@rohanpaul_ai: Spatially Speculative Decoding (SSD) sped up autoregressive image models up to 13.28X by predicting image rows in paral…
Summary
Spatially Speculative Decoding (SSD) accelerates autoregressive image models by predicting entire rows in parallel using small helper networks, achieving up to 13.28x speedup while maintaining benchmark performance.
View Cached Full Text
Cached at: 07/14/26, 08:19 AM
Spatially Speculative Decoding (SSD) sped up autoregressive image models up to 13.28X by predicting image rows in parallel.
By drafting whole rows from nearby spatial information, not single tokens.
Most of the time these models take a 2D image and convert it into a long string of tokens. Then they make each token pass through another expensive model.
SSD adds small helper networks that can look right and down . This allows nearby tokens as well as full rows to be drafted at the same time .
Such helpers don’t try to predict the exact visual token labels. Instead, they predict more detailed internal features that are easier to learn correctly.
The primary model reviews each drafted block at the same time, correcting any errors so that the entire block is not wasted.
The generation time of Janus-Pro, Lumina-mGPT, and Emu3 was 5.74 to 13.28 times faster, but the benchmark scores were comparable to standard decoding.
Lumina-mGPT and Emu3 saw some wall-clock gains from a leaner inference loop, whereas Janus-Pro used matching loops.
Larger token grids increase the gains but the original model is static and small trained heads speed up the process.
– arxiv. org/abs/2606.20543
Title: “SSD: Spatially Speculative Decoding Accelerates Autoregressive Image Generation”
Similar Articles
Speculative Decoding for Autoregressive Video Generation
SDVG adapts speculative decoding to autoregressive video diffusion, using an image-quality router to achieve up to 2.09× speed-up with 95.7% quality retention on MovieGenVideoBench.
SpecLA: Efficient Speculative Decoding for Linear-Attention Models
SpecLA proposes a speculative decoding runtime tailored for stateful linear-attention models, achieving up to 1.70x end-to-end speedup over autoregressive decoding on an NVIDIA H100 with a GDN-1.3B target.
PSD: Pushing the Pareto Frontier of Diffusion LLMs via Parallel Speculative Decoding
This paper introduces Parallel Speculative Decoding (PSD), a training-free framework that accelerates diffusion LLM inference by jointly improving spatial and temporal efficiency, achieving up to 5.5× tokens per forward pass with comparable quality to greedy decoding.
Unlocking Parallelism in Autoregressive Language Models via Speculative Decoding with Progressive Tree Drafting
Proposes Progressive Tree Drafting (PTD), a training-free, model-agnostic speculative decoding method that uses progressive tree structures and stepwise pruning to enable parallel generation of multiple draft paths, achieving up to 2x speedup on various benchmarks.
SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference
SparDA proposes a decoupled sparse attention architecture that adds a lightweight 'Forecast' projection to predict future KV cache needs, enabling lookahead prefetching from CPU to GPU and reducing selection overhead. On 8B sparse-pretrained models, it achieves up to 1.25× prefill and 1.7× decode speedup, with up to 5.3× higher decode throughput over non-offload baselines.