Prompt file imported from xsscx/research (.github/prompts/image-fuzzer-quality.prompt.md). Copyright stays with the author.

Image Fuzzer Output Quality Analysis

Context

xnuimagefuzzer generates fuzzed images via 15 CGBitmapContext configurations and 22+ output formats (PNG, JPEG, TIFFx5, GIF, BMP, HEIF/HEIC, WebP, JP2, DNG, TGA, ASTC, KTX, PDF, ICNS, EXR, plus ICC variant outputs). Quality analysis validates that outputs have sufficient visual diversity for effective fuzzing.

Expected Output

Each --pipeline run produces images across 4 phases:

• Clean: 15 contexts x base formats = ~90 files
• Format: 15 contexts x 22+ multi-format outputs = ~330 files
• Fuzz: 15 contexts x 8 mutation strategies x format outputs = ~2,000+ files
• ICC: Up to 4 ICC variants per TIFF/PNG save (real, stripped, mismatched, mutated)
• Combo: Combined mutation + ICC + multi-format

Total output per full run: 2,000+ images (varies by FUZZ_ICC_DIR availability).

Quality Metrics

1. File validity -- Each output must be a valid image file (not truncated/empty)
2. Dimension diversity -- Mix of sizes from 16x16 to 512x512 (minimum 16x16 enforced)
3. Color diversity -- Each image should have >=5 unique pixel values
4. Format coverage -- All major output formats present (PNG, JPEG, TIFF, GIF, BMP, HEIF minimum)
5. No duplicates -- SHA256 dedup shows 0% duplicate rate
6. No crashes -- Zero ASAN/UBSAN errors during generation
7. ICC variant coverage -- Stripped, mismatched, real ICC, and mutated ICC variants present
8. TIFF subformat diversity -- Uncompressed, LZW, PackBits, JPEG-in-TIFF, Deflate present

Scoring (target: 90+/100)

• 30 pts: File validity (all files parse as valid images)
• 15 pts: Dimension diversity (>=5 distinct sizes)
• 15 pts: Color diversity (>=80% files have >=5 unique colors)
• 15 pts: Format coverage (>=10 distinct formats present)
• 10 pts: ICC variant coverage (all 4 types present)
• 10 pts: No duplicates
• 5 pts: TIFF subformat diversity (>=3 compression types)

Known Limitations

• 32BitFloat and HDR float variants produce near-black images (float->8-bit clamping)
• 1BitMonochrome may not generate on macOS 15 (CGBitmapContext returns NULL)
• Structure-aware mutations add PNG chunk-level corruption (IHDR, PLTE, tRNS, sRGB, gAMA, cHRM)
• HEIF/HEIC requires hardware encoder (may not be available in CI)
• ICC real/mutated variants require FUZZ_ICC_DIR environment variable

Analysis Steps

1. Download fuzzed-images artifact from CI
2. Count files per format: find . -name "*.png" | wc -l (repeat for jpg, gif, bmp, tiff, heif, webp, jp2, etc.)
3. Validate with: file -b "$f" -- should report image type, not "data"
4. Check dimensions: identify "$f" or exiftool -ImageWidth -ImageHeight "$f" or sips -g pixelWidth -g pixelHeight "$f"
5. Check color diversity: convert "$f" -unique-colors -format "%c" histogram:info:
6. Check ICC variants: find . -name "*icc*" -o -name "*no_icc*" -o -name "*mismatch*" | wc -l
7. Check TIFF compression: tiffinfo "$f" | grep Compression for each TIFF variant
8. Compute score against the 8 metrics above

See Also

• mac-catalyst-ci.prompt.md -- Mac Catalyst CI debugging
• improve-fuzzer-coverage.prompt.md -- Coverage gap analysis
• corpus-management.prompt.md -- Corpus storage operations