KAIST's LOWN Cuts Diffusion VLM Decoding FLOPs to 35% While Keeping 99.6% Accuracy
kaist-ai · hf · 2026-10-30
LOWN (Look Only When Needed) from KAIST AI targets inefficiency in diffusion vision-language models (DVLMs), which reprocess the entire input at every denoising step even though visual tokens make up 89–97% of it.
- The authors measure LLaDA-V step by step with and without the image and find removing it changes an unmasked token at only 22.8% of steps on average, as few as 7.9% — most steps don't need the image.
- Method: LOWN predicts masked tokens without the image each step and only reprocesses the image when this prediction disagrees with the last image-conditioned one; agreement certifies the earlier prediction for reuse, and the certified token count sets how many tokens get unmasked.
- Results: with no training or tuned threshold, LOWN retains 99.6% of LLaDA-V's score across ten benchmarks at just 35.3% of decoding FLOPs, generalizes across LLaDA-based DVLMs, and composes orthogonally with visual token pruning.
More from Infra
- Ollama installed ≠ model downloaded: pitfalls when running it on cloud GPUs — NoConflict2854 · 2026-10-30
- NVIDIA launches CUDA Rust: two tracks for writing GPU kernels in Rust — AxSaucedo · 2026-10-30
- Rapid-H3 renders MiniMax-H3 video+audio 3-5x faster on one DGX Spark — vchauhan_ · 2026-10-30
- ByteDance Seed: Effective Learning Rate Trajectories, Not Normalization, Drive WN Training Gains — ByteDance-Seed · 2026-10-30
- Dell workstation with 96GB NVIDIA GPU may cost less than the card alone — chemist_slime · 2026-10-30
- Midwit meme on Databricks vs Snowflake: just put it in S3 — cto_junior · 2026-10-30