OpenMOSS-Team

/

MOSS-Image-SFT

Like 792

Follow

OpenMOSS 1.71k

Hi!
Great model indeed.

MOSS-Image is now in 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 MOSS-Image in the app, choose an install, it downloads the weights, then you type a prompt. Text-to-image, and LoRAs (ComfyUI, ai-toolkit, kohya and diffusers layouts).

• In 8 steps, with the community's LoRA: there is no official fast version yet, so the app runs MOSS-Image-SFT under lodestones' DMD LoRA at strength 1, 8 evaluations, no CFG. The LoRA is kept in fp32 beside the weights, never merged. That's why the app labels MOSS-Image as a beta; it leaves beta when you release your own fast version (+1 to discussion #7).
• On a 16 GB laptop: on an M1 Pro, 1024² in ~108 s and 1024×1536 in ~174 s, with the publisher's bf16 weights (Standard, 27.5 GB) or Comfy-Org's int8 convrot (Compact, 17.9 GB).
• Little memory: the weights are streamed from disk, and the memory need is computed from the model's shapes before the first step. The app's peak: 2.4 GB at 1024², 2.8 GB at 1024×1536.
• Checked against the reference: each stage (the tokenizer and the schedule to the bit, Qwen3.5-9B's 32 layers, the DiT with and without the LoRA, the 8-step trajectory, the FLUX.2 decoder) is compared channel by channel with the reference pipeline in fp32. At 512², the DiT output is 1.3e-6 from the exact (fp64) result; the reference's own error is 1.4e-6.

The app and the model are both Apache-2.0. Notarized DMG: https://github.com/gwenn-ha-dev/Siliconed

[image: moss.jpg — MOSS-Image in the app, Compact, 1248×832, 8 steps, seed 6001, on an M1 Pro]

I'm the author; happy to answer questions. Thanks to OpenMOSS for the model, to lodestones for the 8-step LoRA, and to Comfy-Org for the int8 weights!