gwenn-ha-dev/Siliconed v0.16

Siliconed 0.16

on GitHub

A fifth model: MOSS-Image

• MOSS-Image, by the OpenMOSS team, joins
  Krea 2, Z-Image, Qwen-Image-2.1 and Anima. A single-stream DiT of 5.1 billion parameters that reads
  the prompt with Qwen3.5-9B, a hybrid language model (three layers in four of linear attention,
  Gated DeltaNet, new in the engine), and decodes with FLUX.2's VAE.
• MOSS-Image is a beta, and the app says so (a “Beta” badge on its card and in Models): its 8-step
  recipe is the community's LoRA, not its publisher's. It leaves beta when OpenMOSS releases its own
  fast version.
• Its recipe is the community's: its publisher released no fast version, so Siliconed runs
  MOSS-Image-SFT under lodestones' DMD LoRA
  at strength 1: 8 evaluations, no CFG. The LoRA is kept beside the weights in fp32, never merged:
  merged into the bf16 weights, its deltas would be lost to rounding. Text-to-image; no Detail, no
  edit.
• Your LoRAs stack on top of the 8-step one: ComfyUI's and ai-toolkit's layout (the attention's
  fused qkv, split exactly), kohya's and diffusers'. LoRAs trained on MOSS-Image Base apply as the
  community uses them. Each imported module's ΔW equals PyTorch's from the published file.
• Times: on an M1 Pro with 16 GB, seed 42, 8 evaluations, queued in the app: 1024² in 107.7 s
  in Compact, 107.3 s in Standard; 1024×1536 in 174.2 s in both.
• Faster Rendering takes one half-size evaluation, not three: from the second, MOSS-Image
  smooths skin and softens fine detail where the other models keep it. 8 % less time at 1024²
  (116.9 s with every evaluation at full size).
• The preview shows during the render, as with the other models: a projection of the FLUX.2
  latent's 128 channels to colour, fitted on twelve renders.
• Checked against the reference, stage by stage: the tokenizer and the schedule to the bit, the
  text encoder's 32 layers, the denoiser with and without the LoRA, the 8-step trajectory and the
  decoder, each at the fp64 witness where the reference carries its own error. At 512², the
  denoiser's output is 1.3·10⁻⁶ from the exact result, the reference's own 1.4·10⁻⁶. Under the
  Neural Engine, both versions' denoiser stays within bf16's own error plus 25 %, as for the other
  models.
• Standard or Compact: 27.5 GB (the publisher's bf16, the LoRA in fp32) or 17.9 GB (the DiT and
  the text encoder in int8 convrot, as Comfy-Org publishes them); the Compact is preselected, as for
  the other models (on 8 GB too, since there is no Light to fall back on), with the Neural Engine on by default. No Light yet: a 4-bit DiT is published,
  not yet checked.
• Within the budget: its memory need is computed from the model's shapes before the first
  computation. The app's peak: 2.4 GiB at 1024², 2.8 GiB at 1024×1536, in both versions.
• License: Apache 2.0, the LoRA included. Commercial use allowed.

Every Mac

• The first launch after an update no longer runs slower: when the first gesture was an
  exploration, a grid or silicontrol add, the session ignored this Mac's measure and ran on the
  defaults — the AMX off, about 20 % per step on an M1 Pro. silicontrol status now says which
  measure the session uses.
• A measure taken off rest is taken again: measured while the Mac was warm, in Low Power Mode, or
  with its swap growing (another app pressing on memory), the choices still serve, the Chain's chip
  stays orange, and the next launch at rest measures again before its first render — once.
• An update that leaves the engine untouched keeps this Mac's measure: the measure is tied to the
  engine's sources and compiler, no longer to the app's build; one that moves the engine measures again.
• macOS 26 or later: the only version Siliconed was ever measured and checked on. The app no
  longer installs on macOS 15.