Nightmare

On-device ComfyUI style node graph for Stable Diffusion on Android. Local SD 1.5 and SDXL text to image, img2img and inpainting on the Qualcomm Hexagon NPU. Fully offline, no server, no account.

What it does A canvas built for a phone. Big ports, snap to connect, pinch to zoom, a node palette in a sheet. Not a desktop editor shrunk down. Real decomposition. encode_text, sample, vae_encode, vae_decode and latent_blend are separate nodes. Conditionings and latents move between them as handles that never cross the wire, so a 512 image costs about 40 bytes of JSON instead of 780 KB. Recipes to start from. Text to image, image to image, and inpainting where you paint the area to redo. Batching. Arm seed, steps, cfg, denoise or scheduler on the sampler and Run sweeps them. Two knobs at once gives you a grid. Every run is kept with the exact graph that produced it. Upscalers. RealESRGAN x4plus anime and 4x UltraSharp V2 Lite, loaded per request so they cost no process restart. Bring your own model. Fifteen checkpoints in the catalogue, or import a converted one as a zip. The app carries every HTP architecture tier and picks the build your chip can actually load. Bring your own nodes. A manifest and a script, no toolchain, no app release. Offline and private. Nothing is uploaded, there is no account, and no prompt or picture leaves the phone. The only thing that ever does is a file you explicitly share.

License

CC BY-NC 4.0