AI
SemanTok: Predictable Semantic Tokens for Efficient Autoregressive Video Generation
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 29/09/2026, 20:00:00 UTC
Briefing created: 03/10/2026, 08:08:51 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 15:18:24 UTC.
Briefings
SemanTok introduces a flexible video tokenizer that combines frozen DINO features with lightweight heads to reconstruct token prefixes from each retained prefix alone. This approach achieves high semantic alignment and video fidelity across all autoregressive model sizes, matching or beating VideoFlexTok models 3.4 times their size. The method maintains semantic alignment on out-of-distribution classes and improves decoder alignment at every noise level, including pure noise. Short token prefixes are cheaper to predict and offer better generation fidelity, with pixel detail deferred to later tokens.
Key points
- SemanTok uses frozen DINO features combined with lightweight heads to reconstruct token prefixes from each retained prefix alone, achieving high semantic alignment and video fidelity across all autoregressive model sizes.
- A 201M SemanTok AR model matches or beats a VideoFlexTok AR model 3.4 times its size, demonstrating superior performance at every model scale.
Context and limitations
The study compares SemanTok against existing flexible tokenizers using only available metrics; specific limitations of the REPA loss mechanism and the exact impact of frozen DINO features on out-of-distribution performance remain unquantified.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/10/2026, 08:08:51 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeih2fikzcvfkqtkfvavhcmcweabp2iy7bliiee7i53frea4c77yccq