ForgeBench

Evaluation code for FORGE-Bench, a benchmark of industrial image-to-video generation.

The runner samples each task video at fixed points, collects judge and CV evidence, and writes per-sample scores and model summaries. The scoring method describes the six axes, event coverage, operator caps, and aggregation.

Code

Path	Purpose
eval/run_eval.py	Run evaluation for a model and task manifest
eval/video_protocol.py	Select frames for judges and CV operators
eval/operator_evidence.py	Measure local changes, temporal breaks, and joint motion
eval/llm_judge*.py	Anthropic and OpenAI-compatible judge backends
scoring/per_sample.py	Score one task
scoring/aggregate.py	Aggregate tasks and estimate confidence intervals
scripts/combine_eval_shards.py	Merge evaluation shards against the task manifest
scripts/reaggregate_cached_results.py	Rescore saved judge results

Run

Install Python 3.10 or newer and the dependencies:

python -m pip install -r requirements.txt

Put videos under one directory, named {task_id}.mp4. Category subdirectories are supported. The task manifest should contain paths to the corresponding reference images.

python -m eval.run_eval \
  --model my_model \
  --video_dir /path/to/videos \
  --samples_json /path/to/video_generation_500_samples.json \
  --output_dir results \
  --llm_provider openai_compat

Set OPENAI_COMPAT_API_KEY for an OpenAI-compatible endpoint, or use --llm_provider anthropic with ANTHROPIC_API_KEY. Results are written to results/my_model/ as task records, per_sample.json, aggregate.json, run_metadata.json, and report.json.

The task set and generated videos are available from the dataset release.

Tests

python -m pip install pytest
python -m pytest -q