PromptRoute: Routing Different Image Customization Tasks in a Unified Model via Simple Prompts

Shiwen Zhang
Independent Researcher

[Model]

Abstract

We question whether there exists any conflicts when learning multi-ref subject-driven image generation and content-preserving style transfer given content ref and style ref with a unified model. The conclusion is no and our PromptRoute's solution is simple: we attach "Compose the pictures" for multi-ref subject generation, and "Style Transfer the style of Figure 2 to Figure 1, and keep the content and characteristics of Figure 1." for content-preserving style transfer during training and inference. Such simple prompts distinguish the tasks and eliminate the confusion.

our model outperforms Kling Image 3.0 using their test cases.

Schedule

• Release PromptRoute Diffusers Lora on QIE 2511.
• Release Installation instruction and inference code.
• Release Project Page and tech report.