# Sparse Context Cuts Reference-Conditioned Diffusion Costs by Up to 6.7× via Token Dropping

_Research · published 2026-06-29_

Researchers have introduced Sparse Context, a method for dramatically reducing the computational cost of reference-conditioned image generation in diffusion models. The core insight is that the vast majority of reference tokens are redundant: experiments show that dropping 90% of tokens at inference still preserves coarse scene layout, with only fine details degraded. By fine-tuning with stochastic token dropping — randomly removing 75–95% of reference tokens at varying ratios during training — the model becomes agnostic to which tokens are retained, enabling flexible, task-specific selection strategies at inference without any additional retraining.

The method builds on FLUX-2-Klein-9B, fine-tuned with LoRA on a 63K-image dataset covering single- and multi-reference personalization and instruction-driven editing. At inference, task-aware selection strategies plug in freely: Canny edge maps guide token sampling for structural editing tasks, while saliency maps concentrate the budget on identity-critical regions for subject personalization. With six reference images — where reference tokens account for 86% of total tokens — Sparse Context delivers over 4× wall-clock speedup, and is described as complementary to other efficiency techniques.

## Sources
- [sparsecontext.github.io](https://sparsecontext.github.io/)
---
Canonical: https://genbuzz.news/posts/sparse-context-cuts-reference-conditioned-diffusion-costs-by-up-to-6-7x-via-token-dropping