The model constructs self-distilled triplets with TeleStyle V1, then uses Distribution Matching Distillation to reduce inference cost while preserving Qwen-Image-Edit's general editing ability. It also uses Qwen2.5-VL-7B to generate content and style prompts to reduce content/style reference order confusion.
TeleStyle V2 is useful for image-editing applications where content and style references may be realistic, stylized, or mixed. The project links to model weights, code, and a Hugging Face demo for experimentation.
