5GB or 9GB? A full breakdown of every file in the ComfyUI repack of the MOSS-Image 5B DiT
katefix9 · reddit · 2026-10-22
A detailed Reddit guide untangles the MOSS-Image 5B DiT ComfyUI repack: the 9GB file is bf16 and the 5GB one is an int8 ConvRot quant of the same checkpoint (Base and SFT each ship both). Downloads also need a Qwen3.5-9B text encoder — bf16 (19.31 GB), int8 (9.75 GB) or w4a8 GPTQ (7.34 GB, the ComfyUI template default) — plus a flux2-vae VAE. Base suits LoRA training (no post-training/CFG distillation), while the template's sampler defaults (25 steps, CFG 2.5, ersde) are lighter than the README's 50-step/CFG 3 settings. Community quants cover GGUF, MLX for Macs, and 4/8-step LoRAs. Open question: w4a8 vs bf16 encoder quality on long prompts.
More from Multimodal
- Higgsfield Launches Content Factory to Turn Any Idea into a Viral Reel — SimplyAnnisa · 2026-10-22
- Side-by-side: Seedance 2.5 output vs free Muse looks nearly identical — sven_ai · 2026-10-22
- HeyGen ships a wave of updates: Hyperframes Studio, HeyGen Video and Pro Voice — toolstelegraph · 2026-10-22
- Layout JSON vs plain prompt, 8 seeds each: when bounding boxes are worth it — SquareBat780 · 2026-10-22
- Combating MOSS Image blur with a mid-denoise latent sharpening trick — LMLocalizer · 2026-10-22
- Tencent ARC paper cuts video model training compute 57.9% while boosting physics plausibility 11.2% — _akhaliq · 2026-10-22