DiffSynth-Studio/MOSS-Image-SFT-TurboLoRA

MOSS-Image-SFT Distillation-Accelerated LoRA

This model is trained using OPD (On-Policy Distillation) distillation acceleration based on DiffSynth-Studio, providing inference acceleration for MOSS-Image-SFT.

• Training framework: DiffSynth-Studio
• Base model: MOSS-Image-SFT
• Dataset: ImagePulseV2
• Training resources: 8 ZW810 PPUs
• Training steps: 6000 steps

Showcase

model-8steps.safetensors: An 8-step accelerated LoRA based on MOSS-Image-SFT. It distills the 50-step sampling trajectory into 8 steps, greatly reducing the inference cost while preserving the capabilities of the base model.

	This model (8 steps)	MOSS-Image-SFT (8 steps)	MOSS-Image-SFT (50 steps)
Example 1: Complex composition
Example 2: Text rendering
Example 3: Chinese text rendering

Prompts

• Anime style. A sweet girl with twin tails and a flower crown sits on a park bench, smiling at the viewer while holding a black camera. She wears a red T-shirt and a white tennis skirt. Beside her, an orange kitten rests calmly. The background is a soft blur of sunset colors, keeping the focus entirely on the character and the cat.
• A cinematic movie poster featuring a breathtaking, wide-angle landscape. The background showcases a vast, crystal-clear azure sky stretching upward, dotted with soft, volumetric white cumulus clouds drifting gently. Below the sky rises a majestic range of towering snow-capped mountains, their jagged peaks gleaming bright white against the deep blue atmosphere, rendered with crisp, photorealistic detail and dramatic lighting. In the center foreground, a magnificent eagle soars dynamically through the air, its wings fully extended in mid-flight, feathers catching the sunlight as it glides effortlessly above the alpine scenery. The composition evokes a sense of epic scale, freedom, and adventure, styled like a high-budget Hollywood blockbuster poster. At the bottom of the image, elegantly integrated into the lower third over the mountain base or sky gradient, is the text "DiffSynth-Studio". The typography is bold, modern, and cinematic, using a clean sans-serif font in crisp white or metallic silver, ensuring high legibility and a professional film-title aesthetic.
• In a bright, modern classroom, a young woman dressed in a traditional Japanese-style school uniform (JK uniform) is focused on explaining a technical concept. She holds a marker firmly, her posture poised as she writes '基于DiffSynth-Studio的\nOPD加速训练' across the pristine whiteboard. The lighting highlights the neatness of her script and the professional yet youthful atmosphere of the academic setting.

Inference Code

Install DiffSynth-Studio:

git clone https://github.com/modelscope/DiffSynth-Studio.git
cd DiffSynth-Studio
pip install -e .

Run 8-step accelerated LoRA for MOSS-Image-SFT:

import torch
from diffsynth.pipelines.moss_image import MOSSImagePipeline, ModelConfig

See the full description on the original page