RSTR: Reducing SpatioTemporal Redundancy in Diffusion Transformers

Diffusion Transformers (DiTs) have achieved remarkable success in image generation, yet their deployment is hindered by high computational costs. We identify two sources of redundancy. First, $\textbf{temporal redundancy}$: Classifier-Free Guidance (CFG) applies costly dual forward passes at every timestep, yet guidance matters only at specific steps, and variable scales at critical steps can compensate for skipping others. Second, $\textbf{spatial redundancy}$: under variable guidance, different transformer blocks exhibit heterogeneous sensitivity, yet uniform calibration across all blocks wastes computation while failing to address their varying requirements. We present RSTR, the first framework to jointly reduce spatiotemporal redundancy in diffusion transformers. Stage-1 addresses temporal redundancy through evolutionary search, discovering sparse guidance schedules with variable scales. Stage-2 addresses spatial redundancy through adaptive rank allocation, assigning calibration capacities to transformer regions based on their sensitivity. Experiments on DiT-XL/2, PixArt-$\alpha$, FLUX, and state-of-the-art Qwen-Image demonstrate 50\%-70\% compute savings while maintaining or improving quality. On DiT-XL/2, RSTR achieves 57\% savings with 15\% FID improvement; on Qwen-Image, 3.43$\times$ speedup with preserved quality.

Modern AI image generators produce realistic pictures from a text prompt, but they are slow and energy-hungry because they repeat far more computation than they need. This redundancy takes two forms. One is across time: a technique called "guidance" keeps the image faithful to the prompt, yet the system reapplies it across the dozens of steps it takes to draw a picture, roughly doubling the work, even though only a few steps need it. The other is across the model: it does the same heavy work everywhere, even though some parts matter much less. Our method, RSTR, cuts both. Rather than hand-tuning, we let the computer search for where guidance truly matters, and it finds that strong guidance at a handful of key steps works as well as weak guidance everywhere, like a road trip that needs directions at just a few turns. RSTR then puts effort only into the parts that need it, and reuses earlier results for the rest. The result is the same or better image quality with 50 to 70% less computation, up to 3.4 times faster, with no retraining, so existing models can run faster, cheaper, and greener right away.