Breaking

HeyGen Says Autonomous Agent Triples VAE Decoder Speed

July 1, 2026 at 16:54 EDT

Media Generation
AI Agents
Open Source

HeyGen Research says an autonomous agent built on Claude Code optimized the VAE decoder of the Wan 2.2 video model, delivering roughly a 3x inference speedup on a single NVIDIA H100 GPU, according to its technical report. The company says the agent outperformed expert-led tuning without human intervention.

HeyGen Research · Auto-Optim

An autonomous LLM agent sped up a video model 3.02× — with zero human help

Built on Claude Code, Auto-Optim optimized the Wan 2.2 VAE decoder on a single NVIDIA H100, cutting inference from 14.47s to 4.79s — surpassing the human-aided result in a fraction of the trials.

3.02×

Inference speedup on one H100

Blind experiments by the agent

61.3 dB

PSNR — cleaner than the reference

Same 3× result — far fewer trials

Trials needed to reach the ~3× speedup (each block ≈ 15 trials)

105

Manual (human-aided)

Auto-Optim agent

How the speedup was built up

Inference time per stage — shorter is faster (baseline = 14.47s)

fp32 eager (baseline)

14.47s · 1.00×

bf16

8.42s · 1.72×

+ CUDA graph + native-pad conv

7.42s · 1.95×

+ break-free decoder (fusion)

4.82s · 3.00×

Final refinement

4.79s · 3.02×

Where it stopped — the hardware limit

The agent halted after eight consecutive rounds without improvement, with profiling showing no remaining bottleneck.

99.8%

GPU utilization, dominated by convolution kernels

Frozen gate

Quality must match or beat the fp32 reference on PSNR

3 files

Swap to retarget the toolkit to other models

THE PROMISE

Repetitive, model-specific tuning of dtype, layout, CUDA graph and torch.compile can be handed to an agent — and it can reach the hardware limit unaided.

TO WATCH

Independent benchmarks are still thin. Optimizing for a single metric can invite gaming — reproducibility and generalization to other models and hardware remain open questions.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.