Add CIELAB+Sobel physical-consistency losses for inpainting

[remyx-ai]

What does this PR do?

Adds a pair of physical-consistency losses (CIELAB color + Sobel edge) for structured-object inpainting, packaged as a new research project training script under examples/research_projects/physical_consistency_inpaint/. The losses layer onto components diffusers already ships — SD1.5-inpaint + ControlNet + IP-Adapter — reusing those existing call sites and adding src/loss.py-style CIELAB+Sobel terms. This injects semantic (text/JSON), appearance (IP-Adapter), and geometric (ControlNet) priors together with the two consistency losses to improve fidelity for data-augmentation and precise-editing use cases.

New files include examples/research_projects/physical_consistency_inpaint/requirements.txt (a dependency manifest — please confirm the pins on review).

Test results

Tests could not run in CI — the runner lacks this repo's example dependencies (a collection/import error, not a code failure). Please run the suite locally to validate.

Traceback:
/opt/hostedtoolcache/Python/3.11.16/x64/lib/python3.11/importlib/__init__.py:126: in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
examples/controlnet/test_controlnet.py:23: in <module>
    from test_examples_utils import ExamplesTestsAccelerate, run_command  # noqa: E402
examples/test_examples_utils.py:23: in <module>
    from accelerate.utils import write_basic_config
E   ModuleNotFoundError: No module named 'accelerate'
=========================== short test summary info ============================
ERROR examples/advanced_diffusion_training/test_dreambooth_lora_flux_advanced.py
ERROR examples/consistency_distillation/test_lcm_lora.py
ERROR examples/controlnet/test_controlnet.py
!!!!!!!!!!!!!!!!!!!!!!!!!! stopping after 3 failures !!!!!!!!!!!!!!!!!!!!!!!!!!!
3 errors in 0.11s

Before submitting

• Did you read the contributor guideline?
• Did you write any new necessary tests?

Who can review?

@yiyixuxu @sayakpaul

Drafted by Outrider — paper: arXiv:2609.02348.

Discovery context

Drafted by an autonomous discovery loop — Remyx ranks recent arXiv papers against this team's research interest and shipping history; Claude Code selects the candidate most directly implementable against this repo from the lookback window and drafts it.

Research interest: [crossrepo-eval] huggingface/diffusers

Implements Structured-Prior-Guided Diffusion Inpainting with Physical Consistency for Traffic Sign Augmentation.

Reference: https://github.com/52hz-whale/TrafficSignInpaint

License: Apache-2.0 (class: permissive, compat: 1.00, source: github) — safe to adopt.

Why this paper for this team: The paper is a blueprint for specialized controllable generation and inpainting using diffusers components. It addresses maintaining physical consistency in generated images — important for data augmentation where fidelity matters — by injecting semantic (JSON), appearance (IP-Adapter), and geometric (ControlNet) priors combined with physical-consistency losses (CIELAB, Sobel), showing how to combine popular control mechanisms with custom losses for domain-specific generation.

Why this candidate: The contribution is a pair of self-contained, code-backed physical-consistency losses layered onto components diffusers already ships (SD1.5-inpaint + ControlNet + IP-Adapter), so it drops in as a new research_project training script that reuses existing call sites and adds src/loss.py-style CIELAB+Sobel terms — a clean addition with a permissive (Apache-2.0) source. It was the only pool candidate combining real portable code with a verified existing call site; higher-relevance peers lacked real method code ([6],[7]) or a usable license ([10]), and the best-anchored alternative [4] fails the no-code override (training recipe, not a self-contained signal).

Suggested experiment: Use a Stable Diffusion 1.5 model and ControlNet/IP-Adapter in diffusers. Inpaint a structured object (e.g., a license plate on a car) by providing semantic guidance (text prompt), an appearance prior (IP-Adapter embedding of the desired number style), and a geometric prior (ControlNet from a simple template).

Co-Authored-By: remyx-ai[bot] <289541483+remyx-ai[bot]@users.noreply.github.com>