Modelpublic

LiconStudio/Qwen3.5-9B-t2i_prompt_distillation_GGUF

sourceHugging Faceapache-2.0updated 7mo agoView on Hugging Face

3likes175downloads

Model Card

Overview

Distilled from Qwen3.5-Plus, this model transforms vague user prompts into clear, structured English image generation prompts across multiple dimensions (lighting, composition, style, era, etc.) while staying faithful to the original intent—no over-embellishment.

Distillation Pipeline

Stage	Output
Stage 1	6D element library (135 elements)
Stage 2	1,000 unique combinations
Stage 3	10,000 blurred prompts (7K CN + 2K EN + 1K Mixed)
Stage 4	10,000 structured English prompts

Data Composition

Type	Count
Short Input	5,500
Long Input (detailed)	2,500
Complex Structure (spatial)	2,000

Limitations

—Optimized for image generation prompts only
—May not handle culturally specific references well