Research map / Image generation / Autoregressive and latent generation
SSG: Scaled Spatial Guidance for Multi-Scale Visual Autoregressive Generation
Youngwoo Shin, Jiwan Hur, Junmo Kim
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 |
|---|---|---|---|
| add SSG guidance | VAR-d30 · ImageNet 256x256 | FID: 2.02 -> 1.68 | Improved |
| add SSG guidance | VAR-d16 · ImageNet 256x256 | FID: 3.42 -> 3.27 | Improved |
| add SSG guidance | VAR-d36 · ImageNet 512x512 | FID: 2.70 -> 2.39 | Improved |
| add SSG guidance | HART-0.7B · MJHQ-30K | FID: 8.46 -> 7.28 | Improved |
| add SSG guidance | Infinity-2B · MJHQ-30K | FID: 10.01 -> 9.68 | Improved |
| add SSG guidance | VAR-d30 · ImageNet 256x256 | precision: 0.82 -> 0.81 | No clear effect |
| add SSG guidance | VQ-Diffusion (adapted) · ImageNet 256x256 | FID: 9.39 -> 9.18 | Improved |
| use DSE frequency prior instead of nearest-neighbor spatial prior (ablation) | SSG on VAR-d16 · ImageNet 256x256 | FID: 4.02 -> 3.27 | Improved |
| use DSE frequency prior instead of linear spatial prior (ablation) | SSG on VAR-d16 · ImageNet 256x256 | FID: 3.79 -> 3.27 | Improved |
| extrapolate high frequencies from interpolated logits instead of zero padding in DSE (ablation) | SSG on VAR-d16 · ImageNet 256x256 | FID: 3.34 -> 3.27 | Improved |
| add linear beta_k decay instead of static beta_k (ablation) | SSG on VAR-d16 · ImageNet 256x256 | FID: 3.63 -> 3.27 | Improved |
| add linear beta_k decay instead of static beta_k (ablation) | SSG on VAR-d16 · ImageNet 256x256 | IS: 287.8 -> 285.3 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Infinity-2B — Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025) · better on 2 of 2 shared results
- HART-0.7B — HART: Efficient Visual Generation with Hybrid Autoregressive Transformer (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Rethinking Visual Autoregressive Sampling with Information-Grounding Guidance (ICML 2026 · similarity 0.92)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.91)
- SynVAR: Synergizing Spatial and Semantic Alignment in Visual Autoregressive Model (ECCV 2026 · similarity 0.90)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.89)
- NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering (NeurIPS 2025 · similarity 0.88)
- FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 2025 · 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.
