Rethinking What to Cache in Few-Step Diffusion Transformers: Solver-Aware Target Selection
AutoTarget calibrates which DiT tensor to reuse instead of assuming one cache target fits every solver and model.
The paper argues that few-step distilled diffusion transformers make caching riskier because adjacent sampling steps are farther apart. Its method runs a small uncached calibration set, measures reuse error for candidate tensors, and picks the lowest-error target for the chosen model, solver, and reuse schedule. The authors report that the best target changes with model, resolution, and solver, while AutoTarget cuts DiT evaluations and cache storage with quality close to uncached generation. ArXiv · AI/CL/LG's note
The paper argues that few-step distilled diffusion transformers make caching riskier because adjacent sampling steps are farther apart. Its method runs a small uncached calibration set, measures reuse error for candidate tensors, and picks the lowest-error target for the chosen model, solver, and reuse schedule. The authors report that the best target changes with model, resolution, and solver, while AutoTarget cuts DiT evaluations and cache storage with quality close to uncached generation. ArXiv · AI/CL/LG's note
score 5