Vision and Images

Vision & Images

Real Chat can work with images two ways: process them locally (no model needed) and have a model see them.

Attaching an image

Paste (Ctrl/Cmd+V), drag-drop, or use the Image button in the chat input. A chip appears, and a notice confirms which model will read it.

Attached images are sticky: the chip stays so follow-up questions about the same picture keep reaching the vision model. Tap × on the chip to remove it and return to normal text chat.

How the model sees an image

• If your active model is natively multimodal (most big OpenRouter models, marked with an image icon in the model list), the image goes straight to it.
• If your model is text-only (e.g. DeepSeek), configure a separate Vision provider in Settings (off by default): paste a base URL, key and model — for example the Gemini free tier, or an OpenRouter vision model. While an image is attached, every turn is answered by the vision model and labelled "Vision".

See and act in one turn

The vision model runs the full tool loop — it doesn't just describe an image, it can act on it in the same message. For example:

"Save this screenshot to Attachments and tell me what the text says."

…will save the real picture into your vault (via save_attached_image, returning a clickable ![[link]]) and read the text back. It can also create/edit notes, search your vault, and use any other tool based on what it sees.

save_attached_image

Writes the actual pasted/attached image bytes into your vault (not a description). Optional path (defaults to a name in your download folder; the extension is inferred from the image type) and index (which attached image, if several). Only offered to the model when an image is actually attached.

Local image processing (no model)

Independent of vision, the model can call local image tools:

• process_image — compress to a target KB, resize, convert format (JPEG/PNG/WebP), crop, rotate, flip.
• batch_process_images — apply the same processing to a whole folder.
• image_info — report dimensions and size.

Ask things like "compress this image under 200 KB" or "convert the PNGs in Attachments to WebP".

Notes

• Vision-provider cost is not tracked against DeepSeek/OpenRouter pricing (token counts are shown when the provider returns them).
• Pasted images live in memory for the session and aren't persisted to disk unless you (or the model) save them.