KAIST APT Research Targets Faster, More Efficient Diffusion Transformers

KAIST researchers propose APT for Diffusion Transformers

A KAIST research team has published APT: Accelerating Diffusion Transformers via Attention Probability-Guided Pruning and Quantization, submitted to arXiv on August 26, 2026 and listed by the first author as an ICCAD 2026 publication.

APT is a research accelerator architecture, not a commercially released KAIST chip or software product.

The work targets an increasingly important inference bottleneck: high-resolution Diffusion Transformers can spend large amounts of computation and memory bandwidth on attention operations even when many attention relationships contribute little to the final result.

Attention-guided pruning and precision

The paper introduces Attention Probability-guided Adaptive Dual Thresholding (APDT).

Instead of applying one uniform pruning or quantization rule, APDT uses attention-probability information to decide which interactions can be pruned and which should be processed at lower or higher precision.

The goal is to cut unnecessary attention work while preserving the information that matters most for image generation quality.

APT also includes Timestep-Aware FlashAttention (TAFA). Diffusion models perform many iterative denoising steps, and attention patterns can remain related across neighboring timesteps. TAFA predicts attention probabilities using this temporal behavior so that the accelerator can reduce redundant computation while retaining compatibility with a FlashAttention-style execution flow.

Reported results

The authors evaluate APT with high-resolution Diffusion Transformer workloads including PixArt-α, Stable Diffusion 3 and FLUX.

According to the paper, APT achieves up to 8.16× speedup and 14.98× energy-efficiency improvement relative to the authors' NVIDIA A100 baseline.

The paper also reports up to 3.01× speedup and 2.04× energy efficiency versus the EXION accelerator used for comparison.

These are author-reported research results under the paper's evaluation methodology, not independent production benchmarks or guarantees for general GPU deployments.

Why the research matters

Diffusion Transformers are becoming larger and are increasingly used at higher image resolutions, making attention computation and memory movement important deployment costs.

APT is notable because it treats pruning, precision and hardware execution as one co-designed problem. That approach may be more effective than optimizing only the neural network or only the accelerator.

The work also reflects a broader trend in AI infrastructure: model architecture, sparsity, quantization, memory layout and specialized hardware are being optimized together to reduce inference cost.

Release status

The paper is publicly available on arXiv and the research is associated with KAIST's hardware and AI-acceleration community.

There is no announcement of a generally available APT processor, cloud instance or production SDK. Practical adoption would require implementation, hardware validation, image-quality analysis across broader workloads and comparison against current production GPUs and software stacks.

For developers, the main takeaway is the technique: attention probabilities can potentially guide both which work to skip and what numerical precision to use, while timestep structure in diffusion inference provides another source of exploitable redundancy.