VisSlideBench evaluation
Setup
pip install -r eval/requirements.txt # judge + text matching
pip install paddlepaddle-gpu paddleocr==3.3.2 # OCR, only for run_ocr.py
export OPENAI_API_KEY=... # any OpenAI-compatible endpoint
export OPENAI_BASE_URL=... # optional
export VISSLIDEBENCH_JUDGE=gpt-4.1 # optional, default gpt-4.1
Predictions
Generate one image per case and save it as <pred_dir>/<track>/<id>.png,
where <track> is t2i, template, multi_image, or edit and <id> is the
id field. Our runs generate t2i cases at 1280x720, template and multi_image
cases at 1376x768, and edit cases at the record's width and height (the
source slide size), with seed 2233 + index for case index.
Scoring
cd VisSlideBench
for t in t2i template multi_image; do
python eval/run_ocr.py --task $t --pred_dir $PRED --device gpu:0
python eval/score_generation.py --task $t --pred_dir $PRED --workers 16
done
python eval/score_edit.py --pred_dir $PRED --workers 16
python eval/aggregate.py --pred_dir $PRED # Table 1 layout
python eval/aggregate.py --pred_dir $PRED --by source # real vs synthetic
python eval/aggregate.py --pred_dir $PRED --by prompt_type # short / long / style_json
Per-case scores go to <pred_dir>/scores/<track>.jsonl. A case that fails a
metric (missing image, no OCR text, no required text in the instruction, or a
judge error after three retries) has no value for that metric and is left out
of that metric's average.
Judge settings
| Call | Prompt | Temperature | Max tokens |
|---|---|---|---|
| Required-text extraction (CF) | in score_generation.py |
0.2 | 800 |
| Input-image preservation (multi_image CF) | prompts/content_faithfulness.txt |
0.2 | 1200 |
| Instruction Following | prompts/style_instruction_following_{text,json}.txt |
0.2 | 2000 |
| Aesthetic Quality | prompts/aesthetic_quality.txt |
0.2 | 2000 |
| Edit success + consistency | prompts/edit_evaluation_source_control.txt |
0.7 | 512 |
| Edit presentation quality | prompts/edit_evaluation_quality.txt |
0.7 | 512 |
Images sent to the judge are JPEG-encoded with the long side capped at 1,600
pixels and detail: high.