MOSS-Image 5B (Apache 2.0) tested on a 12GB RTX 3060 with 8-step DMD turbo, 6x faster
mearbode · reddit · 2026-10-17
A Reddit user benchmarked open-source MOSS-Image (5B DiT, Qwen3.5-9B text encoder, Apache 2.0) on a 12GB RTX 3060 with 35 hard prompts, one seed, no cherry-picking. INT8 quant peaks at 9.7GB VRAM; lodestones' 8-step DMD turbo—unlike a normal LoRA, it must run unmerged in fp32 with its own sampler—cuts 1MP generation to 24s (6x faster). Verdict: excellent multilingual text rendering and Layout JSON support; turbo drifts on lighting and fine constraints; both fail negation and clock times. Prompts and ComfyUI workflows are public.
More from Multimodal
- PRFL ablation: full-trajectory reward feedback beats early-only or late-only training — gowthami_s · 2026-10-17
- PRFL: reward feedback learning entirely in latent space, no VAE decoding needed — gowthami_s · 2026-10-17
- AndrewDai launches Glimpse, a 'visual thinking model' claiming frontier-level visual reasoning vs Astra — AndrewDai · 2026-10-17
- speech.cpp v0.8.8 adds Fish Audio drama-3 and M5 Mac speedups, no Python/CUDA needed — solyarisoftware · 2026-10-17
- Open-source Skill 'Anatomy' turns one prompt into explainer videos of any mechanism — dotey · 2026-10-17
- Open-source Claude Code skill renders Khan Academy-style blackboard explainer videos — _haiderfarooq · 2026-10-17