Try It: Loss Masking

1. Image

Choose image… Open job bundle…

No image loaded. You can also drop a file on the canvas.

2. Layers

Click a layer to select it; the tools paint into it. Higher layers win where layers overlap. Layers set to None are not exported. Hidden layers are still exported.

Name

Colour

Thread width 4.5 px Angle 0° (0 = vertical threads)

Knot radius 13 px

Coil pitch 7 px

3. Mask tool

Box Brush Magic wand

Eraser (subtract), or hold Alt

Brush radius 16 px

Tolerance 32 Non-contiguous (whole image)

4. Quantise by colours

Splits the image into flat colour areas, one new layer each. Then pick a stitch for each layer.

Colours 6 Smoothing 2 px Merge areas under 0 px Only inside layer Replace layers from the last quantise

To quantise just a loss, mark it as a layer first, choose it under “Only inside layer”, then set that layer to None or delete it.

5. Custom stitches

For a stitch LoRA you trained yourself.

Add custom stitch

Name Trigger word Init style LoRA file

Saved as loras/<file> in the job. Put the .safetensors file in a loras folder in the bundle, or upload it alongside the zip in Colab. It must suit the model it runs on (a FLUX LoRA for flux_base).

6. Procedural preview

Preview fill Preview seed

Each layer's colour and shape settings are in section 2.

7. Export

Seed Variants per region

Variants: 0 = one fill per region; more saves extra fills to choose between.

Colab: choose Runtime → Run all, and pick the zip when asked.

Locally: python pipeline/staged_reintegrate.py --job <file>.zip --model flux_base --lora_variant trigonly_v2, or drop zips in a folder and run python pipeline/run_jobs.py <folder>.

Other downloads

Per-stitch masks are binary PNGs (white = selected) named <stem>__<stitch>.png, one per stitch in use.

Drop an image or job bundle here