Can a closed image model make editable, completed poster layers?
A precomputed baseline on the same 48 public Vietnamese marketing posters.
What ran
GPT-6 Astra plans independent objects, coherent text blocks, logos, decorations and their stacking order. Each color-layer request receives the original poster and asks for one element at its original relative position, including plausible hidden parts. A separate request removes the planned foreground layers and fills the background.
All model work runs through codex exec. Image editing uses Codex’s built-in image tool. The tool does not expose its exact image model, snapshot, seed or quality settings. This is therefore a Codex workflow baseline, not a model-locked GPT Image 2 evaluation.
Why there are three representations
| View | What it contains | How to interpret it |
|---|---|---|
| Native image | The first image-tool output, unchanged | A painted checkerboard is still an opaque image. |
| GPT opacity mask | A second image-tool output, conditioned on the native color layer | White means opaque; black means transparent. It can be misaligned or contain holes. |
| Derived RGBA | Native RGB plus the mask’s grayscale values as alpha | RGB is unchanged. This is a packed result, not native transparent model output. |
No SAM, Qwen, OCR rerendering, thresholding, morphology or manual edge cleanup is used in this branch. A mask is resized with bilinear interpolation only when its dimensions differ from the color layer; that conversion is recorded.
What the pilots showed
The three pilot reconstructions were visibly poor. Generated people, cars, text and graphic panels changed scale and position. Separately predicted opacity masks also shifted or retained unwanted content, creating fringes, holes and pieces of painted checkerboard. More complete individual objects did not produce faithful poster reconstruction.
Overlapping cars, VI_033: the blue car’s hidden body and wheels were generated. The packed result also showed light edge fringes and a problematic plate region. Completion and clean matting are separate problems.
Vietnamese headline, VI_001: the main lettering was recognizable, but the rendering added a shadow and changed its extent. A separate mask made a transparent file while introducing its own edge mismatch.
Background, VI_019: the foreground was removed and the stadium/graphic background was filled. The fill is generated content and should be checked for changes to visible scenery.
A simpler design, VI_021: the large text on a flat colored panel stays closer to its original layout, although small logos still drift or duplicate. In denser posters, such as VI_004, independently generated objects can grow dramatically and cover neighboring content.
These are qualitative visual observations, not accuracy measurements over all 48 posters. Use the full overview to inspect the range of results.
How to read reconstruction
Layers are composited in the GPT-predicted order. Whole canvases are normalized to the original aspect ratio; there is no automatic alignment correction. Missing results and foregrounds without usable alpha are excluded by default and the reconstruction is labeled partial. Their raw outputs remain available.
The earlier Qwen outputs use different automatically selected candidates and resolutions. They provide a visual reference; this is not a controlled comparison with matched masks, layer counts or compute.
Protocol and data
The first color and opacity outputs are retained, including poor results and rejected calls. A separate minimal transparency probe is retained outside the main layer inventory. Early pilot opacity prompts used explicit subject descriptions; later calls use the common template. All available instructions, outputs and conversion metadata accompany the data. A quota interruption left 15 image files without final result JSON; these were recovered unchanged from their Codex session directories and explicitly marked as recovered metadata. Their exact submitted image prompts and tool-call counts are unavailable. Only stages without retained outputs were retried after quota was replenished.
Model documentation: GPT-6 Astra · GPT Image 2. Documentation describes the named models; it does not establish the hidden backend used by Codex’s image tool.