× OUTPAINTER 1.5
More room for your photo. Open a photo and add space using the four edge sliders. For a cropped person, try Extend down in small increments. The AI uses the visible photo, without a text prompt in Model A. Model B accepts a description of the finished photo. Prepare AI once, then generate. Try a new seed for a different result. Use Continue editing to make the result your next starting point. Brush over small areas to repair them. Export your result as a PNG. Klein 4B cloud mode Model C runs FLUX.2 Klein 4B on Cloudflare Workers AI. Add an AI binding named AI and an encrypted APP_ACCESS_KEY secret (16+ characters), then redeploy. No BFL key or credits are needed. Each request sends two photo references below 512 pixels and generates a 1024-pixel image at fixed 4 steps. Estimated usage is 115 neurons per call; actual billing is shown in your Cloudflare dashboard. More variations and staged strips use more calls. No automatic retry occurs.
New: local Klein in version 1.5 Select Model D to download about 5.3 GB once and run experimental image-to-image editing locally. Requires WebGPU, shader-f16 and substantial memory. Local mode never switches automatically to cloud. GPU execution and offline use need testing on your device. Model C remains available as the cloud option.
Cloud Klein Klein supports reference-image editing, not a native inpainting mask. The app provides a magenta selection guide and asks it to complete those regions, then copies only selected pixels into your working image. Unselected pixels stay unchanged at working resolution. Framing, seams, body completion and identity can still vary. Full-scene mode with one variation is the default. An accepted cloud request can use neurons even if you cancel.
Version 1.2 models A: Moebius, original weights, with sequential guidance and staged sessions to reduce memory use. B: SD 1.5 Inpainting with merged LCM adapter, prompt support and 4–8 steps. Model B is an experimental browser integration and requires shader-f16. Fewer steps do not imply a smaller model or guaranteed compatibility. All local models work at 512 pixels. Switching unloads the previous model; its cache stays.
Your Surface Laptop 4 Use an up-to-date Edge or Chrome with graphics acceleration enabled. Connect to power and close memory-heavy tabs. WebGPU availability is checked here, but memory and driver compatibility can only be confirmed by loading and running the AI. 16 GB of system memory is a more practical target; 8 GB may run out of memory. Model B can need more memory than Model A despite using fewer steps. No performance guarantee is made for your particular configuration.
Private and local In local Models A and B, AI processing happens on this device and photos are not uploaded. In optional Mode C, the working image, mask and prompt are sent to Cloudflare Workers AI after you opt in. Your original full-resolution file is not uploaded by this app. Model A downloads about 1.27 GB. Optional Model B downloads about 2.24 GB compressed and uses about 2.55 GB of browser storage. Each has a separate cache. Only the selected model is downloaded. Cached files can be reused; browsers may remove them to recover space. Offline operation requires the app and all AI files to finish caching first. This private hosted link may still need a connection for sign-in; installation and offline launch depend on the hosting environment.
The working image is reduced to fit 512 × 512. Unselected working pixels are preserved exactly; full-resolution originals are not retained in the exported file. Large extensions and hands, feet, faces and anatomy may look incorrect. No pose guidance is included in version 1.
Model B provenance ONNX pack: Chandanku/lcm-sd15-inpainting , revision a74c020. Its manifest identifies SD 1.5 Inpainting and latent-consistency/lcm-lora-sdv1-5. The adapter merge is declared by the publisher and has not been independently re-exported here. Browser downloads check file lengths and ZIP CRCs. File integrity does not guarantee model quality.
Base weights: CreativeML Open RAIL-M . LCM adapter model card: OpenRAIL++ . No image is sent to the model host. Browser GPU inference remains unverified on your Surface.
Credits Created by LTBZ. Browser inference adapted from simonw/moebius-web . Moebius by Duan, Xu and collaborators / HUST Vision Lab. Model and browser port: Apache 2.0. Runtime: ONNX Runtime Web (MIT). Notices · License .