All stories
MOSS-Image-Base 8-Step DMD Ecosystem Community reports

MOSS-Image-Base gets an eight-step community checkpoint and a simple LoRA route

Lodestones recommends loading the adapter onto the stock Base transformer, with CFG 1 and four to eight steps; the full sampler has stricter precision requirements.

By AITrending2 min read
Sources (3)

Why now The repository appeared October 18 at 01:27 UTC and received its update-12000 EMA checkpoint at 07:24 UTC, a fresh Base-model distillation distinct from the previously covered SFT turbo demo.

A new community MOSS-Image-Base distillation offers an adapter-only route to four-to-eight-step sampling. Lodestones’ English card describes a work-in-progress conversion from the Base model’s 50-step, CFG-3 configuration to eight network evaluations without classifier-free guidance.

The repository was created October 18 at 01:27 UTC. Its update-12000 EMA checkpoint followed at 07:24 UTC. This is a Base-model release, distinct from the MOSS-Image-SFT lineage behind the previously covered four-step turbo demo; those older results should not be transferred to this checkpoint.

Start with the LoRA, says the author

The recommended workflow is unusually concise: load lora.safetensors on the stock MOSS-Image-Base transformer, for example through a ComfyUI LoRA loader at strength 1. Use the usual flow-matching Euler sampler, shift 5, CFG 1 and four to eight steps. The author says this route needs none of the repository’s other adapter or sampler files.

The alternative training-time data-space sampler is more demanding. Its supplied transformer must be loaded and run in FP32 without autocast. According to the card, the distillation changes are smaller than BF16 rounding increments, so casting that transformer to BF16 discards most of the distillation. The documented faster BF16 route keeps the upstream Base transformer and applies unmerged FP32 adapters; the replacement time embedder and two modulation biases must remain together in FP32.

The release carries Apache-2.0 terms and targets users already running the upstream image model. It does not turn MOSS into a standalone desktop application or remove the upstream runtime and hardware requirements.

Eight network evaluations are a sampling target, not a measured wall-clock speedup. The card provides neither a latency comparison nor evidence of quality parity with the 50-step Base configuration. AITrending reviewed the English instructions and publication history without running inference. The practical news is the new downloadable checkpoint and its simpler adapter path, with precision handling still a material setup detail.

Related reading