Reve details image API for create, edit and remix after 2.0 launch

Why it matters

Reve is trying to compete on controllability, not just image quality. If its layout system works as claimed, it targets a gap that still limits AI image tools in production workflows.

Reve, the image-generation startup, launched Reve 2.0 in a thread on X, pitching the release as a 4K image model built around precise layouts rather than only text prompts. The company has also published API documentation for create, edit and remix endpoints, giving developers a more concrete view of the product surface that Reve had said was coming "in a week or so."

Reve's core model claim is that each image is segmented and labeled, making regions and elements addressable for editing. In the thread, Reve said images are "represented as code," so parts of an image can be edited and manipulated after generation. That is the company's answer to a persistent weakness in image models: they can generate impressive frames, but fine-grained control often breaks when users ask for specific placement, edits or composition changes.

The API docs list three main endpoints: POST /v1/image/create for text-to-image generation, POST /v1/image/edit for modifying an uploaded base64 reference image with text instructions, and POST /v1/image/remix for combining prompts with one to six reference images. All three require bearer-token authentication.

The create endpoint accepts a prompt of up to 2,560 characters, aspect-ratio controls, model-version selection and optional postprocessing. The edit endpoint requires an edit_instruction and reference_image, while the remix endpoint requires a prompt plus reference images and can use XML img tags to refer to specific images by index. The docs list supported aspect ratios including 16:9, 9:16, 3:2, 2:3, 4:3, 3:4 and 1:1.

Reve's API defaults to JSON responses with a base64 PNG and metadata, but developers can request direct image/png, image/jpeg or image/webp bytes through the Accept header. The docs also describe postprocessing options including upscaling, background removal, fit-to-dimension resizing and effects such as cmyk_halftone. A test_time_scaling parameter can spend more model effort for additional credits, according to the docs.

For operations teams, the API exposes request and billing metadata through fields and headers, including request IDs, content-violation flags, model version, credits used and credits remaining. The documented error surface includes invalid parameters, missing or invalid API keys, insufficient credits, unprocessable inputs, rate limits and internal server errors.

Reve's leaderboard framing now has an outside source: Arena.ai, the model-evaluation leaderboard operator, posted that Reve 2.0 is in the top two on its Text-to-Image leaderboard. Reve's own launch thread said that placement puts its research lab ahead of Nano Banana 2 and GPT-Image-1.5. As with any public leaderboard, the ranking is a snapshot rather than a substitute for workload-specific testing, but it gives Reve a third-party benchmark to attach to the API launch.

The product is live at reve.com.