CLI
CLI Options Reference
Every VisionSqueezer CLI flag, including --max-tokens token budgets, --model provider profiles, --format, --quality, --smart-crop, --json and --dry-run.
Reference
| Flag | Description |
|---|---|
--model <name> | Target model alias; core profiles plus popular families such as kimi, glm, pixtral, gemma, internvl, llava, and minimax. |
--format jpeg|webp|avif | Output encoding. AVIF is the default for new pipelines. |
--quality 1-100 | Output quality (default 75). |
--auto-quality 0.0..1.0 | Binary-search quality in [40,95] to hit an SSIM target. |
--smart-crop | Edge-energy (Sobel-lite) crop. Best for photographic content. |
--ops '<JSON>' | Execute Sandbox operations. |
--output <path> | Custom output destination (single-file mode). |
--output-dir <path> | Output root (batch mode, mirrors structure). |
--recursive | Walk subdirectories in batch mode. |
--max-tokens <N> | Token budget. The image is downscaled until the target model's token estimate fits (Claude when no --model is set). 0 disables. The MCP server defaults to 1600; the CLI has no cap unless you pass this. |
--max-tiles <N> | Hard cap in the target model's own unit (28px patches for Claude, tiles for Gemini/Llama). Prefer --max-tokens (token budget guide). |
--mode auto|standard|ocr | auto (default) behaves like standard and keeps colour. ocr is the only mode that converts to black and white. |
--json | Machine-readable JSON output (single-file or batch aggregate). |
--dry-run | Run the full pipeline without writing to disk or updating the stats DB. |
Quality vs. auto-quality
--quality is a fixed encoder setting. --auto-quality is smarter: it binary-searches the quality range and lands on the smallest file that still passes a perceptual SSIM threshold.
Terminal
# Fixed quality
vision-squeezer image.png --quality 80
# Target perceptual fidelity, minimize bytes
vision-squeezer image.png --auto-quality 0.95
Use --auto-quality 0.95 when bandwidth matters and you can tolerate slight perceptual loss.
Crop strategy
- Default (corner-tolerance): strips solid-color padding. Best for screenshots with uniform borders.
--smart-crop: keeps the high-information region using gradient energy. Best for photos and saliency-heavy content.
Estimate before writing
Terminal
vision-squeezer image.png --model gpt6 --json --dry-run
--json --dry-run reports token impact with zero side effects — ideal for pipelines that gate on savings before committing.