Project layout — this bundle contains five stage directories from the AI-Designer pipeline: research/ (literature survey), architect/ (blueprint + ModelConfig), coder/ (PyTorch implementation), validator/ (tests + benchmarks), and documenter/ (this README plus docs/ and CHANGELOG.md). An optional paper/ directory holds the NeurIPS-format writeup when the paper-generation step was triggered.

The original research request that produced this bundle is preserved verbatim in prompt.md — if any URLs in the prompt were fetched server-side for additional context, their cleaned contents are appended there too.

---

ArtRestore-Diffusion

A unified latent diffusion architecture for style-consistent inpainting of heterogeneous damage types in historical artworks, with interpretable per-patch attribution to artist corpus exemplars.

ArtRestore-Diffusion addresses the problem that existing inpainting methods treat all damage types identically and provide no guarantee of artist-specific style consistency. The architecture combines (a) mask-adaptive FiLM conditioning that handles cracks, scratches, large missing regions, fading, and text overlay in a single forward pass, (b) artist-specific style embedding injection via AdaLN and cross-attention, (c) retrieval-augmented denoising through a style exemplar memory bank of undamaged artist patches, and (d) per-patch attribution maps tracing each generated region to its nearest source exemplar.

Status: Pre-experimental infrastructure. All unit tests and domain benchmarks pass on randomly initialized models. Performance targets (LPIPS, DISTS, style certificate) require trained-model evaluation.

Highlights

• Unified damage-type handling — A single mask-adaptive CNN + FiLM module processes binary (crack, scratch, hole, text) and continuous (fade) masks with no per-type routing; see ARCHITECTURE.md#3-the-core-component
• Style certificate — Formal bound on artist-specific style deviation via embedding distance from artist manifold centroid; violation rate < 10% patches exceeding 3σ is the target; see ARCHITECTURE.md#section-7
• Interpretable by construction — Cross-attention weights to the style exemplar memory bank are rendered into attribution maps linking each generated patch to a real source patch; see docs/ARCHITECTURE.md#43-attribution-map-rendering
• Linear-time inference — 50-step DDIM sampling (vs. 1000-step DDPM) with deterministic, reproducible attribution maps; see BENCHMARKS.md#profiling

Quick start

pip install -r requirements.txt
python smoke_test.py        # 13 tests, verifies all component shapes + gradients + losses
pytest -v test_model.py     # 18 unit tests (shapes, gradients, correctness, numerics)
pytest -v test_benchmarks.py  # 19 domain-specific benchmarks

Repository layout

coder/
  config.py              — ArtRestoreConfig dataclass (all hyperparameters)
  layers.py              — Shared UNet layers (ResBlock, SpatialAttention, etc.)
  mask_encoder.py        — Mask-adaptive CNN + FiLM modulation
  style_block.py         — Artist embedding table + AdaLN + cross-attention injection
  exemplar_memory.py     — Style exemplar memory bank + cross-attention retrieval
  boundary_gate.py       — Learned per-pixel blending at mask edges
  diffusion_schedule.py  — Linear beta schedule + DDIM sampling
  model.py               — ArtRestoreBlock, ArtRestoreUNet, ArtRestoreModel
  attribution.py         — Attribution map rendering + style certificate
  inference.py           — End-to-end restoration pipeline
  trainer.py             — Training harness with EMA, AdamW, LR schedule
  smoke_test.py          — 13-test smoke suite for shape/gradient/loss verification
validator/
  test_model.py          — 20 pytest unit tests (shapes, gradients, correctness, numerics)
  test_benchmarks.py     — 19 domain-specific benchmarks (damage types, style, memory, etc.)
  run_ablations.py       — 11 single-field ablation configs from architect traceability table
  profile_model.py       — GPU memory, operator timing, FLOP estimation
  research_eval/         — Research-quality scoring, claim grounding, experiment coverage

Documentation

• docs/ARCHITECTURE.md — Design, inductive biases, equations, shape evolution, design decisions, domain-specific considerations, known limitations
• docs/TRAINING.md — Environment setup, hyperparameters, training recipe, expected behavior, troubleshooting
• docs/BENCHMARKS.md — Damage-type uniformity, style certificate, reconstruction fidelity, ablation studies, profiling
• docs/API.md — Module-level API reference with shape contracts

Citation

@misc{artrestore-diffusion,
  title  = {ArtRestore-Diffusion: Unified Style-Consistent Art Restoration with Interpretable Attribution},
  author = {TODO},
  year   = {2026},
  note   = {Generated via ml-designer pipeline}
}