Research map / Image generation / Fast diffusion sampling
Improved Training Technique for Shortcut Models
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 iSM, a training framework for shortcut generative models that adds intrinsic (input-conditioned) guidance, interval guidance, multi-level wavelet loss, scaling optimal transport, and twin EMA, reaching FID 5.27 (1-step) and 1.93 (8-step) on ImageNet 256x256.
Problem
Shortcut models for one-, few-, and multi-step generation suffer from compounding and inflexible classifier-free guidance, low-frequency bias from pixel-wise losses, EMA-induced inconsistency in self-consistency targets, and high-curvature flow trajectories from random noise-data pairing.
Approach
Make the guidance scale an explicit network input trained to output guided velocities directly (with interval guidance, stop-gradient), replace pixel L2 with a multi-level DWT loss, scale OT by pooling K mini-batches, and use two EMA networks (fast for consistency targets, slow for inference).
Key results
- iSM reaches FID 5.27 at 1 step and 2.05 at 4 steps on ImageNet 256x256, versus 10.60 and 7.80 for the original shortcut model.
- iSM achieves FID 1.93 at 8 steps and 1.88 at 128 steps, with 8-step performance comparable to VAR-d30 (2B).
- On ImageNet 512x512 with FlowDCN, 1-step recall rises from 0.11 to 0.55 and 4-step FID drops from 12.16 to 9.94.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace fixed guidance with intrinsic guidance (w_max 3.5) (ablation) | Shortcut Model · ImageNet 256×256, 1 and 4 NFE | FID: 21.38 -> 9.62; 13.46 -> 3.17 | Helped |
| Set guidance interval threshold to 0.3 (ablation) | iSM with intrinsic guidance · ImageNet 256×256, 1 and 4 NFE | FID: 9.62 -> 8.49; 3.17 -> 2.81 | Helped |
| Raise t_interval from 0.3 to 0.5 (ablation) | iSM with interval guidance · ImageNet 256×256, 1 NFE | FID: 8.49 -> 19.22 | Hurt |
| Use 5-level rather than no wavelet decomposition (ablation) | iSM with interval guidance · ImageNet 256×256, 1 and 4 NFE | FID: 8.49 -> 8.12; 2.81 -> 2.64 | Helped |
| Increase wavelet levels from 1 to 3 to 5 (ablation) | iSM with interval guidance · ImageNet 256×256, 1 and 4 NFE | FID: 8.21 -> 8.17 -> 8.12; 2.79 -> 2.75 -> 2.64 | Helped |
| Use scaling OT with K=32 instead of no OT matching (ablation) | iSM with 5-level wavelet loss · ImageNet 256×256, 1 and 4 NFE | FID: 8.12 -> 7.97; 2.64 -> 2.23 | Helped |
| Increase scaling OT from K=1 to K=8 to K=32 (ablation) | iSM with 5-level wavelet loss · ImageNet 256×256, 1 and 4 NFE | FID: 8.07 -> 8.03 -> 7.97; 2.51 -> 2.28 -> 2.23 | Helped |
| Use Twin EMA target decay 0.95 rather than 0.9999 (ablation) | iSM with scaling OT · ImageNet 256×256, 1 and 4 NFE | FID: 7.97 -> 6.56; 2.23 -> 2.16 | Helped |
| Set intrinsic guidance w_max to 3.5 instead of 2.0 or 5.0 (ablation) | iSM · ImageNet 256×256, 1 and 4 NFE | FID: 1 NFE: 10.10 -> 9.62 -> 10.38; 4 NFE: 3.21 -> 3.17 -> 3.34 | Mixed |
| Increase t_interval from 0.0 to 0.1 to 0.3 (ablation) | iSM with intrinsic guidance · ImageNet 256×256, 1 and 4 NFE | FID: 9.62 -> 8.58 -> 8.49; 3.17 -> 3.14 -> 2.81 | Helped |
| Use iSM rather than SM | SiT-XL/2 · ImageNet 256×256, 1 NFE | FID-50k: 10.60 -> 5.27 | Helped |
| Use iSM rather than SM | SiT-XL/2 · ImageNet 256×256, 4 NFE | FID-50k: 7.80 -> 2.05 | Helped |
Surprising findings
- Traditional mini-batch OT lowered FID compared with not using OT in the ablation, while the scaled OT (K=32) improved it further.
- Resuming training from flow-matching weights gave limited improvement for the original shortcut model but considerable gains for iSM.
Limitations
- No error bars or multiple seeds reported due to evaluation cost.
- Evaluation limited to ImageNet class-conditional generation; text-conditioned settings not studied.
- Hyperparameters (w_max, t_interval, L, K, EMA rates) selected via ablations on the same benchmark.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet 256x256 ImageNet 512x512
Metrics: FID-50k FD-DINOv2 Inception Score Precision Recall
Baselines: Shortcut Models IMM iCT VAR-d20 VAR-d30 MAR DiT-XL/2 SiT-XL/2 REPA-XL/2 FlowDCN StyleGAN-XL BigGAN-deep
Builds on: shortcut models flow matching classifier-free guidance inductive moment matching SiT
Closest papers by meaning
- One Step Diffusion via Shortcut Models (ICLR 2025 · similarity 0.91)
- Do ImageNet-trained Models Learn Shortcuts? The Impact of Frequency Shortcuts on Generalization (CVPR 2025 · similarity 0.89)
- On the Foundations of Shortcut Learning (ICLR 2024 · similarity 0.88)
- Visual Generation Without Guidance (ICML 2025 · similarity 0.88)
- Increasing the Utility of Synthetic Images through Chamfer Guidance (NeurIPS 2025 · similarity 0.87)
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (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.