Elfsong/poster-layer-decomposition-48

Automatic masks and layer extraction

An English viewer for 48 Vietnamese marketing posters, with all images and precomputed results stored in this Space. Select a poster and a candidate to compare the original-pixel mask cutout with the raw Qwen RGBA output. Click an image to enlarge it, switch the preview background, or download a mask or layer.

• —457 object candidates: Grounding DINO base + SAM 2.1 Hiera large.
• —694 text-line candidates: Hi-SAM-L / HierText stroke masks, grouped into lines.
• —33 unassigned stroke groups: preserved for inspection, not sent to Qwen.
• —1,151 Qwen extraction attempts: Qwen-Image-Layered-Control-V2, 30 steps, seed 777, area resolution 640; each call used the original poster.

The interface is in English. The original Vietnamese poster artwork is unchanged. The Space serves a completed experiment and loads image assets on demand.

Findings

Four overlapping cars in VI_033 were selected and extracted separately. Occluded parts remain missing, and generated details may change. Text generation remains unreliable: 298 of 694 text outputs were nearly blank; other examples contain background fringes or colored blocks. Of the 457 object outputs, 115 were nearly blank.

“Nearly blank” means no alpha value exceeds 32/255. These counts are diagnostics, not a semantic accuracy metric. Automatic masks can miss elements, merge objects or miss diacritics. Original-pixel cutouts use binary alpha and still need edge cleanup. The results are not a guaranteed reconstructable layer stack.

Illustrated notes

Data and downloads

• —data/vi_###/: original poster, one lossless RGB image, overview overlays, all 1,184 candidate masks, raw Qwen PNGs, and a viewer manifest.
• —48-page mask overview
• —Masks and split plans
• —Original posters, masks, hints, Qwen layers and run metadata
• —Batch artifact validation

The browser reconstructs each cutout from the lossless RGB image and its mask. PNG downloads preserve the exact RGBA pixels, including RGB beneath transparent pixels. This avoids storing the same poster pixels in hundreds of separate files. The full archive also includes the original posters and masks needed to reproduce the cutouts. It retains unsuccessful Qwen outputs. Model weights and interactive session histories are not part of this viewer.

Local preview

python -m http.server 7860 --directory .

Open http://localhost:7860. The app uses relative asset paths and has no runtime dependency on the original GPU worker or temporary Gradio tunnel.