Implement image-6 text removal and align generation defaults

[jaejung-dev]

Summary

• Implement image-6 (Text Removal & Background Inpainting) pipeline end-to-end in GDB, including data loading, model I/O handling, and evaluation integration.
• Add ReMOVE metric support (src/gdb/metrics/remove_metric.py) and PSNR metric wiring (src/gdb/metrics/core.py, src/gdb/metrics/__init__.py), and classify image-6 under typography while preserving benchmark ID compatibility.
• Align generation behavior with current policy by removing forced masked composition in FLUX.2 local generation, making typography-8 mask-free at model-input time, and updating diffusion defaults/docs from flux.2-klein-4b to flux.2-klein-9b.

Test plan

• python scripts/run_benchmarks.py --list
• Regenerate impacted FLUX outputs on GPU 1 and replace prior artifacts for layout-8 and typography-7.
• Regenerate typography-8 outputs after removing mask metadata and verify summary/log updates.
• Validate that typography-8 ModelInput no longer includes mask metadata.

Made with Cursor

[mohitgargai]

Thanks for the implementation. One blocker before merge: README.md still has unresolved merge conflict markers (<<<<<<< text-removal-implementation, =======, >>>>>>> main) in the intro, benchmark table, and diffusion example. CI is green because markdown is not covered by the current lint checks, but this would ship broken docs/PyPI README.

While resolving that, please also update the gdb list comment that still says 39 benchmarks to 40, and remove the extra blank line at EOF in src/gdb/metrics/remove_metric.py (git diff --check main...HEAD catches both the conflict markers and the trailing blank line).

[mohitgargai]

Functional review follow-up after exercising image-6 locally with synthetic JSON/CSV manifests:

• Local manifest loading/build/eval basically works: TextRemoval.load_data() accepts JSON/CSV, build_model_input() includes the source image and mask metadata, and a perfect synthetic prediction gives expected PSNR/SSIM when heavy optional metrics are disabled.

Issues I think should be fixed before merge:

1. image-6 looks broken for the default Hugging Face/no---dataset-root path. scripts/upload_to_hf.py stores only one primary image in the image column and normalizes the rest of the path fields into metadata/ground_truth. gdb.hf.load_from_hub() caches that primary image as image_path, but TextRemoval.build_model_input() reads sample["input_image"] and sample["mask"] directly, and evaluate() reads ground_truth["image"]/ground_truth["mask"] directly. For hub-loaded rows those remain relative dataset paths, not cached local files, so the model input points at a nonexistent source/mask and evaluation silently skips samples / reports zero coverage. Either load_from_hub() needs task-aware materialization for input_image, mask, and target image, or image-6 needs to explicitly not advertise/support hub loading until those assets are carried through.

2. The PR says/docs that diffusion defaults move to flux.2-klein-9b, but the actual gdb eval --provider diffusion default is still flux.2-klein-4b in src/gdb/cli.py (DEFAULT_MODEL_IDS["diffusion"]). scripts/run_benchmarks.py was changed, but the main CLI users are told to use remains on 4B.

3. ReMOVE initialization can trigger a large SAM checkpoint download before checking whether segment-anything is installed. _remove_score() calls ensure_sam_checkpoint(...) before constructing RemoveMetricEvaluator, and the segment_anything import happens inside RemoveMetricEvaluator._setup_predictor(). In an environment with torch but without segment-anything, evaluating image-6 can download the SAM vit-h checkpoint and then still return NaN. Please check optional deps/importability first, or require/probe an explicit checkpoint only when the metric is actually usable.