AbrahamPJ/nightmare-moss-image

MOSS-Image SFT, Q4_0 GGUF (for Nightmare Mobile)

The diffusion transformer of OpenMOSS-Team/MOSS-Image-SFT (Apache 2.0), quantized to Q4_0 for the Hexagon NPU of Snapdragon phones, as used by Nightmare Mobile.

• moss-image-sft-q4_0.gguf: made from realrebelai/MOSS-Image_GGUFs' MOSS-Image-SFT-Q8_0.gguf with stable-diffusion.cpp's sd-cli -M convert --type q4_0. Tensor names and layout unchanged (layers.N.proj.q/k/v apart).
• Nightmare runs it with the 8-step DMD LoRA from lodestones/MOSS-Image-SFT-8-steps-dmd applied at run time (8 steps, CFG 1). The LoRA is NOT merged in: its change to each weight is about a thousandth of a Q4_0 rounding step, so a merge would erase it.
• Text encoder: unsloth/Qwen3.5-9B-GGUF Q4_0, with the tokenizer.json from OpenMOSS-Team/MOSS-Image-Base. VAE: the FLUX.2 VAE.

All upstream work is Apache 2.0; this conversion is published under the same licence.