nightmare-mobile

Introduction: On-device ComfyUI style node graph for image/video diffusion models on Android. Local SD1.5, SDXL, Flux, Z-image, Anima, Qwen Image 2.1, Krea 2, MOSS Image, Ideogram 4, inpainting, outpainting, controlnet, LoRA, IPAdapter, LivePortrait and WAN 2.2 14B video on the Qualcomm Hexagon NPU. Fully offline, no server, no account.
More: Author   ReportBugs   
Tags:

ComfyUI-style image and video generation on an Android phone, running on the Snapdragon NPU. FLUX.2, Z-Image, Qwen Image, SDXL, Ideogram 4, Wan 2.2 video and 50 more models, as a node graph made for touch. No server, no account, no cloud.

Download the APK · Demo video · Telegram group · All releases · User guide · Report a problem

21 text to image models racing on one Galaxy S25 Ultra, from Anything V5 in 3 seconds to Qwen Image 2.1 in 4 minutes 48; links to the 65 second demo video

Why it is worth a look

  • 21 models, one phone, real times (a Galaxy S25 Ultra, above): SD 1.5 in 3 to 7 s, SDXL at 1024x1024 in 6 to 47 s, NitroFusion 2 and FLUX.2 Klein 4B in under 25 s, Z-Image Turbo in under 2 min, Qwen Image 2.1 and Ideogram 4 (readable text in pictures) in under 5.
  • Video too: Wan 2.2 makes a 5 s clip at 832x480 in under 3 minutes, MiniMax H3 makes clips with speech, sound and music, and Motion transfer moves your picture the way a video you pick moves.
  • A real node graph, made for fingers: ComfyUI's nodes and wires, with a one-step Default view when you just want a picture. Every result reopens as the flow that made it.
  • Your own models: a CivitAI FLUX.2, Z-Image, Qwen Image, Krea 2, Anima or MOSS-Image fine-tune (.safetensors or GGUF) runs as it is, with no conversion. Search and download LoRAs from CivitAI and Hugging Face inside the app.
  • The desktop toolbox: inpaint and outpaint, tap to select (Segment Anything), ControlNet, IP-Adapter, LoRAs, hires fix, Fix faces and hands (ADetailer), editing by instruction (FLUX.2 Klein, Qwen Image, Fibo 2.0), upscalers, seed and setting sweeps as grids, and A1111 parameters in every saved PNG.
  • Offline and private: nothing is uploaded. A local API (off by default) also speaks AUTOMATIC1111's /sdapi/v1/txt2img, so other apps can use your phone as their image generator.

Nine screens of the app on a phone: a picture from a text prompt, the node graph, a MiniMax H3 clip with sound, painting an inpaint mask, a render in progress, sending a picture to another flow, the Models and Flows pages, and Results

Everything it does, with measurements, and the user guide, which explains each part.

What we built

The app started from Local Dream's NPU server (SD 1.5 and SDXL) and its early DiT engine (FLUX.2 Klein and Z-Image). The rest is Nightmare's own work, and every engine change is public in backend-patches/:

  • Most of the models, brought to the phone here: Qwen Image 2.1, Krea 2, Anima, MOSS-Image, Iris, Ideogram 4, FLUX.3 Klein, Fibo 2.0 Edit and FLUX.1 dev, and all of the video: Wan 2.2 5B and 14B, motion transfer (Wan 2.2 Fun Control) and MiniMax H3 with sound. Where the engine lacked a model or ran it wrong on the NPU, we ported or fixed it: 85 patches to stable-diffusion.cpp and ggml's Hexagon backend.
  • A 14B video model on a 12 GB phone: Wan 2.2 14B's two experts stream from storage instead of sitting in RAM.
  • Faster NPU kernels, measured: tensor concat about 23 times faster, a cheaper softmax and earlier prefetch in flash attention (a 5 s Wan clip renders 8.6 % faster from that alone), threaded copies.
  • DiT models below the flagships: the engine runs on the 8 Gen 2 and the 8s chips, and every Snapdragon gets FLUX.2 Klein and NitroFusion 2 on the Adreno GPU, in pieces small enough that the screen keeps responding. On the 8s Gen 3 and 8s Gen 4 the NPU's matrix unit turned out to return wrong results on every model, so those chips run with it off.
  • Drop-in fine-tunes: a CivitAI DiT checkpoint imports as it is; int8, fp8, ConvRot and mxfp8 files become Q8_0 on the phone, and a self-made GGUF the engine cannot read is repaired on import.
  • LoRA, ControlNet and IP-Adapter chosen per render on SD 1.5 and SDXL, on an NPU whose models are compiled ahead of time (the Swap models, converted with our npuforge).
  • Diagnose a wrong picture: it renders again with each part of the NPU moved to the CPU in turn, and the one that comes out right names the faulty part on your chip.
  • A node graph for fingers: a touch canvas, plugin nodes in JavaScript (QuickJS), and a scheduler that orders a graph with several checkpoints so each one loads once.

What your phone needs

Android 12 or newer and a Snapdragon with a Hexagon NPU. Models are not in the APK: each one downloads when you pick it, and the app shows which ones your phone can run.

chip runs, among others
Snapdragon 888 and newer SD 1.5
8 Gen 1, 8+ Gen 1 FLUX.2 Klein 4B and NitroFusion 2 on the GPU; Anima on the CPU (slow)
8 Gen 2, 8s Gen 3, 8s Gen 4 Z-Image Turbo, NitroFusion 2, FLUX.2 Klein 4B and 9B, Anima Turbo, Iris (much slower on the 8s chips for now)
8 Gen 3 and newer SDXL, Qwen Image 2.1, MOSS-Image, FLUX.3 Klein, Wan 2.2 video; Krea 2, Fibo 2.0 Edit and MiniMax H3 with 12 GB of RAM
8 Elite and newer all of the above, plus Ideogram 4 (12 GB), FLUX.1 dev and Neodragon video

Without a Snapdragon (MediaTek Dimensity, Exynos, Google Tensor) Text to image and Image to image still run with an imported SDXL MNN or SD 1.5 CPU model, on the GPU or CPU. The full table has every model and caveat.

Credits

Built on Local Dream. The C++ inference server started as a fork of xororz/local-dream: its QNN pipelines, model conversions, per-chip build tiers and device gating, and the SD 1.5, SDXL and upscaler archives the app downloads. Anyone interested in how Stable Diffusion runs on a Hexagon NPU should start there. The DiT and video models run on stable-diffusion.cpp (leejet) with happyyzy's Hexagon work, on ggml. Every engine, model and helper, with its authors and licence: credits and NOTICE.

Building it

JDK 17 and Android SDK 35: gradlew assembleDebug. The NPU backend is a native binary that is not in this repository; without it the app builds and the canvas works, but nothing renders. The QNN runtime is not in the APK: the app downloads the one for the phone's NPU (QAIRT 2.51, about 28 MB) once, with the first model, and keeps it through updates. Prompt translation needs Foxlet built once: tools/build_foxlet.sh (Git Bash) clones and builds it and copies the AAR into app/libs/.

Licence

CC BY-NC 4.0, inherited from xororz/local-dream rather than chosen; no further restrictions are added. Share it, fork it, build on it. Don't sell it or ship it inside something you sell. Full text in LICENSE; third-party components in NOTICE. The Qualcomm AI Runtime is under Qualcomm's own terms and is not in this repository.

Something failed?

Tap the share icon on the red error message: it shows a full report with the error, the backend's log, your phone and model files, but never your prompt. Copy or share it as a .txt, then attach it to an issue, or ask in the Telegram group. More in the troubleshooting guide.

Support

buymeacoffee.com/abrahampaulj: tips are welcome and change nothing: the app stays free, offline and CC BY-NC, and nothing is gated behind them.

Keywords: on device AI, offline Stable Diffusion, ComfyUI for Android, mobile Stable Diffusion, local image generation, on device video generation, node editor, Qualcomm Hexagon NPU, Snapdragon, QNN, FLUX.2, Z-Image, Qwen Image, Krea 2, Iris, MOSS-Image, Ideogram 4, SDXL, Anima, Illustrious, ControlNet, IP-Adapter, LoRA, img2img, inpainting, outpainting, Segment Anything, npuforge, image to prompt, Danbooru tags, no cloud, private.

Apps
About Me
GitHub: Trinea
Facebook: Dev Tools
AI Daily Digest