Research map / Image generation / Fast diffusion sampling
Deeply Supervised Flow-Based Generative Models
Inkyu Shin, Chenglin Yang, Liang-Chieh Chen
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 deep supervision (ablation) | SiT-B/2 · ImageNet-256 (B/2, 80 ep, uniform, no CFG) | FID: 34.4 -> 33.0 | Improved |
| add time-gap between branch time-steps (ablation) | SiT-B/2 with deep supervision · ImageNet-256 (B/2, 80 ep, uniform, no CFG) | FID: 33.0 -> 31.1 | Improved |
| add inter-layer acceleration in VeRA block (ablation) | SiT-B/2 with deep supervision and time-gap · ImageNet-256 (B/2, 80 ep, uniform, no CFG) | FID: 31.1 -> 29.9 | Improved |
| add cross-space attention in VeRA block (ablation) | DeepFlow-B/2-2T · ImageNet-256 (B/2, 80 ep, uniform, no CFG) | FID: 29.9 -> 28.1 | Improved |
| modulate concatenated velocity and acceleration instead of adding modulated acceleration to velocity (ablation) | DeepFlow-B/2-2T VeRA block · ImageNet-256 (B/2, 80 ep, uniform, no CFG) | FID: 29.3 -> 28.1 | Improved |
| deeper ACC_MLP {2048, 4096, 2048, 768} instead of {2048, 768} (ablation) | DeepFlow-B/2-2T · ImageNet-256 (B/2, 80 ep, uniform, no CFG) | FID: 30.4 -> 28.1 | Improved |
| vary maximum time-gap alpha from 0.1 to 0.001 (ablation) | DeepFlow-B/2-2T · ImageNet-256 (B/2, 80 ep, uniform, no CFG) | FID: 28.9 (0.1), 28.6 (0.05), 28.1 (0.01), 28.3 (0.005), 28.5 (0.001) | No clear effect |
| vary acceleration loss weight lambda from 0.5 to 1.5 (ablation) | DeepFlow-B/2-2T · ImageNet-256 (B/2, 80 ep, uniform, no CFG) | FID: 27.8 (0.5), 28.1 (1.0), 28.8 (1.5) | Mixed |
| replace SiT-B/2 with DeepFlow-B/2-2T (no SSL, lognormal) | SiT-B/2 · ImageNet-256 (B/2, 80 ep, no CFG) | FID: 29.7 -> 23.1 | Improved |
| add DINOv2 SSL alignment to DeepFlow-B/2-2T (lognormal) | DeepFlow-B/2-2T · ImageNet-256 (B/2, 80 ep, no CFG) | FID: 23.1 -> 17.2 | Improved |
| replace SiT-XL/2 with DeepFlow-XL/2-3T (no SSL, 80 ep) | SiT-XL/2 · ImageNet-256 (XL/2, 80 ep, no CFG) | FID: 13.8 -> 10.3 | Improved |
| apply DINOv2 alignment twice in SiT | SiT · ImageNet-256 | FID | No clear effect |
Closest papers by meaning
- Mean Flows for One-step Generative Modeling (NeurIPS 2025 · similarity 0.90)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.88)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.88)
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.87)
- Asymmetric Flow Models (NeurIPS 2026 · similarity 0.87)
- HiFlow: Training-free High-Resolution Image Generation with Flow-Aligned Guidance (NeurIPS 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.
