Research map / Image generation / Autoregressive and latent generation

AHPA: Adaptive Hierarchical Prior Alignment for Diffusion Transformers

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: High-resolution diffusion synthesis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AHPA, which aligns intermediate SiT/DiT features to multi-level frozen SD-VAE encoder features routed by a timestep-conditioned dynamic router, accelerating diffusion transformer training without external encoders.

Problem

Representation alignment for diffusion transformers uses a fixed, timestep-agnostic supervision target and granularity, mismatched to the changing SNR along the denoising trajectory, and external encoders add cost.

Approach

Mine multi-level intermediate features of the frozen VAE encoder as a hierarchical prior bank and use a timestep-conditioned dynamic router to softly weight intra-group layers and inter-group (mid vs. deep) features as alignment targets, with zero inference overhead.

Key results

• AHPA on SiT-XL/2 at 400K iterations reaches FID 8.1, matching vanilla SiT-XL/2 trained for 7M iterations (8.3) and beating REPA (8.5) at 400K.
• With CFG on ImageNet 256 at 800 epochs, AHPA reaches FID 1.40 versus REPA 1.42 and REG 1.36 (REG is better).
• On MS-COCO T2I with MMDiT (SDE, NFE=250), AHPA achieves FID 4.52 versus REPA 4.74 and SRA 2 4.91.

Surprising findings

• A timestep-only router outperforms content-adaptive routing, which the paper reports as failing.
• Alignment to shallow VAE groups (G1, G2) hurts generation quality relative to mid/deep groups.

Limitations

• Gains on ImageNet 256 at 400K steps are slightly below REPA in recall and IS for some scales (e.g., SiT-L/2 IS 92.1 vs 95.0 for REPA).
• Crossover point of diagnostic gradient curves is architecture-dependent, so hand-designed schedules do not transfer.
• Results with CFG at full training are behind REG at 800 epochs (FID 1.40 vs 1.36).

Experimental setup

Datasets: ImageNet MS-COCO

Benchmarks: ImageNet 256x256 ImageNet 512x512 MS-COCO

Metrics: FID sFID IS Precision Recall PickScore G-SNR

Baselines: Vanilla SiT REPA SRA 2 SRA REG DiT-XL/2 iSSD MaskDiT SD-DiT DUPA MMDiT

Builds on: REPA SRA 2 SiT DiT MMDiT

Closest papers by meaning

• Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.92)
• REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 2025 · similarity 0.90)
• Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.89)
• Rethinking Cross-Layer Information Routing in Diffusion Transformer (NeurIPS 2026 · similarity 0.89)
• HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.88)
• Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.88)