Quantization of MOSS-Image SFT aiming to reduce storage and RAM/VRAM.
This version uses a mix of INT8 and 4-bit weights and takes up approximately 3.72 GiB. It was created by testing different precision settings against the official INT8 model.
107 layers use INT4 — W4A4 ConvRot MSE.
61 layers use INT8 — the original INT8 ConvRot weights.
Workflow in image uses Lodestone's 8 Step Lora so my GPU won't scream for mercy
Tested on:
NVIDIA GTX 1660 Super 6 GB
32 GB System RAM
Test settings:
8/12 steps
CFG 1
Euler / Simple
696 × 1048 resolution
Approximately 5.7 s/it on a GTX 1660 Super


