PhotoLoop
AI Image to Image Generator
Upload a photo, describe what should change, and get a new version back — then send it straight into video if you want.
Your photo guides, it does not survive
The reference anchors composition, proportion and broad colour, but every pixel in the output is generated. This is what lets the model relight a scene or change a material, and equally why fine detail drifts. Faces and small lettering are where you will notice it first, so those are what to check before calling a result done.
Describe the change, not the picture
The most common mistake is re-describing what is already visible. The model can see your photo; spending the prompt on its contents wastes the instruction. Say what should be different — the palette, the medium, the light, the season — and let the reference carry everything you did not mention.
Contradictions get resolved arbitrarily
Asking to keep a subject exactly as they are while also changing their pose sets two goals that cannot both hold. The model will satisfy one and drop the other, and which one is not predictable. When a result ignores half your instruction, the fix is usually to split it into two passes rather than to rewrite it more forcefully.
Restyle, then animate, without leaving
This is why image and video live in one generator here. Rewrite a photo, then flip the output to video and use the new version as your first frame. Two credits for the image, then whatever the video model costs — 38 credits on Wan 3.0 at 720p for five seconds, up to 266 on Seedance 2.5 at 1080p.
- 2 credits an image with GPT Image 2
- Reference guides composition and colour
- Result can become the first frame of a video
- Upload and result deleted after 24 hours
Questions
How much of my original photo survives?
More than text-to-image and less than you probably expect. The reference anchors composition and broad colour, but the model still regenerates every pixel. Faces and small text are where drift shows first, so check those before deciding a result is usable.
How is this different from a filter?
A filter maps your pixels to new pixels. This regenerates the image from scratch using yours as guidance, which is why it can add and remove objects, change materials and relight a scene — and also why it will not preserve fine detail the way a filter does.
What should the prompt describe?
The change, not the whole picture. If your photo is already a portrait by a window, say what should be different — the light, the palette, the medium — rather than re-describing the portrait. Restating what the model can already see wastes the instruction.
What does it cost?
Two credits an image with GPT Image 2, the same as text-to-image. One credit equals one cent of supplier cost.
Can I take the result straight into video?
Yes, and that is the main reason this lives in the same generator rather than a separate tool. Restyle a photo, then switch the output to video with the result as your starting frame — no re-uploading, no second tab.
What happens to my uploaded photo?
It goes to the model provider to generate from, and both your upload and the result are deleted after 24 hours. We never use either to train models. The photo does reach our provider, so we only make that promise for ourselves.
Why did my result ignore part of the instruction?
Usually because the instruction contradicted the reference. Asking to keep a subject identical while also changing their pose gives the model two goals it cannot satisfy at once, and it will pick one. Split contradictory changes across two passes.