MOSS DMD 8-Step — QKV Forge Neo Fix
Experimental LoRA Compatibility Conversion | MOSS-Image-SFT
An experimental QKV-compatible conversion of the original MOSS-Image-SFT 8-Step DMD LoRA, developed by lodestones.
This version adapts the original separate Q, K and V LoRA projections into fused QKV projections for improved compatibility with MOSS-Image INT8 checkpoints in Forge Neo.
What's Different?
Original Q/K/V projections converted into fused QKV LoRA layers.
All 42 transformer blocks processed.
168 LoRA modules recognized by Forge Neo.
No unmatched LoRA keys reported during loading.
Successfully tested with 8-step image generation.
No additional training performed.
Tested Configuration
Base model: MOSS-Image-SFT INT8
Text encoder: Qwen3.5-9B W4A8 GPTQ
VAE: FLUX.2
Interface: Forge Neo
Steps: 8
CFG: 1.0 1.5
LoRA strength: 1.0
GPU: NVIDIA RTX 3060 Laptop 6 GB
Important Notes
This is an experimental compatibility conversion, not a new DMD distillation or independently trained acceleration LoRA.
The original DMD project uses a specialized sampling algorithm. This conversion improves LoRA layer compatibility but does not implement the complete official DMD sampler.
Image quality and compatibility may vary across checkpoints, samplers and Forge versions.
Credits
Original DMD LoRA and distillation: lodestones
https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd
Original MOSS-Image model: OpenMOSS Team
https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT
QKV compatibility conversion and Forge Neo testing: Vince
License
Apache License 2.0, following the original project license. This modified distribution must retain applicable license and attribution notices.
Independent community modification. Not an official lodestones or OpenMOSS release.