← sources

Built a native macOS app for local Krea 2 / FLUX / Qwen image inference on Apple Silicon

source: reddit (reddit:StableDiffusion)published 2026-09-14 16:01ingested 2026-09-15 02:38

retained text



i've been building Radiant Canvas, a native macOS image generation app for Apple Silicon

i started it mostly because i wanted something i could just open and use on my Mac without keeping a Python env, backend process and a pile of dependencies around. i also wanted to understand how far i could push newer image models through MLX directly instead of wrapping an existing stack

the backend is C++20 / Objective-C++ using MLX + Metal directly, with Swift on the app side. there's no Python, diffusers or ComfyUI process running underneath

right now it supports Krea 2 Turbo, FLUX.2 Klein 4B, Z-Image Turbo, Qwen Image / Edit and ERNIE-Image Turbo

i'm integrating the model families separately instead of trying to squeeze everything through one generic runtime. that takes more work, but it also means i can deal with the weird parts of each architecture properly

for example the FLUX.2 path has its own MLX CPU/GPU streams, keeps the engine around between generations, explicitly releases heavier model parts afterwards, clears caches when needed, tracks peak memory use, handles OOM / GPU failure cases, supports live previews and can cancel from inside the actual inference loop

it also takes unified memory into account when deciding what resolution is safe to render internally, which matters quite a bit once you start pushing these models on 16/24 GB Macs

LoRAs are loaded directly by the native engine too, currently up to 10 at once for FLUX.2 with separate strengths

the other half of the app is Studio. it has 70+ nodes now, but it's not just a visual preset editor. the graph is actually used to schedule execution, so generated images, processing nodes and downstream generations resolve in dependency order

switches can prune branches that aren't used, and stuff like Cartesian Matrix, Zip and parameter/model/prompt sweeps create actual execution branches rather than just changing some UI fields

Draw Things is obviously the closest comparison on macOS and it's much more mature / broader in model support. i'm going a bit narrower with Radiant Canvas and spending more time on the native implementation for each model plus the workflow side

and compared to ComfyUI it's much more constrained, but the upside is that it's just one normal Mac app, no Python env, no separate backend, no node packages to keep alive

everything runs locally once the models are downloaded

**benchmarks**

Krea2Turbo 8-bit on MacBook M5 Max 48GB:

cold run:

**ComfyUI**: 45.95s  
**Radiant Canvas**: 25.9s

warm run (second generation):

**ComfyUI:** 32.02s  
**Radiant Canvas**: 24.5s



the application is free, with an optional PRO option, but it's only for a few advanced features, all models can be used with no limits without it:

[https://apps.apple.com/us/app/radiant-canvas-ai-image-gen/id6802973075](https://apps.apple.com/us/app/radiant-canvas-ai-image-gen/id6802973075)

document fetched 2026-09-15 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1wg7hca/built_a_native_macos_app_for_local_krea_2_flux