Skip to content
Build Perch
Browse guides

How to generate images from a reference image

Short answer: Attach or drag an image into the chat, then tell the AI what to change: it feeds your image straight into the image model as a reference and generates a new picture built from it. Image generation runs on OpenRouter’s models, so it needs an OpenRouter API key, but the chat itself can be on any provider.

Steps

  1. Make sure an OpenRouter API key is configured under Integrations. Your chat can be on any provider (OpenRouter, local Ollama, or Ollama Cloud); the image request routes to OpenRouter either way. You don’t have to pick an image model yourself (Build Perch falls back to a low-cost default), but you can set one under Chat Configuration → Models & Search if you want a specific one.
  2. Attach your reference image: drag a file into the prompt box, paste one from your clipboard (Ctrl+V), or drag one in from the workspace’s Files panel.
  3. Describe the change in the same message: “turn this into a watercolor painting,” “same pose, snowy forest background,” “make the jacket blue.” The AI reuses your attached image as the input rather than describing it back to itself, then generates a new image based on it.
  4. The result appears inline in the chat as a new image. Ask for another pass if it’s not right; each request starts fresh from your original reference plus your latest instruction.
  5. You can also point at an image already saved in the workspace instead of one you just attached: name the file, and the AI uses that as the reference.

Notes & gotchas

  • Like all AI-generated media (speech, music, and video included), this runs on OpenRouter’s models and sends your reference image there, even if the conversation is on a local model. What stays local either way is media editing (trimming, reframing, captioning existing video/audio) and speech-to-text.
  • You can sketch on an attached image first (the in-app sketchpad) to mark up what should change, then ask the AI to generate from your markup.
  • Mention an orientation if it matters (“wide,” “tall,” “square”); otherwise the model picks.
  • To save or reuse the result, drag it into the workspace, into a card, or right-click it and choose “Save to folder” to write it elsewhere on disk.