A ComfyUI style node graph for Stable Diffusion — and now video — on Android, running entirely on device.
Local text to image, image to image, inpainting and text to video on the Qualcomm Hexagon NPU. No server, no account, no cloud, no network. Built on LocalDream's NPU backend.
Early. It works — text to image, image to image, inpainting and text to video all render on device — but this is a young project from one developer, so expect rough edges.
Download the APK — Android 12 or newer, arm64, and a Snapdragon with a Hexagon NPU. Models are not in the APK; pick one in the app and it downloads on first use.
Stable Diffusion 1.5, SDXL and Anima run locally on a Snapdragon NPU, offline. You draw a graph, press Run, and the picture appears on the node that made it. Everything stays on the phone.
⚠ Video is much narrower than pictures, and needs ~8.6 GB of models. It runs only on a Snapdragon 8 Elite (gen 4) or 8 Elite Gen 5 (gen 5) — nothing older, because a compiled model runs on the architecture it was built for and every newer one. The app decides by running a small real model on your chip rather than by reading its name, so it tells you before you spend the 8.6 GB — and on a phone that cannot, video is hidden everywhere rather than offered and then failing.
⚠ None of that applies to pictures. Text to image, image to image, inpainting and upscaling run on far more phones and need none of those models.
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, SDXL, Anima, text to image, text to video, image to video, img2img, inpainting, Segment Anything, SAM 2.1, npuforge, safetensors to QNN, on device model conversion, no cloud, private.
- 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.
- Text to video, on the NPU. A prompt in, 49 frames at 1024x640 out — about two seconds of clip in about 25 seconds on an 8 Elite. ⚠ See the requirements above: ~8.6 GB of models and an 8 Elite or 8 Elite Gen 5. Image to video animates a photo instead, and is the faster of the two — same three nodes, with a photo wired in. The clip loops on the node that made it and plays full screen.
- Seven node types, not seventy. A flow is
prompt → generate → output, plus a photo where one is wanted. Cropping, masking, encoding, sampling, blending and decoding happen inside the sampler, because there is no runtime compiler on an NPU and nobody can recombine the inside of a pipeline anyway — so the whole process is one node. Inpainting used to be ten nodes and is four. Tap to select: the Inpaint flow comes with a Segment model node wired in and tap what you want redone — Segment Anything 2.1 on the phone's CPU, an 87 MB download under Models, Tools. - The ops are still decomposed underneath.
encode_text,vae_encode,sample,latent_blendandvae_decodeare separate endpoints, and conditionings and latents move between them as handles that never cross the wire — a 512 image costs about 40 bytes of JSON instead of 780 KB. A plugin reaches the ops; a person reaches the nodes. - Two checkpoints in one graph. Each sampler carries its own, and the executor orders the run to minimise backend relaunches — finish everything under the model already loaded, then switch. The run bar says what the relaunches will cost before you press Run.
- Recipes to start from. Text to image, image to image, inpainting where you paint the area to redo, upscale a photo, and both video flows — each laid out so no wire crosses and the whole graph fits the screen when it opens.
- Batching. Arm
seed,steps,cfg,denoiseorscheduleron 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 — or bring your own converted
.bin. - Embeddings. Import a textual-inversion
.safetensorsand name it in a prompt. - Pick your size. SD 1.5 renders any resolution its checkpoint ships a patch for — 512² up to 1024², portrait and landscape; SDXL and Anima crop their fixed 1024² canvas to the shape you choose.
- Bring your own model. Twenty-four checkpoints in the catalogue — five SD 1.5, ten SDXL and nine Anima — or import a converted one as a zip, including SDXL converted by npuforge, which turns an SD safetensors checkpoint into a QNN model on the phone itself, with no PC involved. ⚠ Anima is 8 Gen 3 or newer, ~4.3 GB per checkpoint, and slow: about 80 s a picture on an 8 Elite, and Android may stop it while other apps are busy. Each sampler has its own checkpoint picker, grouped by family, listing what is actually on the phone; switching family rewrites that node and keeps every wire. The app carries every HTP architecture tier and picks the build your chip can actually load. The video models are their own download under Models, resumable per file because 8.6 GB over a phone connection will be interrupted.
- English, 中文 and Русский. The interface follows your phone's language. Adding another
is a file drop — copy
app/src/main/res/values/strings.xmlinto avalues-<code>/folder and translate it; no code changes. ⚠ Model prompts are never translated: SD 1.5 reads English tags, so a translated prompt would render something else. - 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.
Contributors can add their own node without an app release: a JSON manifest plus a small
JavaScript file, zipped and pushed to the app's plugin directory. No toolchain, no compile
step. Four worked examples — from a one-op wrapper to a multi-node toolkit — are in
examples/.
This is early: the plugin format can still change between versions, there's no host op yet
that lets a node run a model of its own, and an imported script should be treated like any
other code you didn't write. Read an example in examples/ for the manifest shape and the
handful of image/latent ops a node can call.
JDK 17 and Android SDK 35.
gradlew.bat assembleDebug
The NPU backend is a native binary and the QNN runtime libraries are not in this repository. Without them the app builds and the canvas works, but nothing renders.
The NPU work is not ours. This app forks the C++ inference server from xororz/local-dream, which is where the QNN pipelines, the model conversions, the per chipset build tiers and the device gating all come from. The checkpoint and upscaler archives it downloads are published by the same author. Anyone interested in how Stable Diffusion runs on a Hexagon NPU at all should start there rather than here.
LocalDream is the consumer app built on that backend, and it stays the simpler way to generate a picture on a phone. Its mask editor, brush behaviour and batch strip are ported here rather than reinvented, on purpose: someone who has both installed should not have to learn the same tool twice.
The video is not ours either. It is Neodragon by Qualcomm AI Research, released under BSD-3-Clause-Clear with additional terms on the model card. The pyramidal schedule, the autoregressive MMDiT loop and the streaming VAE decode are theirs; the NPU conversions this app runs were made from their weights. Its first frame comes from SSD-1B by Segmind, Apache 2.0.
What is new here is the graph. Pipelines were decomposed into ops, latents and conditionings became handles, node types became something a contributor can add without an app release, and the video path was taken apart into the same kind of nodes as the picture one.
CC BY-NC 4.0, inherited rather than chosen. This project derives from xororz/local-dream, which is released under that licence, and it continues under the same terms. No further restrictions are added here.
Share it, fork it, build on it. Do not sell it or ship it inside something you sell.
Full text in LICENSE; attribution and third party components in NOTICE.
The Qualcomm AI Runtime libraries the app needs at runtime are covered by Qualcomm's own terms, which this licence does not override. They are not in this repository.
Tips are welcome and change nothing: the app stays free, offline and CC BY-NC, and nothing is gated behind them. A donation is not a purchase of the software and does not grant commercial rights the licence withholds.