Diffusion models typically spend the same computational effort on every token, even though some tokens need more capacity than others. Prior attempts to mix models allocated work by denoising step or by assigning drafting and verification roles. HybridStitch, a new training-free framework from researchers at arXiv, instead uses token-level feedback to decide which model should handle each adopted update during generation.

The method maintains a shared generation state for both continuous image refinement and discrete language commitment, while keeping model-specific key-value buffers to supply attention context for selected queries. This lets different models contribute at different token positions without retraining or fine-tuning.

On four main benchmarks, HybridStitch reports 1.63–1.68× faster image generation and 1.15–1.58× higher throughput for diffusion large language models. The authors estimate that dLLM decode query cost falls by 50–66% while generation quality remains comparable to using a single model throughout.