Image to Image AI Editor & Retouch Workspace
Upload a photo, write an edit instruction, and restyle it with Flux 2, Qwen, Nano Banana, Midjourney, and GPT-4o. Retouch, relight, or swap backgrounds.
How image to image works on SeedVR2
Image to image edits a picture you already have: retouch a portrait, change an outfit or background, or relight a product shot from a text instruction while the original subject stays recognisable. The uploaded image is the anchor, and the prompt describes only what should change. Because the subject, pose, and lighting come from your file rather than from text, edits stay close to the real photo instead of drifting into a newly invented scene.
From source image to edited result
Each run pairs your uploaded image with a written instruction, sends both to the selected model, and returns an edited copy that keeps the original on file. The source is never overwritten, so the untouched photo is always available for another attempt with different wording or a different model. One variable at a time, such as the outfit or the background, makes it easy to see which instruction produced which change and to rebuild a result you liked.
1. Write the edit instruction
Describe the change, not the whole picture: swap the outfit, replace the background, add rim light, remove an object. Naming only the edit keeps the face, pose, and composition close to the source image instead of drifting into a new photo.
2. Screening and validation
The uploaded file and the instruction are checked before the model runs. Unsupported formats, oversized files, or content that breaks the acceptable-use policy are rejected up front, so a blocked request is stopped before any credit is deducted.
3. Pick a model and see the price
The selector shows every model with its per-image price, and the workspace repeats that price before you run it. Flux 2, Qwen, Nano Banana, Midjourney, and GPT-4o are all available here, from fast low-cost edits to slower precision models.
4. What to upload
This route is built around a base image. Upload one to five PNG, JPG, or WebP files, at about 10 MB each, and select the one the edit should start from. Sharper, larger sources give the model more to keep, while a soft or heavily compressed source limits how clean the result can be.
5. How long a run takes
Edit time depends on the model and the image size. In the current catalog a default run is typically reported between 5 and 18 seconds, and lighter models return first. You can keep working in the workspace while the edit finishes in the background.
6. When a different route fits better
Use image to image when the picture should change but stay recognisable. If the photo only needs to be larger, sharper, or cleaner, the seedvr2 upscaler restores it without altering the content, and text to image is the right start when there is no source picture at all.
Open the upscalerWhat image to image costs
Every model in the selector is billed per image at its default entry settings, from 20 to 244 credits. The price of the model you pick is shown in the workspace before you run it, and a low-cost model can validate an instruction before you spend on a precision model.
| Model | Credits per image |
|---|---|
| Nano Banana | 30 |
| Nano Banana Pro | 240 |
| Flux 2 Pro | 49 |
| Flux 2 Flex | 98 |
| Flux 2 Dev | 20 |
| Qwen | 50 |
| Midjourney | 244 |
| GPT-4o | 40 |
Image to image FAQ
Short answers about uploads, model choice, credits, and failed runs.
How are credits charged for an image to image run?
A run bills the selected model price once, taken from the table above, when you start it. The price reflects the model default settings, so a larger output or a heavier model costs more, and a run that never starts is never charged.
What happens if I run out of credits?
The edit does not start and nothing is charged. Add a credit pack or a subscription on the pricing page, then come back to the same image and instruction; the workspace keeps your upload and prompt while you top up.
Why was my image or instruction rejected?
Uploads and prompts are screened against the acceptable-use policy before the model runs. If either is blocked, replace the image or rewrite the instruction and resubmit; the rejected request is not charged because the edit never began.
What if the provider times out or the edit fails?
Confirmed system or provider failures are covered by the refund policy, which explains how a job is reviewed and repaid. You can also retry the same instruction, or start from a smaller source image and a lighter model to lower the chance of another failure.
Do I need to upload an image?
Yes. Image to image always works from a base image, and the editor accepts up to five files per run so you can pick the cleanest source. If you have no picture yet, start on text to image instead.
Can I keep editing the result?
Yes. Feed the edited image back in as the new source and describe one further change, which keeps each step small and predictable. Change a single variable per run so you can tell what caused the difference.