Research map / Image generation / Text-to-image diffusion
Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation
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 Angle Domain Guidance (ADG), which rotates the conditional x0-prediction in angle space instead of linearly extrapolating it, to curb CFG's norm amplification and color distortion at high guidance weights.
Problem
Classifier-free guidance at high guidance weights amplifies latent norms in text-to-image latent diffusion, causing oversaturation, color distortion and degraded images.
Approach
Replace linear extrapolation of the conditional x0-prediction with a rotation in the angular domain (angle between conditional and unconditional x0 predictions scaled by guidance weight, capped at pi/3), which preserves latent norm while enhancing text alignment.
Key results
- On SD v3.5 at 10 NFE, ADG reaches ImageReward 0.970 at omega=6/10 versus best CFG 0.843 with FID 16.6 comparable to baselines.
- At omega=15 CFG drops to ImageReward -0.659 while ADG holds 0.940 with FID 16.8.
- On SD v2.1 with DPM-Solver (25 NFE, omega=8), ADG improves CLIP 0.322 vs 0.313 and ImageReward 0.467 vs 0.399 over CFG with FID tied at 16.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace CFG with ADG | SD v3.5 text-to-image generation · COCO10k, best performance | ImageReward: 0.843 -> 0.970 | Helped |
| replace CFG with ADG | SD v3.5 text-to-image generation · COCO10k, best performance | CLIP: 0.318 -> 0.324 | Helped |
| replace CFG with ADG | SD v3.5 text-to-image generation · COCO10k, best performance | FID: 16.6 -> 16.6 | No clear effect |
| replace CFG with ADG | SD v2.1 with DPM-Solver · COCO10k, ω=8 | CLIP: 0.313 -> 0.322 | Helped |
| replace CFG with ADG | SD v2.1 with DPM-Solver · COCO10k, ω=8 | ImageReward: 0.399 -> 0.467 | Helped |
| replace CFG with ADG | SD v2.1 with DPM-Solver · COCO10k, ω=8 | FID: 16.6 -> 16.6 | No clear effect |
| remove maximum turning angle constraint (ablation) | ADG · COCO10k, SD v3.5, ω=8 | CLIP: 0.322 -> 0.275 | Hurt |
| remove maximum turning angle constraint (ablation) | ADG · COCO10k, SD v3.5, ω=8 | ImageReward: 0.970 -> -0.782 | Hurt |
| remove maximum turning angle constraint (ablation) | ADG · COCO10k, SD v3.5, ω=8 | FID: 16.7 -> 28.6 | Hurt |
| add normalization of x̂₀,ω (ablation) | ADG · COCO10k, SD v3.5, ω=8 | CLIP: 0.322 -> 0.322 | No clear effect |
| add normalization of x̂₀,ω (ablation) | ADG · COCO10k, SD v3.5, ω=8 | ImageReward: 0.970 -> 0.958 | No clear effect |
| add normalization of x̂₀,ω (ablation) | ADG · COCO10k, SD v3.5, ω=8 | FID: 16.7 -> 16.6 | No clear effect |
Surprising findings
- Removing the maximum turning-angle constraint causes catastrophic failure (ImageReward -0.782 at omega=8), showing the angle cap is essential.
- CFG++ with lambda=omega/12.5 degrades sharply at high guidance (ImageReward -0.980 at omega=6).
Limitations
- The method is heuristic and lacks formal theoretical guarantees, as the authors acknowledge.
- The theoretical analysis uses Gaussian mixture assumptions that may not hold for real latent distributions.
- Evaluation is limited to COCO prompts and two Stable Diffusion models, with only FID, CLIP and ImageReward metrics.
Experimental setup
Datasets: COCO
Benchmarks: COCO10k
Metrics: FID CLIP score ImageReward
Baselines: CFG CFG++ APG
Builds on: classifier-free guidance CFG++ APG DDIM DPM-Solver
Closest papers by meaning
- Rethinking the Spatial Inconsistency in Classifier-Free Diffusion Guidance (CVPR 2024 · similarity 0.89)
- Towards Transformer-Based Aligned Generation with Self-Coherence Guidance (CVPR 2025 · similarity 0.88)
- Towards a Golden Classifier-Free Guidance Path via Foresight Fixed Point Iterations (NeurIPS 2025 · similarity 0.88)
- Color Alignment in Diffusion (CVPR 2025 · similarity 0.87)
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.87)
- Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models (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.