K2

K² · Artificial intelligence

Throw away most of the reference image, and the AI still draws it right — 4x faster

The authors built Sparse Context, a method that constructs sparse reference representations for reference-based diffusion models by retaining only a reduced subset of reference tokens.

RP
AR

Rishubh Parihar, Ayush Raina, R. Venkatesh Babu et al.

4 authors · cs.CV

arXiv preprintArtificial intelligenceJun 2026 · ~70s read

The 30-second scan

Some image generators don't just take a text prompt.

  1. They observe that dropping a significant portion of reference tokens at inference time largely preserves generation capabilities even without modifying the model.
  2. The method fine-tunes the model with random token dropping at varying ratios to encourage robustness to partial reference representations.
  3. The training strategy decouples the model from any specific token selection rule, allowing flexible control at inference time.
4x increase in inference speed for multi-reference generation2x increase in inference speed for single reference generation
Read the originalPDF