Images

Read this when you attach or paste images, use a vision model, or hit image size limits.

What the read tool does with images

The read tool attaches images to the conversation. When the active model supports vision (claude, gpt-4o/4.1/4.5/5, o3/o4, glm-5.3-flash/flashx, glm-5v …), a read of a jpg/png/gif/webp sends the image inline — the model sees it. Detection is by content (magic bytes), never by file extension.

  • Text-only models still read successfully: the image is replaced by a placeholder at request time, and the tool notes the omission.
  • BMP converts to PNG (EXIF orientation baked in); jpg/jpeg mime labels normalize.

Attach images

  • Print mode: ink @shot.png @notes.txt "what is this" — text files embed as <file> blocks, images attach to the first message through the same processor.
  • TUI: Ctrl+V pastes the clipboard at the cursor, in this order: copied files (e.g. a Finder Cmd+C) paste as their real paths — an image file never pastes its file icon; image data (a screenshot) is written to a tmp file and that path is inserted; otherwise plain text pastes. Use Ctrl+V — the terminal owns Cmd+V, and only text can travel that channel. Read via osascript (macOS), wl-paste/xclip (Linux), PowerShell (Windows — best-effort readers, not yet verified on that platform). A line that starts with an absolute path (a pasted tmp path, for example) submits as a normal message, never as a slash command. You can also drag files into the terminal or type the path directly.

Oversize handling (resize ladder)

Oversized images (>4.5 MB encoded) are resized automatically through the photon (Rust/WASM) ladder — 2000×2000 cap, PNG/JPEG candidates, quality steps, dimension decay — with a coordinate-mapping note telling the model how to map back to original pixels. Set images.autoResize: false in ~/.ink/settings.json to ship original bytes (oversize then gets a teaching note instead).

Vision models per provider

  • z.ai Coding Plan: glm-5.3-flash is the vision model of the plan (GLM-5.3 itself is text-only).
  • Anthropic claude models, OpenAI gpt-4o/4.1/4.5/5 and o3/o4 support images natively; see the model picker (/model, Ctrl+L) — the catalog entry lists input: image capability.