Comparison Results
* The text prompt is only used for text-based image editing. Proposed method (PIF) does not use the text prompts as input.
Reference Style
Loading prompt...
Text Prompt*
Reference Style
Loading prompt...
Text Prompt*
Reference Style
Loading prompt...
Text Prompt*
Results with Residual One-step Diffusion
After finetuning with the proposed residual one-step diffusion paradigm, the model demonstrates the ability to adjust the photographic style conditioned on the given text prompt, while maintaining the fidelity of the original image content.
Content
Comparison with Average
Loading prompt...
Average Style Prompt (Left)
Loading prompt...
Modified Style Prompt (Right)
Incremental Results
With the concepts optimized from reference image, PIF renders them onto content image incrementally.
Content
+Tint
Incremental Result
Reference
Results with Conflicting Concepts
Reference images may contain conflicting concepts beyond the shared style. PIF learns them like vector summation: since optimization only updates pseudo-word embeddings, and word embeddings are roughly linear, conflicting concepts cancel out, leaving embeddings close to an “average.”
Content
Reference A
Result
Reference A†
Reference B†
†: The image of Reference A shows a blue tint, light shadow, low contrast, and fuzzy highlight. Reference B shares the photographic style
in contrast, highlight, sharpness and shadow, but differs in the tint, by leaning towards orange.
#: Reference A+B means the text embeddings are optimaized on both Reference A and Reference B