A multiscale benchmark for latent style fidelity in generative art models
Python
0
278 commits
updated Oct 4, 2026
Does painter-name prompting recover differences between painters, or mainly produce a shared painting-like appearance?
This project compares generated images with digital reproductions of paintings by Monet, Sisley, Pissarro and Cézanne using 31 interpretable color, spatial and texture features, with CLIP and CSD representations as a same-image check. It separates the appearance common to all painter names from reference-aligned painter variation. It measures agreement with digital reference contrasts; perceptual style and internal model mechanisms remain unvalidated.
The primary prospective experiment has 1,008 images: six image models × 14 scenes × six prompt clauses × two repeats. The models are GPT Image 1, GPT Image 2, GPT Image 2.5 Flare, GPT Image 2.5 Sunburst, Nano Banana 2 and FLUX.2 Max.
docs/STATUS.md has the full findings, the retrospective analyses, open decisions and limits.
| Manuscript | Source | Scope |
|---|---|---|
| TMLR submission (lead) | paper/tmlr/main.tex | Anonymous TMLR-format manuscript: the shared/between-name decomposition, learned embeddings, readout disagreement and the SD-Turbo check; its PDF is built locally with make paper-tmlr |
| Korean translation of the TMLR manuscript | paper/tmlr_ko/main.tex | Translation by an AI assistant, made anew on 2026-10-04 and not yet reviewed by the author; same numbers, tables and figures as the TMLR manuscript; built locally with make paper-tmlr-ko |
The full-length paper of 2026-09-15 and its Korean translation were deleted on 2026-10-04 at the user's request; Git history keeps them (last in commit a55a2b4). The paper guide explains the layout and builds.
Use the recorded Python 3.13.11 runtime and uv. The package metadata's
broader Python floor does not qualify every frozen study on older interpreters.
uv sync --locked --extra analysis --extra dev --inexact
make install-hooks # Reject staged files over GitHub's 100 MiB limit
make restore-analysis # Extract and checksum-verify the frozen transfer result
make specificity-check # Primary numerical results
make review-check # Post-result diagnostics
make reference-quality-check
make retrospective-check # Retrospective analyses (direct naming, timing, learned, transfer, covariance)
make tmlr-check # TMLR tables, figure and every number quoted in its text
make figures-check # Retained palette figure against its replay
make check # Ruff and the routine test suite
These commands make no generation requests and need no API key or model weights.
The frozen transfer result is stored as a lossless analysis.json.gz archive.
make restore-analysis restores its original path and bytes without changing
any scientific hashes; make retrospective-check and make tmlr-check run this step automatically.
Run it before invoking the transfer Python scripts or numerical bundle tools directly.
A Git checkout contains compact measurements, not image pixels; commands that
re-extract features or verify pixel hashes need the retained local archive.
The analysis catalog lists every replay command.
| Location | Purpose |
|---|---|
| paper/ | TMLR manuscript with its asset builder, and the frozen earlier manuscripts |
| src/latent_art_bench/ | Versioned analysis code and shared measurement primitives |
| studies/ | Protocols, fixed plans and methodological boundaries |
| data/manifests/ | Compact measured vectors, request records, hashes and receipts |
| reports/ | Numerical results, report plots, release receipts and review records |
| tests/ | Offline analysis and integrity checks |
| scripts/, tools/ | Record audits, artifact checks, historical collectors and release adapters |
| configs/ | Executed study configurations |
| docs/ | Status, handover, analysis catalog and retention policy |
| critics/, literature_reviews/ | 2026-09-13 external reviews; original literature base |
Historical protocols, source, results and local raw evidence are needed for reproducibility and are bound by path and hash. Read the artifact policy and contributing guide before moving or deleting research files.
Earlier studies have versioned public releases; the six-model experiment has not been released yet. Code is distributed under the repository license. Full-resolution artwork is not redistributed; the paper includes four reduced reference examples with recorded public-domain metadata.
A multiscale benchmark for latent style fidelity in generative art models
Python
0
278 commits
updated Oct 4, 2026
Does painter-name prompting recover differences between painters, or mainly produce a shared painting-like appearance?
This project compares generated images with digital reproductions of paintings by Monet, Sisley, Pissarro and Cézanne using 31 interpretable color, spatial and texture features, with CLIP and CSD representations as a same-image check. It separates the appearance common to all painter names from reference-aligned painter variation. It measures agreement with digital reference contrasts; perceptual style and internal model mechanisms remain unvalidated.
The primary prospective experiment has 1,008 images: six image models × 14 scenes × six prompt clauses × two repeats. The models are GPT Image 1, GPT Image 2, GPT Image 2.5 Flare, GPT Image 2.5 Sunburst, Nano Banana 2 and FLUX.2 Max.
docs/STATUS.md has the full findings, the retrospective analyses, open decisions and limits.
| Manuscript | Source | Scope |
|---|---|---|
| TMLR submission (lead) | paper/tmlr/main.tex | Anonymous TMLR-format manuscript: the shared/between-name decomposition, learned embeddings, readout disagreement and the SD-Turbo check; its PDF is built locally with make paper-tmlr |
| Korean translation of the TMLR manuscript | paper/tmlr_ko/main.tex | Translation by an AI assistant, made anew on 2026-10-04 and not yet reviewed by the author; same numbers, tables and figures as the TMLR manuscript; built locally with make paper-tmlr-ko |
The full-length paper of 2026-09-15 and its Korean translation were deleted on 2026-10-04 at the user's request; Git history keeps them (last in commit a55a2b4). The paper guide explains the layout and builds.
Use the recorded Python 3.13.11 runtime and uv. The package metadata's
broader Python floor does not qualify every frozen study on older interpreters.
uv sync --locked --extra analysis --extra dev --inexact
make install-hooks # Reject staged files over GitHub's 100 MiB limit
make restore-analysis # Extract and checksum-verify the frozen transfer result
make specificity-check # Primary numerical results
make review-check # Post-result diagnostics
make reference-quality-check
make retrospective-check # Retrospective analyses (direct naming, timing, learned, transfer, covariance)
make tmlr-check # TMLR tables, figure and every number quoted in its text
make figures-check # Retained palette figure against its replay
make check # Ruff and the routine test suite
These commands make no generation requests and need no API key or model weights.
The frozen transfer result is stored as a lossless analysis.json.gz archive.
make restore-analysis restores its original path and bytes without changing
any scientific hashes; make retrospective-check and make tmlr-check run this step automatically.
Run it before invoking the transfer Python scripts or numerical bundle tools directly.
A Git checkout contains compact measurements, not image pixels; commands that
re-extract features or verify pixel hashes need the retained local archive.
The analysis catalog lists every replay command.
| Location | Purpose |
|---|---|
| paper/ | TMLR manuscript with its asset builder, and the frozen earlier manuscripts |
| src/latent_art_bench/ | Versioned analysis code and shared measurement primitives |
| studies/ | Protocols, fixed plans and methodological boundaries |
| data/manifests/ | Compact measured vectors, request records, hashes and receipts |
| reports/ | Numerical results, report plots, release receipts and review records |
| tests/ | Offline analysis and integrity checks |
| scripts/, tools/ | Record audits, artifact checks, historical collectors and release adapters |
| configs/ | Executed study configurations |
| docs/ | Status, handover, analysis catalog and retention policy |
| critics/, literature_reviews/ | 2026-09-13 external reviews; original literature base |
Historical protocols, source, results and local raw evidence are needed for reproducibility and are bound by path and hash. Read the artifact policy and contributing guide before moving or deleting research files.
Earlier studies have versioned public releases; the six-model experiment has not been released yet. Code is distributed under the repository license. Full-resolution artwork is not redistributed; the paper includes four reduced reference examples with recorded public-domain metadata.