Running Anima Turbo v1.1 natively on Mac (Swift/Metal, no Python)
#243
by gwenn-ha-dev - opened
Hi! I've added Anima Turbo v1.1 to Siliconed, a free, open-source macOS app that runs image models natively on Apple silicon, in Swift and Metal: no Python, no ComfyUI, nothing to wire. Pick Anima in the app, it downloads the weights, then you type a prompt. Text-to-image, image-to-image and LoRAs (kohya and ComfyUI layouts) work, and artists are written @name, as the model card says.
- Fast on a laptop: on an M1 Pro with 16 GB, 8 steps at CFG 1: 12 s at 512², 44 s at 1024², 76 s at 1024×1536 ( 49 s at 1024² with a LoRA).
- Checked against the reference: each stage (Qwen3-0.6B and the adapter, the DiT, the Qwen-Image VAE) is compared channel by channel with the reference implementation; at 1024² the DiT output is 1.4e-5 from an fp64 computation, the fp32 reference itself at 1.5e-5. A LoRA is checked against ComfyUI's own merge.
- Small Macs welcome: 5.65 GB in all with the full weights, or 4.10 / 3.31 GB with the GGUF Q8_0 / Q4_K_M DiT from vanes430, kept at their published precision.
- Civitai-ready: the PNGs carry the settings Civitai reads, and link to Anima Turbo v1.1 there.
The app is Apache-2.0; the model keeps its CircleStone Labs Non-Commercial license : the app shows it and asks you to accept it before the first render. Notarized DMG: https://github.com/gwenn-ha-dev/Siliconed
I'm the author; happy to answer questions. Thanks to CircleStone Labs and Comfy Org for Anima, and to vanes430 for the GGUF!