Research map / Image generation / Autoregressive and latent generation
Exploring DCN-like architecture for fast image generation with arbitrary resolution
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 FlowDCN, a purely convolution-based flow-matching diffusion model built on group-wise multiscale deformable convolution, for fast arbitrary-resolution image generation with linear complexity.
Problem
Arbitrary-resolution image generation with diffusion transformers suffers from quadratic attention cost and poor resolution extrapolation due to absolute position embeddings.
Approach
Decouples deformable convolution's sampling field into a learnable scale and a fixed direction, assigns different scale priors to channel groups (group-wise multiscale deformable block), removes long residual connections, and rescales the maximum sampling scale S_max by the train/test resolution ratio for resolution extrapolation.
Key results
- FlowDCN-XL/2 with CFG 1.375 and ODE solver reaches 2.13 FID and 4.30 sFID on ImageNet 256x256 at 1.5M steps, versus SiT-XL/2 at 2.06 FID and 4.50 sFID (CFG).
- FlowDCN-XL/2 uses about 21% fewer FLOPs (93.24G vs 118.64G) and 8.4% fewer parameters (618M vs 675M) than SiT-XL/2.
- On 320x320 with base-size models, FlowDCN-B/2 reaches 34.4 FID versus 61.4 for FiT-B and 95.5 for DiT-B.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FlowDCN-S/2 instead of SiT-S/2 | CIFAR10 32×32 image generation · CIFAR10 32×32 | FID: 7.42 -> 5.47 | Helped |
| remove group-wise multiscale deformable convolution (ablation) | FlowDCN-S/2 · CIFAR10 32×32 | FID: 5.47 -> 5.72 | Hurt |
| use random prior initialization instead of predefined direction grids and linearly increased scale priors (ablation) | FlowDCN-S/2 · CIFAR10 32×32 | FID: 5.47 -> 5.68 | Hurt |
| increase kernel size from 9 to 32 sampling points (ablation) | FlowDCN-S/2 · CIFAR10 32×32 | FID: 5.47 -> 5.13 | Helped |
| learn direction priors instead of fixing them (ablation) | FlowDCN-S/2 with fixed relative scale · CIFAR10 32×32 | FID: 5.6 -> 6.01 | Hurt |
| learn relative scale rather than fixing it (ablation) | FlowDCN-S/2 with fixed direction priors · CIFAR10 32×32 | FID: 5.6 -> 5.47 | Helped |
| replace RMSNorm and SwiGLU with LayerNorm and FFN (ablation) | FlowDCN-B/2 · ImageNet 256×256, 400k steps | FID: 28.5 -> 29.1 | Hurt |
| replace deformable convolution with standard group-wise convolution (ablation) | FlowDCN-S/2 · ImageNet 256×256, 400k steps | FID: 54.6 -> 59.0; 54.6 -> 63.0 | Hurt |
| adjust Smax for inference resolution (ablation) | FlowDCN-B/2 · ImageNet 320×320 and 224×448 | FID: 34.4 -> 35.7; 71.7 -> 81.1 | Mixed |
| use FlowDCN-XL/2 instead of SiT-XL/2 | ImageNet 256×256 generation with CFG=1.375 and ODE · ImageNet 256×256 | sFID: 4.50 -> 4.30 | Helped |
| increase FlowDCN training from 384M to 486M total images | FlowDCN-XL/2 with CFG=1.375 and ODE · ImageNet 256×256 | FID: 2.13 -> 2.01 | Helped |
| use FlowDCN-XL/2 instead of SiT-XL/2-G | ImageNet 512×512 generation · ImageNet 512×512 | FID: 2.62 -> 2.44 | Helped |
Surprising findings
- Fixed direction priors outperform learnable direction priors for the deformable field.
- Base-size FlowDCN extrapolates poorly to 224x448 without various-aspect-ratio training (71.7 FID vs 44.7 for FiT-B), and S_max adjustment does not consistently improve FID/sFID.
Limitations
- Deformable convolution backward pass is inefficient relative to attention.
- Resolution extrapolation works only within a reasonable aspect-ratio range set by training data.
- Scale Adjustment does not uniformly improve FID/sFID; FID/sFID may not capture semantic quality.
Experimental setup
Datasets: ImageNet CIFAR-10
Benchmarks: ImageNet 256x256 ImageNet 512x512 CIFAR-10 32x32 ImageNet arbitrary-resolution extrapolation (320x320, 224x448, 160x480)
Metrics: FID sFID Inception Score Precision Recall FLOPs Latency
Baselines: DiT-XL/2 SiT-XL/2 FiT-XL/2 DiffusionSSM-XL ADM-U LDM-4 U-ViT-H/2 StyleGAN-XL BigGAN-deep
Builds on: DiT SiT deformable convolution (DCN) DCNv4 FiT
Closest papers by meaning
- Native-Resolution Image Synthesis (NeurIPS 2025 · similarity 0.88)
- DemoFusion: Democratising High-Resolution Image Generation With No $$$ (CVPR 2024 · similarity 0.88)
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · similarity 0.86)
- DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.86)
- STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis (NeurIPS 2025 · similarity 0.86)
- Boosting Flow-based Generative Super-Resolution Models via Learned Prior (CVPR 2024 · similarity 0.85)
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.