FangSen9000 commited on
Commit
362b97b
·
verified ·
1 Parent(s): a4bb8ea

Add README

Browse files
Files changed (1) hide show
  1. README.md +68 -0
README.md ADDED
@@ -0,0 +1,68 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ pretty_name: MAGICK Transparency (WSM v1 raw)
4
+ task_categories:
5
+ - image-segmentation
6
+ - image-to-image
7
+ tags:
8
+ - transparency
9
+ - alpha-matting
10
+ - layer-diffuse
11
+ - magick
12
+ - wsm
13
+ ---
14
+
15
+ # MAGICK Transparency — raw WSG samples
16
+
17
+ Raw (pre-VAE) samples for the WSM v1 transparency-estimation pilot
18
+ (`magick_transparency_v1`): predict a foreground **alpha / transparency** layer from an RGB
19
+ image. The matching FLUX.2 latents live in the companion latent repo
20
+ `World-Snapshot/WSM_v1_Latent_PT_WSG_v1.8_for_Flux2`
21
+ (`custom_pilots/magick_transparency_v1/cache_named/`).
22
+
23
+ ## Source
24
+
25
+ Built from the **MAGICK** RGBA asset set (transparent foreground objects). Each foreground is
26
+ composited over random backgrounds so the model must recover the soft alpha of the object from a
27
+ realistic RGB image.
28
+
29
+ ## How each sample is produced
30
+
31
+ Each self-describing named-key `.npz` holds two modalities (`uint8`, `H×W×3`; the trainer maps
32
+ pixels with `x/127.5 − 1`):
33
+
34
+ | key | meaning |
35
+ |---|---|
36
+ | `0-2` | **RGB** — the foreground composited over a random background. |
37
+ | `LayerDiffuse_Transparency` | **target** — the RGB-replicated soft alpha / transparency of the foreground (1-channel alpha stored replicated to 3 channels). |
38
+
39
+ This is an estimation task: RGB → transparency. No caption, no reference image.
40
+
41
+ ## Layout
42
+
43
+ ```text
44
+ wsg_named/ the named-key .npz samples (0-2 + LayerDiffuse_Transparency)
45
+ ```
46
+
47
+ `split_manifest.jsonl` records the deterministic train/val/test assignment used by WSM v1 training
48
+ (see `split_summary.json`); the split is per-sample (npz path), reproducing the trainer's hash split
49
+ at `val_split_ratio=0.05`.
50
+
51
+ ## Packaging
52
+
53
+ Streamed as multi-part `MAGICK_Transparency_partNNN.tar.zst` (each < 15 GB) with
54
+ `..._parts_manifest.jsonl` (path → part index) and `..._parts_summary.json`.
55
+
56
+ ### Extract
57
+
58
+ ```bash
59
+ mkdir -p data/custom_pilots/magick_transparency_v1
60
+ for f in MAGICK_Transparency_part*.tar.zst; do
61
+ tar -xf "$f" -C data/custom_pilots/magick_transparency_v1/
62
+ done
63
+ # -> data/custom_pilots/magick_transparency_v1/wsg_named/
64
+ ```
65
+
66
+ ## License
67
+
68
+ MIT for the packaging/processing; underlying MAGICK assets follow their original terms.