AI
HiRAE: Hierarchical Representation Autoencoding with Residual Budgets
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 03:23:50 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 02:18:24 UTC.
Briefings
HiRAE is a hierarchical autoencoder that learns residual corrections across encoder layers to improve image reconstruction fidelity. On ImageNet-256, it reduced the FID score from 0.299 to 0.209 compared to RAEv2 while maintaining guided generation quality. For text-to-image generation, HiRAE-24 improved alignment metrics on GenEval, DPG-Bench, and GenAI-Bench both before and after supervised fine-tuning, with post-fine-tuning GenEval increasing from 84.86 to 87.70 under identical generator-training protocols.
Key points
- HiRAE groups encoder layers by depth and learns residual corrections to the deepest representation.
- Group-wise norm caps bound residual corrections relative to the deep anchor with tighter budgets for shallower groups.
Context and limitations
Coverage is limited to ImageNet-256 and text-to-image generation benchmarks; results are preprint status and may vary with different generator-training protocols.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 03:23:50 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibofjbdgk7bmantdbduavxk64nwmc3zwqmqtzairq7ondbmbhku2q