Get Started
Home
Topics
Search
Library
Research questionHow can high-resolution diffusion Transformers prune tokens without sacrificing image fidelity or predictable compute?At high resolution, self-attention cost grows quadratically with the number of image tokens. Pruning tokens can reduce this burden, but removing information may harm generated-image fidelity and make computation harder to predict.
Diffusion Models
Image Generation
Inference Optimization
Machine Learning
Latest papersRecent research connected to this question, newest first.Shiva-DiT: Residual-Based Differentiable Top-$k$ Selection for Efficient Diffusion TransformersThe source studies Shiva-DiT, which uses hard top-k token selection in the forward pass and differentiable training of token scores and retention budgets. Experiments cover SD3-Medium, Flux.1-dev, and PixArt-Σ, reporting FLOP and latency reductions, including a 1.54× wall-clock speedup on SD3-Medium with competitive fidelity.research paper · Sep 2, 2026
Related questions
How can low-rank compression preserve text-to-image quality in large diffusion transformers?How can attention heads be pruned in text-to-image diffusion transformers without losing prompt-specific object identity?How can few-step diffusion defer high-resolution denoising without causing transition artifacts?How should diffusion samplers allocate limited neural evaluations without sacrificing generated image quality?