Research map / Image generation / Fast diffusion sampling
CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Sampling
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 CADS, a training-free inference method that anneals the conditioning vector with scheduled Gaussian noise to boost diffusion sample diversity at high guidance while keeping quality, reaching FID 1.70 on ImageNet 256x256.
Problem
Conditional diffusion models lose output diversity under high classifier-free guidance or when trained on small datasets, trading diversity for image quality.
Approach
Add scheduled, monotonically decreasing Gaussian noise to the condition vector during inference (strong corruption early, zero by the end) so early sampling relies less on the condition, with optional rescaling of the noisy condition; it plugs into any pretrained model and sampler without retraining.
Key results
- CADS achieves a new state-of-the-art FID of 1.70 on class-conditional ImageNet 256x256 and 2.31 at 512x512 using an existing pretrained DiT-XL/2.
- On ImageNet with DDPM at high guidance, CADS improves FID from 20.83 to 9.47 and Recall from 0.32 to 0.62; Dynamic CFG reaches FID 18.42.
- CADS keeps top-1 class accuracy at 0.96 versus 0.98 for DDPM, and pose MPJPE and CLIP-score are unchanged.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add condition-annealed Gaussian noise during sampling instead of standard DDPM sampling | DiT-XL/2 · ImageNet class-conditional generation | FID: 20.83 -> 9.47 | Helped |
| Add condition-annealed Gaussian noise during sampling instead of standard DDPM sampling | DiT-XL/2 · ImageNet class-conditional generation | Recall: 0.32 -> 0.62 | Helped |
| Use condition-annealed sampling instead of Dynamic CFG | DiT-XL/2 · ImageNet class-conditional generation | FID: 18.42 -> 9.47 | Helped |
| Use condition-annealed sampling instead of Dynamic CFG | DiT-XL/2 · ImageNet class-conditional generation | Recall: 0.39 -> 0.62 | Helped |
| Add condition-annealed noise during sampling | DiT-XL/2 · ImageNet class-conditional generation | Top-1 Class Accuracy: 0.98 -> 0.96 | Hurt |
| Add condition-annealed noise during sampling | Pose-to-image model · Pose-to-image generation | MPJPE: 0.02 -> 0.02 | No clear effect |
| Add condition-annealed noise during sampling | Stable Diffusion · Text-to-image generation | CLIP-Score: 0.31 -> 0.31 | No clear effect |
| Set τ₂ to 0.5 instead of 0.9 (ablation) | CADS · ImageNet class-conditional generation, CFG 5 | FID: 9.47 -> 8.21 | Helped |
| Set τ₂ to 0.5 instead of 0.9 (ablation) | CADS · ImageNet class-conditional generation, CFG 2.5 | FID: 5.02 -> 6.14 | Hurt |
| Set τ₂ to 0.5 instead of 1 (ablation) | CADS · Pose-to-image generation | FID: 10.93 -> 8.26 | Helped |
| Use a polynomial annealing schedule instead of the piecewise-linear schedule (ablation) | CADS · ImageNet class-conditional generation | FID: Baseline 14.20; Polynomial, d=1 15.85; Polynomial, d=2 14.98; Polynomial, d=3 14.49; Polynomial, d=4 14.32 | Hurt |
| Use minimal conditioning noise (s=0.025) (ablation) | CADS · Conditional generation | diversity | Hurt |
Surprising findings
- Standard IS and Precision give artificially high scores to low-diversity DDPM samples (Precision 0.92 vs 0.82 for CADS), so FID and Recall are more reliable here.
- Training the diffusion model with noisy conditions reduced CADS effectiveness, suggesting the diversity issue lies mainly in inference.
Limitations
- Applying CADS to dense spatial conditions such as segmentation maps is left as challenging future work.
- Rescaling trades diversity for stability; noise scale and thresholds require tuning per task and guidance scale.
- Adding noise to pose images during training instead of inference reduces CADS effectiveness.
Experimental setup
Datasets: ImageNet DeepFashion SHHQ COCO
Benchmarks: ImageNet 256x256 ImageNet 512x512 DeepFashion SHHQ COCO
Metrics: FID Recall Precision Inception Score Mean Similarity Score Vendi Score top-1 accuracy MPJPE CLIP-Score
Baselines: DDPM Dynamic CFG MDT DDIM PNDM DPM++
Builds on: classifier-free guidance DDPM DDIM Dynamic CFG
Closest papers by meaning
- A Simple Approach to Unifying Diffusion-based Conditional Generation (ICLR 2025 · similarity 0.91)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.91)
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.90)
- EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.89)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.89)
- One Step Diffusion via Shortcut Models (ICLR 2025 · similarity 0.89)
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.