimatedWEBP.
• lased_point_picker.json — LASED Point Picker driving LASED Swarm Edge: refresh the preview, click the region you want traced, queue.

Tests

The simulation core (lased_core.py) is torch- and ComfyUI-free, with acceptance tests mirroring the upstream plans (seeded determinism, latch coverage, mask persistence, TTF freeze, cleanup behaviour, classic-mode regression):

python tests/test_core.py        # or: pytest tests/

Provenance

• lased_core.py is the reference LASED.py from the LASED repository with the Stage 2 masking mode added. Classic mode keeps the reference RNG stream and field updates byte-for-byte, so cross-implementation agreement per seed holds against the Python, TypeScript and Rust cores.
• Node family per grounding §11; deferred there as stub I-01 and built here as its own project.

License

Apache-2.0 (same as the upstream LASED repository).