Alibaba's Representation Guidance: 459M-Parameter Pixel Diffusion Transformer Beats Larger Models
alibabagroup · hf · 2026-10-12
Visual representations are largely built for understanding, but what makes them useful for generation is unclear. Studying pixel diffusion transformers, the authors find two key insights:
- Better representation quality does not necessarily translate into greater generative utility, as externally aligned semantics can be progressively lost during pixel synthesis.
- Architectural separation with specialized pixel decoders mainly redistributes semantic and spatial modeling without yielding representations better suited to generation.
They propose Representation Guidance (RG), which forms and exploits internal representations while retaining a plain Transformer: a low-dimensional bottleneck forces generation-relevant information into a compact representation, then explicitly guides decoder generation by modifying the flow-matching target according to its induced prediction offset.
Results: on ImageNet 256×256, RG-L/16 with only 459M parameters outperforms substantially larger pixel diffusion transformers, and RG-XL/16 reaches FID 1.55; FID 1.74 at 512×512; extended to text-to-image, it achieves 0.89 on GenEval and 87.1 on DPG-Bench.
More from Multimodal
- Anthropic's bet on code over imagegen is being vindicated, observer argues — IndraVahan · 2026-10-12
- AI-made 5-min Indian history doc uses Claude, Gemini and Flux — deedydas · 2026-10-12
- MixTok hybrid tokenizer hits gFID 1.95 on ImageNet with just ~131 tokens per image — ZhejiangUniversity · 2026-10-12
- AI-generated mud hut photo goes viral: the tiny clay oven is the most convincing object in it — misovalko · 2026-10-12
- Lux In Music adds Yue2 model and multi-track MP3/FLAC separation, fully local — alexcovo_eth · 2026-10-12
- One prompt, one shot: Capsule agent produces a 15s kinetic-type launch video — DavidBennett__ · 2026-10-12