A sample dataset creation tool for the Qwen-Image-Layered trainer on fal.ai. Generates layered image datasets from the PrismLayersPro dataset on Hugging Face.
uv syncuv run main.py \
--subset cartoon \
--number-of-layers 3 \
--seed 42 \
--sample-count 100 \
--output ./my_dataset| Option | Required | Description |
|---|---|---|
--subset |
Yes | PrismLayersPro style category (e.g., cartoon, anime, pixel_art, 3D) |
--number-of-layers |
Yes | Number of output layers (>= 2) |
--solid-background-only |
No | Only select samples with solid-color backgrounds |
--seed |
Yes | Random seed for reproducibility |
--sample-count |
Yes | Number of samples to generate |
--output |
Yes | Output folder (must not exist) |
3D, Pokemon, anime, cartoon, doodle_art, furry, ink, kid_crayon_drawing, line_draw, melting_gold, melting_silver, metal_textured, neon_graffiti, papercut_art, pixel_art, pop_art, sand_painting, steampunk, toy, watercolor_painting, wood_carving
For each sample (e.g., sample 000):
000.txt # Caption describing the complete image
000_start.png # Complete composite image (reference/target, fully opaque)
000_end.png # Base/background layer
000_end2.png # First foreground layer
000_end3.png # Second foreground layer
...
000_endM.png # Remaining layers composited together
All output images are RGBA PNGs with identical dimensions. Stacking _end*.png files with alpha compositing recreates the _start.png image.
- M = 2:
end.png= base,end2.png= all foreground layers composited - M > 2:
end.png= base,end2throughend(M-1)= first M-2 layers (preserving z-order),endM= remaining layers composited
Generate a 2-layer dataset from pixel art:
uv run main.py --subset pixel_art --number-of-layers 2 --seed 123 --sample-count 50 --output ./pixel_2layerGenerate a 4-layer dataset with solid backgrounds only:
uv run main.py --subset anime --number-of-layers 4 --solid-background-only --seed 456 --sample-count 25 --output ./anime_4layer