Research map / Image generation / Autoregressive and latent generation

Diff-CA: Separating Common and Salient Factors with Diffusion Models

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: High-resolution diffusion synthesis · 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 Diff-CA, a diffusion-based contrastive analysis method that trains an image-conditioned flow-matching generator on DINOv3-derived tokens and a separator encoder that splits the conditioning into common and salient codes under weak binary supervision.

Problem

Contrastive Analysis: separating factors common to two unpaired image distributions from factors salient to only one, using only dataset-level binary labels; prior CA methods built on VAEs/GANs suffer from poor reconstruction and image quality.

Approach

Decompose diffusion conditioning tokens (not a global latent) additively into a salient code and a common code, enforced by reconstruction, pinning, cycle-consistency and gradient-reversal adversarial losses, with a theorem showing identifiability of the additive factors under these constraints.

Key results

• Diff-CA reconstruction on FFHQ glasses: SSIM 0.610, LPIPS 0.115, ID-Sim 0.522, versus best baseline DoubleInfoGAN SSIM 0.326 and LPIPS 0.430.
• Glasses swap fine-grained accuracy 94.5% for Diff-CA versus 47.0% for DoubleInfoGAN.
• Common-factor preservation after swap: gender 92.5%, smile 92.9%, pose MAE 1.87 degrees, versus 68.5%, 79.2%, 7.35 for DoubleInfoGAN.

Surprising findings

• Prompt-based T2I editors (FLUX.1 Kontext, Nano Banana Pro) are reported to add generic attributes or alter anatomy on medical images, though these are qualitative failure cases only.

Limitations

• Limited to 2D images; not adapted to 3D volumetric medical data.
• Assumes additive factorization; settings where salient factors appear in both distributions are not covered.
• Identifiability proof relies on assumptions (e.g., moment conditions, linear additive latent space) that are only empirically motivated.

Experimental setup

Datasets: FFHQ CelebA-HQ AFHQ BraTS 2023

Benchmarks: FFHQ glasses swapping CelebA-HQ reconstruction AFHQ BraTS 2023

Metrics: SSIM LPIPS ID-Sim accuracy MAE

Baselines: MM-cVAE SepVAE DoubleInfoGAN DiffAE EncDiff

Builds on: DiffAE Perceiver-IO OT-CFM DINOv3 MM-cVAE

Closest papers by meaning

• NoiseCLR: A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Directions in Diffusion Models (CVPR 2024 · similarity 0.88)
• Erasing Undesirable Influence in Diffusion Models (CVPR 2025 · similarity 0.88)
• High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.88)
• Adapting Diffusion Models for Improved Prompt Compliance and Controllable Image Synthesis (NeurIPS 2024 · similarity 0.87)
• Contrastive Denoising Score for Text-guided Latent Diffusion Image Editing (CVPR 2024 · similarity 0.87)
• Unsupervised Region-Based Image Editing of Denoising Diffusion Models (AAAI 2025 · similarity 0.87)