---
license: mit
pretty_name: MAGICK Transparency (WSM v1 raw)
task_categories:
  - image-segmentation
  - image-to-image
tags:
  - transparency
  - alpha-matting
  - layer-diffuse
  - magick
  - wsm
---

# MAGICK Transparency — raw WSG samples

Raw (pre-VAE) samples for the WSM v1 transparency-estimation pilot
(`magick_transparency_v1`): predict a foreground **alpha / transparency** layer from an RGB
image. The matching FLUX.2 latents live in the companion latent repo
`World-Snapshot/WSM_v1_Latent_PT_WSG_v1.8_for_Flux2`
(`custom_pilots/magick_transparency_v1/cache_named/`).

## Source

Built from the **MAGICK** RGBA asset set (transparent foreground objects). Each foreground is
composited over random backgrounds so the model must recover the soft alpha of the object from a
realistic RGB image.

## How each sample is produced

Each self-describing named-key `.npz` holds two modalities (`uint8`, `H×W×3`; the trainer maps
pixels with `x/127.5 − 1`):

| key | meaning |
|---|---|
| `0-2` | **RGB** — the foreground composited over a random background. |
| `LayerDiffuse_Transparency` | **target** — the RGB-replicated soft alpha / transparency of the foreground (1-channel alpha stored replicated to 3 channels). |

This is an estimation task: RGB → transparency. No caption, no reference image.

## Layout

```text
wsg_named/     the named-key .npz samples (0-2 + LayerDiffuse_Transparency)
```

`split_manifest.jsonl` records the deterministic train/val/test assignment used by WSM v1 training
(see `split_summary.json`); the split is per-sample (npz path), reproducing the trainer's hash split
at `val_split_ratio=0.05`.

## Packaging

Streamed as multi-part `MAGICK_Transparency_partNNN.tar.zst` (each < 15 GB) with
`..._parts_manifest.jsonl` (path → part index) and `..._parts_summary.json`.

### Extract

```bash
mkdir -p data/custom_pilots/magick_transparency_v1
for f in MAGICK_Transparency_part*.tar.zst; do
  tar -xf "$f" -C data/custom_pilots/magick_transparency_v1/
done
# -> data/custom_pilots/magick_transparency_v1/wsg_named/
```

## License

MIT for the packaging/processing; underlying MAGICK assets follow their original terms.