Check reference support
GPT-Image-2, Nano Banana 2 and Seedream 5.0 Lite support reference-led image work. The number of references and settings vary.
Your image. A new setting.
Bring a portrait or product photo. Describe the changes you want, then review the result against your reference.
Reference images guide the result. Check faces, labels and fine details.
See the edits
Keep a reference. Change the scene.
Compare an existing portrait with its generated scene, then explore a second reference across three settings. Results can change identity details, so review each output.
ReferenceGenerated scene
100%
Be specific about the change.
A clear source and a focused edit request give you a useful starting point. Review the output before using it.
- 01
Upload a clear reference
Choose a portrait, product photo or other image with visible subject details.
- 02
Describe the edit
Name what you want to change, such as the setting, wardrobe or lighting, and what should stay recognizable.
- 03
Compare with the source
Check each result against your reference. Keep a useful variation or adjust the prompt for another pass.
Model guidance
Choose by the work you need.
Compare the available model settings and credit estimate before each run. Access depends on your plan.
Describe what matters
Name the features you want to keep and the changes you want to make. A reference guides the model; it does not guarantee exact preservation.
Check before publishing
Compare faces, product shapes, logos and labels against the source. Refine or reject outputs that change important details.
Starting from a photo?
Use a reference when the existing subject matters. Text to image is useful for new concepts; image to video adds motion to a still.
Best when you need
- New scenes from an existing subject
- Identity continuity from a reference
- Controlled wardrobe or background changes
- Reference-guided restyles and refinement
Consider another starting point for
- Prompt-only ideation from scratch
- Outputs with no reference continuity
- Motion from a still frame
- Large coordinated output runs
Keep creating
Explore related tools, creative examples and more starting points.
Image-to-image FAQ
How much of the original image can I preserve?
That depends on the strength of the reference and how aggressive the requested change is. Image-to-image works best when the prompt targets specific changes instead of rewriting the entire scene.
What kind of reference image works best?
Use a clean, sharp image with readable subject detail. The stronger the source frame, the easier it is to preserve composition and identity where needed.
Can I edit the same image multiple times?
Yes. A common workflow is to keep the best variation from one pass, then use it as the starting point for the next refinement round.
Do I need a paid plan?
Generation requires enough credits and access to the selected model. Accounts without a paid plan can use Lite-supported models while they have spendable credits. Other models require an eligible paid plan. Review the credit estimate before submitting.
HummingBytes