Research map / Image generation / Autoregressive and latent generation
V-Co: A Closer Look at Visual Representation Alignment via Co-Denoising
Han Lin, Xichen Pan, Zun Wang, Yue Zhang, Chu Wang, Jaemin Cho, Mohit Bansal
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use fully dual-stream architecture instead of shared-backbone token concatenation (ablation) | V-Co co-denoising JiT-B/16 · ImageNet-256 (200 ep, CFG=1.0) | FID: 14.70 -> 8.86 | Improved |
| allocate more feature-specific blocks in token-concat single-stream (2 to 6) (ablation) | single-stream TokenConcat · ImageNet-256 (200 ep, CFG=1.0) | FID: 14.70 -> 12.35 | Improved |
| use token concatenation instead of direct addition for fusion (ablation) | single-stream JiT-B/16 · ImageNet-256 (200 ep, CFG=1.0) | FID: 15.15 -> 14.70 | Improved |
| add co-denoising with Latent Forcing-style single-stream over plain JiT-B/16 | JiT-B/16 · ImageNet-256 (200 ep, CFG=1.0) | FID: 32.54 -> 13.06 | Improved |
| use structural semantic-to-pixel masking instead of zero-embedding dropout for unconditional CFG (ablation) | V-Co dual-stream · ImageNet-256 (200 ep, CFG=2.9, independent dropout) | FID: 6.69 -> 3.59 | Improved |
| use semantic-to-pixel mask only instead of bidirectional cross-stream mask (ablation) | V-Co dual-stream · ImageNet-256 (200 ep, CFG=2.9, joint dropout) | FID: 5.66 -> 3.18 | Improved |
| use joint dropout of labels and semantic features instead of zero-embedding independent dropout (ablation) | V-Co with semantic-to-pixel mask · ImageNet-256 (200 ep, CFG=2.9) | FID: 3.59 -> 3.18 | Improved |
| jointly drop labels and semantic features with input-level zero embedding (ablation) | V-Co without structural mask · ImageNet-256 (200 ep, CFG=2.9) | FID: 6.69 -> 24.75 | Hurt |
| add perceptual-drifting hybrid loss to V-Co (ablation) | V-Co-B/16 · ImageNet-256 (300 ep, guided) | FID: 2.96 -> 2.44 | Improved |
| add REPA loss on top of co-denoising (ablation) | V-Co-B/16 · ImageNet-256 (300 ep, guided) | FID: 2.96 -> 2.91 | No clear effect |
| add RMS-based feature rescaling of semantic features (ablation) | V-Co-B/16 · ImageNet-256 (200 ep, guided) | FID: 5.28 -> 2.52 | Improved |
| use RMS scaling instead of noise-schedule shifting (ablation) | V-Co-B/16 · ImageNet-256 (200 ep, guided) | FID: 2.93 -> 2.52 | Improved |
Closest papers by meaning
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.90)
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.89)
- Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think (NeurIPS 2025 · similarity 0.89)
- DSAQuant: Denoising-Stage-Aligned Quantization-Aware Training for Video Generation (NeurIPS 2026 · similarity 0.88)
- Denoising Autoregressive Representation Learning (ICML 2024 · similarity 0.88)
- ViewFusion: Towards Multi-View Consistency via Interpolated Denoising (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
