How AI Understands Images
What a model actually takes from an uploaded image, and what it will never copy exactly.
On this page
When you upload an image in the Studio, the model does not paste it into your design. It reads the image as description — subject, palette, lighting, composition, texture — and uses that description to steer the next generation. This page explains what survives that translation and what does not.
Overview
An uploaded image is interpreted, not inserted. The model builds an internal reading of what the picture contains and how it looks, then generates new artwork influenced by that reading.
Strong signals come through clearly: overall palette, lighting direction, mood, level of realism, and rough composition. Weak signals rarely survive: exact faces, precise logos, small text, and one-off details in the background.
If you need an element reproduced exactly rather than interpreted, that is a layer job, not a reference job. See Layers and Add a Logo.
Why this matters
Most disappointment with reference images comes from expecting a copy and receiving an interpretation. Knowing the difference tells you instantly which tool to use for what you want.
It also improves your results: once you know the model reads mood and palette most reliably, you can choose reference images for those qualities instead of for their content.
What transfers and what does not
| Quality | Transfers? | Notes |
|---|---|---|
| Color palette | Reliably | One of the strongest signals in any reference. |
| Lighting and mood | Reliably | Direction, softness, and time of day carry over well. |
| Art style | Usually | Painterly, photographic, and graphic styles read clearly. |
| Composition | Partly | Approximate placement, not pixel-accurate framing. |
| Exact faces | No | Expect a similar look, not the same person. |
| Logos and text | No | Add these as layers instead. |
Step-by-step guide
Step 1: Choose a reference for one quality
Decide before uploading whether you want its palette, its lighting, or its style. References chosen for a single quality work far better than ones chosen for everything at once.
Step 2: Upload it in the Studio
Use the reference control in the Studio to attach the image to your generation.
Open the AI StudioStep 3: Write the prompt anyway
The reference steers; the prompt still decides the subject. A reference with no prompt gives the model too much freedom.
Step 4: Name the quality you borrowed
If you uploaded a reference for its palette, say "same teal and amber palette" in the prompt. Reinforcing in words strengthens the effect.
Step 5: Generate and compare
Look at whether the borrowed quality came through. If it did not, the reference is probably too busy — try a simpler image.
Step 6: Add exact elements as layers
Logos, names, and precise shapes go on top as layers after the artwork is right.
Best practices
- Use high-resolution, clean references. Compressed screenshots transfer noise, not style.
- One reference at a time. Multiple references average into something neither of them looks like.
- Crop the reference to the part you actually care about before uploading.
- Reinforce the reference in words so the prompt and the image agree.
- For a wide mousepad, prefer wide reference images. Portrait references fight the canvas.
Common mistakes
| Mistake | Why it happens | What to do instead |
|---|---|---|
| Expecting an exact copy of the uploaded image | The model interprets rather than reproduces. | Use layers for anything that must be exact. |
| Uploading a busy collage | Too many competing signals average out. | Upload one clean image with a clear dominant look. |
| Uploading a portrait reference for a 930 mm wide pad | The framing has nowhere to go on a very wide canvas. | Crop to a wide aspect first, or describe the composition in the prompt. |
| Uploading someone else's artwork | It may infringe copyright and breaks the acceptable use rules. | Use your own work or licensed material. See Copyright Explained. |
Frequently asked questions
Will my reference image appear in the final design?
No. It influences the generation but is never pasted in. Anything that must appear exactly should be a layer.
Can I upload a photo of myself and get a portrait?
You will get something in a similar spirit, not a likeness. Image models do not reproduce specific faces reliably.
How many references can I use?
Use one strong reference. Additional images tend to cancel each other out rather than combine.
Does a reference image cost extra credits?
Generation cost depends on the model you choose. The current cost is shown in the Studio before you generate — see Generation Costs.
Why did the colors transfer but not the composition?
Palette is a much stronger signal than framing. State the composition in the prompt to reinforce it.
Can I use a reference to keep a character consistent across designs?
It helps, but it is not exact. See Character Consistency for the full technique.
What image formats work best?
Standard high-quality JPEG or PNG files. Avoid heavily compressed images and screenshots of screenshots.
Related articles
- How Image References Work — Use an uploaded image to steer style, palette, and mood — and know where the limits are.
- Add a Reference — Use a reference image to anchor the style or subject of a generation.
- Layers — How generated elements stack, and how to arrange them into a finished design.
- How AI Understands Prompts — What actually happens to the words you type, and why small wording changes can move a generation a long way.
- Reference Image Workflows — Reference Image Workflows in the Design Academy section of the CursorCulture documentation.
Related blog articles
- Prompt writing — Prompt breakdowns, before-and-after rewrites, and prompt patterns that hold up in print.
- Design inspiration — Desk setups, art directions, and design walkthroughs from the community.
Need more help?
- Contact support — for orders, billing, and account questions.
- Join the community — ask other creators and share what you make.
- Chat for Ideas — brainstorm a direction before spending credits.
- Open the AI Studio — put any of this into practice right now.