enhanced by silicone
Local text-to-image generation and AI image editing for Apple silicon Macs. Z-Image Turbo, Qwen-Image-2.1 and Anima Turbo, native in Swift and Metal: no Python, no server, no cloud.
Free and open source (Apache 2.0) · macOS 15 or later · M1 and later · signed and notarized
Every stage is compared channel by channel with diffusers and transformers in fp32. What cannot be checked against the reference does not ship.
Z-Image Turbo at 1024² in 113–119 s on an M1 Pro, in fp32, with the 8 evaluations of its recipe.
Weights stream from disk as needed. Siliconed never takes more than ¾ of the Mac's memory minus 2 GB (10 GB on a 16 GB Mac): each render computes what it needs before it starts and checks its own peak against that figure.
Nothing leaves your Mac, and it works offline once a model is installed.
Install full size, in 8 bits or in 4 to 6 bits; import LoRAs from Civitai or Hugging Face. Licenses and disk space are shown before the first byte.
silicontrol drives the running app from a terminal with JSON output; silicontrol help is all an agent needs.
New in 0.10: Anima Turbo, a 2-billion-parameter model for illustration, painting and anime. It knows artists by name and takes LoRAs; here the inCase LoRA at 0.8, a 1024² image in 49 s on an M1 Pro. It makes no photorealism, by its publisher's choice — and the images it makes may be used commercially.
Give Qwen-Image-2.1 one to three images and an instruction — “convert this photo to a western comic book style with cel shading”, “put the dog of image 2 next to her”. Its encoder sees the images. Drag the curtain across the result and its original to see exactly what changed.
The second image is a source: “dress the woman of image 1 in the clothes of the woman in image 2”. The surfer from Z-Image Turbo keeps her face, her beach and her light, and wears the blouse and skirt of the reader Anima painted.
Any two images, side by side in one frame: slide the curtain to see a photo by Z-Image Turbo become, in Qwen-Image-2.1's hands, a cel-shaded comic book — same face, same pose, same beach.
A grid across seeds, steps, LoRA strengths, models or prompt variants, drawn in the canvas with its axes as labels. The cell you pick becomes the finished image, with the same seed.
Standard, Compact or Light: the publisher's weights, 8-bit versions, or a 4-bit transformer. Every file is pinned to a revision and checked by sha256, and downloads resume where they stopped. The next file downloads while the previous one is prepared: Qwen-Image-2.1 installs in about 6½ minutes on a fast connection.
| Model | What it does | License |
|---|---|---|
| Z-Image Turbo | text-to-image, image-to-image, LoRA | Apache 2.0 |
| Qwen-Image-2.1 Turbo | text-to-image, editing by instruction with 1 to 3 images, LoRA | Qwen Research (non-commercial) |
| Anima Turbo | text-to-image (illustration, painting, anime), image-to-image, LoRA | CircleStone Labs Non-Commercial (the images may be used commercially) |
Timings on an M1 Pro with 16 GB, Standard weights: Z-Image Turbo 1024² 113–119 s (8 steps); Qwen-Image-2.1 Turbo 1024² ~134 s (8 steps); an edit with one reference ~210 s; Anima Turbo 1024² 43.7 s (8 steps).
Nothing is preinstalled: pick a model in the app, read its license, and it downloads.