Alibaba's Representation Guidance: plain pixel-diffusion Transformers beat bigger models
bdsqlsz · x · 2026-10-14
Alibaba, with Sun Yat-sen University, HKU and SJTU, introduced Representation Guidance (RG), an end-to-end pixel-space diffusion framework that scales plain Transformers by forming and exploiting compact internal representations to guide generation.
- Key insight: better representation quality doesn't mean better generative utility; specialized pixel decoders merely redistribute semantic/spatial modeling
- Method: a low-dimensional bottleneck squeezes generation-relevant info into a compact representation, which explicitly modifies the flow-matching target to guide the decoder
- Results: 1.55 FID on ImageNet 256×256, 1.74 at 512×512, 0.89 GenEval, 87.1 DPG-Bench
- RG-L/16 with just 459M parameters outperforms substantially larger pixel diffusion transformers; RG-XL/16 scales to 1024×1024 with far less compute
More from Multimodal
- Stunning Midjourney Generations: Cosmic Light Tunnels in Warm Tones — ciguleva · 2026-10-14
- Redditor combines H3, UVR and YuE2-3B into a full AI music video pipeline — Delicious-Farmer-234 · 2026-10-14
- AI-generated NERDVANA tee mashes Nirvana's smiley with a dead Finder icon — Delahuntagram · 2026-10-14
- The Prompting Zone Debuts Episode 01: The Last Human — Inevitable_Emu2722 · 2026-10-14
- Narrative AI video on a single RTX 3090: 720p Wan2GP workflow with 2x2 storyboard references — Inevitable_Emu2722 · 2026-10-14
- NVIDIA's PhyProbe scores physical consistency of AI-generated videos, beating prior evaluators — nvidia · 2026-10-14