
AI generator
Image to Image: Start from a Picture, Not a Blank Page
Some ideas are easier to show than to describe. Upload a rough sketch, a phone photo or an earlier generation, then say what to keep and what to change: the layout, the character, the palette or the whole style.
Made with OpenDream
See more in the galleryHow it works
- 01
Upload one or more references
Add a sketch, a photo or a previous result. FLUX.2 and Seedream 5 accept reference images, and Nano Banana 2 takes up to 14, which helps when combining a character, an outfit and a location.
- 02
State what must stay
Be explicit: 'keep the pose and camera angle', 'keep the face and hairstyle' or 'keep the room layout'. Without this, the model decides for you.
- 03
Describe the transformation
Then name the change, like 'render as a sunlit watercolor' or 'turn the pencil sketch into a finished 3D render in brushed aluminum'. Compare results and tighten the instruction.
Why create image to image with OpenDream
Sketch to finished image
A few lines on paper can pin down the composition. The reference holds the layout while the prompt supplies materials, lighting and style.
Restyle a photo
Turn a holiday snapshot into a paper-cut scene, a clay diorama or an oil painting while the place stays recognizable.
Consistent characters
Reuse one character reference across several images to keep a face, outfit or mascot stable through a comic, storyboard or campaign.
Several references at once
Nano Banana 2's multi-image input lets you bring together a product, a person's pose and a background from separate sources in a single generation.
Getting predictable results from reference images
The reference and the prompt share control, and results improve when they agree. If your sketch shows a figure facing left and the prompt says 'looking over her right shoulder at the camera', the model has to choose. Write the prompt as if describing the reference with only the intended changes applied.
Clean references transfer better. A sketch with confident outlines and clear shapes beats one covered in construction lines; a photo with even lighting is easier to restyle than one with blown highlights. For character consistency, choose a reference where the face is well lit and unobstructed, and where the full outfit is visible if clothing matters.
With several references, give each a job in the prompt: 'use the first image for the character, the second for the jacket, the third for the street'. Models handle multiple inputs better when told what each is for, and it stops features from one picture bleeding into another.
Only upload images you have the right to use. Your own sketches, photos and earlier generations are ideal. Avoid photos of other people without their consent, and do not upload copyrighted characters or artwork as a shortcut to recreating them.
Prompt examples
Sketch to product render
Using the reference sketch for shape and proportions, render a finished desk lamp concept in brushed aluminum and matte white, soft studio lighting, seamless light grey background, three-quarter view, realistic 3D render
The sketch controls form while the prompt controls materials and light, so the two never compete.
Try this promptPhoto to paper-cut
Keep the composition and main buildings of the reference photo, restyle it as a layered paper-cut illustration in five layers, warm evening palette of coral, gold and deep blue, soft shadows between the layers
Saying what to keep before naming the new style produces a recognizable scene with a clearly different look.
Try this promptSame character, new scene
Same character as the reference image, same face, hairstyle and green raincoat, now standing at a rainy bus stop at night, neon reflections on wet pavement, cinematic medium shot, shallow depth of field
Listing the identity features to preserve helps the model hold the character steady while everything else changes.
Try this promptMulti-reference composite
Use the first image for the woman's face and hair, the second for the knitted cream sweater and the third for the autumn park setting; she sits on a bench reading, soft overcast light, natural candid photograph
Assigning a role to each reference keeps traits from mixing between images.
Try this promptRoom restyle
Keep the exact layout, windows and furniture positions of this living room photo, restyle it as a warm Scandinavian interior with light oak floors, a linen sofa, paper lantern lights and plants, late morning daylight
Locking the layout while changing materials is the quickest way to explore interior options for the same space.
Try this prompt
Start from a template
Frequently asked questions
- Which models accept reference images?
- FLUX.2 and Seedream 5 accept references, and Nano Banana 2 accepts up to 14 in one generation. Smart mode can pick a suitable one when you upload a picture.
- Is this the same as editing?
- They overlap. Editing makes a targeted change to one picture, like swapping the background or adjusting the light. Working from a reference uses the picture as a guide for a new result, which can differ a lot from the original.
- How do I keep a character looking the same?
- Reuse one clear reference of the character, describe the features that must not change, and keep the style words identical between generations.
- Are my uploads public?
- No. Uploads and results are private by default and only appear publicly if you explicitly publish them.
Keep exploring
Ready to create?
Pick a plan and you only spend credits on images you actually receive.
Start creating


