Published November 4, 2026by _HaDeS_

HaDeS MOSS Suite v0.1.1 for Comfyui is out

151 views6 reactions0 comments on CivitAI3 collected

mossannouncementnodesloramergecomfyui

HaDeS MOSS Suite v0.1.1 is out (Comfy Registry: hades-moss-suite, GitHub: https://github.com/Devildonia/HaDeS-MOSS-Suite)

12 ComfyUI nodes for MOSS-Image. Every default and tooltip claim comes from measurements, all published in tools/RESULTS.md.

Mergers: Block Merger, with one ratio per group of 6 layers, and Multi-Model Merger (TIES, DARE, SLERP, Model Stock…). Both work as lazy patches and accept int8 and fp16-fixed files.
LoRA tools: Extractor, Block Weight, Merger and Resize, for every community MOSS LoRA format.
Model Saver: saves bf16, fp32 or int8 ConvRot in the Comfy-Org layout.
Pipeline: Latent (128-channel, 16 px grid), Layout Compiler (OpenMOSS drafts, byte-identical to the official compiler), and a KSampler with the official defaults.
Telemetry: checks your pipeline against what was measured.

A few things we measured:

layers 24–35 carry most of the SFT → Base difference;
shift 5.0 stays clean from 768² to 2048²;
SageAttention is about 22% faster than PyTorch attention at 2048²;
to upscale a MOSS latent, use “Upscale Latent By”; “Upscale Latent” doubles the size.

Example workflows are included.

More complete info soon.

HaDeS