Research map / Image generation / Text-to-image diffusion

CASL: Concept-Aligned Sparse Latents for Interpreting Diffusion Models

NeurIPS 2026 · Image generation › Text-to-image diffusion · approach: Multi-concept composition · 0 citations (OpenAlex)

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 CASL, which trains a sparse autoencoder on frozen diffusion U-Net bottleneck activations, aligns latent dimensions to concepts via a linear mapping, and uses CASL-Steer as a causal probe plus a new Editing Precision Ratio metric.

Problem

Unsupervised SAE features in diffusion models are hard to map to human-understandable concepts, limiting reliable semantic control and interpretation of generated images.

Approach

Supervised alignment of SAE latent dimensions with semantic concepts via a lightweight linear map predicting activation-space edit directions (trained with CLIP and L1 losses), validated by a steering probe and scored with EPR, a ratio of target-attribute change to non-target-attribute change.

Key results

• On CelebA-HQ smiling, CASL-Steer reaches EPR 4.465 versus 3.359 for the next best (SwiftEdit) and LPIPS 0.243 with ArcFace 0.566.
• Single top-1 latent dimension linear SVM probe achieves 0.655-0.757 accuracy across five concepts, rising to above 0.93 with top-16 dimensions.
• SAE with expansion ratio 128 reaches MSE 1.91e-2, cosine similarity 0.993 and DAR 1.08e-2 on CelebA-HQ.

Surprising findings

• The authors report EPR varies substantially across classifier backbones (ResNet18, VGG16, MobileNetV2, ViT-B/16), with ViT-B/16 producing unstable scores, so EPR should be read as a relative metric.
• CASL-Steer does not achieve the best CLIP-Score on all concepts (e.g., Big Nose and Blond Hair where MasaCtrl or Asyrp score higher) and lower ArcFace than Slider for Young.

Limitations

• EPR depends on labeled attributes and the choice of classifier, so absolute values are not comparable across backbones or datasets.
• Editing strength alpha is chosen by visual inspection rather than by a quantitative criterion.
• Evaluation is limited to facial attributes plus a few LSUN-Church and AFHQ-Dog concepts; the results rely on 32 sampled test images per attribute.

Experimental setup

Datasets: FFHQ CelebA-HQ LSUN-Church AFHQ-Dog CelebA

Benchmarks: CelebA-HQ

Metrics: EPR CLIP-Score LPIPS ArcFace similarity MSE cosine similarity DAR SVM top-k accuracy

Baselines: BoundaryDiffusion Asyrp Concept Sliders MasaCtrl SwiftEdit

Builds on: Asyrp DiffusionCLIP Sparse Autoencoders DDIM

Closest papers by meaning

• You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.90)
• The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.90)
• Unveiling Concept Attribution in Diffusion Models (NeurIPS 2025 · similarity 0.89)
• Concept Reachability in Diffusion Models: Beyond Dataset Constraints (ICML 2025 · similarity 0.89)
• Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning (NeurIPS 2026 · similarity 0.89)
• Unsupervised Region-Based Image Editing of Denoising Diffusion Models (AAAI 2025 · similarity 0.89)