Research map / Image generation / Autoregressive and latent generation
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
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 FlowAR, a scale-wise autoregressive image generator with a simple doubling scale schedule, off-the-shelf continuous VAE latents, and a spatially adaptive conditioned flow matching head, reaching FID 1.65 on ImageNet-256.
Problem
VAR's next-scale prediction relies on a rigid, non-uniform scale schedule and a multi-scale residual discrete VQGAN tokenizer that is tightly coupled to the generator, limiting generalization and tokenizer flexibility.
Approach
Use a doubling scale sequence obtained by downsampling the finest VAE latent, and replace categorical per-scale prediction with a flow matching head that is conditioned on the autoregressive Transformer's per-scale semantics through position-wise adaptive LayerNorm (Spatial-adaLN).
Key results
- FlowAR-H (1.9B) reaches FID 1.65 on ImageNet-256, versus VAR-d30 (2.0B) at 1.97.
- FlowAR-L (589M) achieves FID 1.90 versus VAR-d20 (600M) at 2.95.
- With the simplified scales {1,2,4,8,16}, FlowAR reaches FID 3.61 while VAR fails to train on that scale set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace VAR with FlowAR-H at comparable scale | scale-wise autoregressive image generation · ImageNet-256 | FID: 1.97 -> 1.65 | Helped |
| use FlowAR-L instead of VAR-d20 | scale-wise autoregressive image generation · ImageNet-256 | FID: 2.95 -> 1.90 | Helped |
| use a VAE tokenizer instead of VAR's multi-scale residual discrete tokenizer (ablation) | FlowAR · ImageNet-256 | FID: 5.81 -> 4.22; 5.81 -> 3.94; 5.81 -> 3.61 | Helped |
| downsample latents rather than input images to construct scales (ablation) | FlowAR · ImageNet-256 | FID: 12.19 -> 3.61 | Helped |
| reduce VAR's ten-scale sequence to {1, 2, 4, 8, 16} (ablation) | VAR · ImageNet-256 | FID | Hurt |
| reduce FlowAR scales to {1, 4, 8, 16} (ablation) | FlowAR · ImageNet-256 | FID: 3.61 -> 4.88 | Hurt |
| reduce FlowAR scales to {1, 4, 16} (ablation) | FlowAR · ImageNet-256 | FID: 3.61 -> 6.10 | Hurt |
| use per-scale instead of per-token prediction (ablation) | FlowAR scale probability model · ImageNet-256 | FID: 6.85 -> 3.93; 6.15 -> 3.61 | Helped |
| use flow matching instead of diffusion for per-token prediction (ablation) | FlowAR scale probability model · ImageNet-256 | FID: 6.85 -> 6.15 | Helped |
| use flow matching instead of diffusion for per-scale prediction (ablation) | FlowAR scale probability model · ImageNet-256 | FID: 3.93 -> 3.61 | Helped |
| add Spatial-adaLN instead of adaLN (ablation) | FlowAR semantic injection · ImageNet-256 | FID: 14.28 -> 3.61 | Helped |
| use Spatial-adaLN instead of sequence concatenation (ablation) | FlowAR semantic injection · ImageNet-256 | FID: 9.22 -> 3.61 | Helped |
Surprising findings
- VAR training fails entirely when its scale schedule is simplified to {1,2,4,8,16}, indicating strong tokenizer-generator coupling.
- Downsampling the VAE latents yields FID 3.61 while downsampling the image before the VAE gives FID 12.19.
Limitations
- Compute and hardware for training are not reported.
- Evaluation limited to class-conditional ImageNet-256.
Experimental setup
Datasets: ImageNet-256
Benchmarks: ImageNet-256
Metrics: FID IS precision recall
Baselines: VAR LlamaGen DiT SiT MAR StyleGAN BigGAN GigaGAN U-ViT ADM LDM RQTransformer
Builds on: VAR flow matching MAR DiT
Closest papers by meaning
- Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens (ICLR 2025 · similarity 0.88)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.87)
- Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective (NeurIPS 2024 · similarity 0.87)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.87)
- EditAR: Unified Conditional Generation with Autoregressive Models (CVPR 2025 · similarity 0.87)
- Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching (ICLR 2025 · similarity 0.87)
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.