Faster on large Macs
- The denoiser runs as one stream of GPU work, for the five models: a block's work is queued on the GPU without a wait between operations, the Neural Engine's rows beside it, and where the memory budget holds the model's weights they are read in place instead of streamed. The capacity map measures that pairing as a render runs it. Render times on the reference machine: Numbers.
- The Compacts compute in int8 × int8 (W8A8) on the GPU's tensor units where this Mac's measure finds it faster, the way their publishers compute them: the activation rotated and quantized to int8 per row by the operation that makes it, the weights read in place. Checked against ComfyUI's own recipe on the published integers (
siliconed check <model> dit --variant compact --int8-activations). - The decoders run in bf16 where the map finds it faster, and the end of each phase waits only for what Metal still holds.
The app
- A ✕ on the history's thumbnails: hover one, click the ✕, and that image leaves the history; ⌘Z brings it back.
- The Prompt Lab in a window of its own: a button in its header detaches the unfolded plate into a window as large as you make it, its text growing with it; closing the window folds the lab back to its line under the image (
silicontrol lab detach,lab attach). - Choose Steps, a Beta option (Settings › Advanced › Beta), off by default: on, the denoiser's box offers the step count, for a LoRA distilled to fewer steps (Krea 2 at 4); off, every render takes the model's own count again.
- The Prompt Lab opens folded while no model is installed, so that the models to download are what is seen; it unfolds when the first one is ready.
- After the update, a model whose weights changed in 0.21 (the Compacts of Z-Image and Anima, every Light) no longer renders until it takes them: the welcome says which, with the version you had chosen, and offers Update…; so does the Models window when another model is ready.
The Lights, in our own 4 bits
- The Lights of Z-Image, Qwen-Image-2.1, Anima and Krea 2 are now our own 4-bit (SQ), in place of the third-party GGUF Q4_K_M: more faithful to the publisher's model than the Q4_K_M it replaces, at about the Compact's speed and with less memory. One file per family, published on Hugging Face (
gwenn-ha-dev/…-Siliconed-SQ4). An installed Light takes the new weights at its next installation, which removes the old ones (the image model weighs 5.0 GB for Z-Image, 4.4 GB for Qwen-Image-2.1, 1.7 GB for Anima, 7.6 GB for Krea 2).
The Compacts of Z-Image and Anima
- Z-Image's and Anima's Compacts are now int8 convrot, like Qwen-Image-2.1's: Z-Image's image model from Comfy-Org, Anima's from Bedovyy, in place of unsloth's and vanes430's GGUF Q8_0. Their text encoders do not change, and both are checked against diffusers (
siliconed check). An installed Compact takes its new image model at its next installation (6.2 GB for Z-Image, 2.6 GB for Anima), which removes the old one.