mlsanityData to 14 Oct 2026

Representation Fr'echet Loss for Visual Generation

arXiv 2604.28190 ↗First posted 30 Apr 2026cs.CVcited by 37

Not on the map. Why?

Abstract (arXiv)

We show that Fr'echet Distance (FD), long considered impractical as a training objective, can in fact be effectively optimized in the representation space. Our idea is simple: decouple the population size for FD estimation (e.g., 50k) from the batch size for gradient computation (e.g., 1024). We term this approach FD-loss. Optimizing FD-loss reveals several surprising findings. First, post-training a base generator with FD-loss in different representation spaces consistently improves visual quality. Under the Inception feature space, a one-step generator achieves0.72 FID on ImageNet 256x256. Second, the same FD-loss repurposes multi-step generators into strong one-step generators without teacher distillation, adversarial training or per-sample targets. Third, FID can misrank visual quality: modern representations can yield better samples despite worse Inception FID. This motivates FDrk, a multi-representation metric. We hope this work will encourage further exploration of distributional distances in diverse representation spaces as both training objectives and evaluation metrics for generative models.

Key facts. Cited by 37 papers in the map (data through 14 Oct 2026). Builds on 51 of its 74 references that are in the map. First posted on arXiv 30 Apr 2026.

Builds on

51 of 74 references.

References · 51

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  2. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,974
  3. 2019Momentum Contrast for Unsupervised Visual Representation Learningcited by 7,513
  4. 2021Masked Autoencoders Are Scalable Vision Learnerscited by 7,403
  5. 2015Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shiftcited by 7,005
  6. 2022Scalable Diffusion Models with Transformerscited by 6,054
  7. 2015Rethinking the Inception Architecture for Computer Visioncited by 5,475
  8. 2024Scaling Rectified Flow Transformers for High-Resolution Image Synthesiscited by 4,181
  9. 2017Deep reinforcement learning from human preferencescited by 3,912
  10. 2022Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flowcited by 3,700
  11. 2022A ConvNet for the 2020scited by 3,579
  12. 2016Improved Techniques for Training GANscited by 2,970
38 more
  1. 2018Unsupervised Feature Learning via Non-Parametric Instance-level Discriminationcited by 1,845
  2. 2023Consistency Modelscited by 1,822
  3. 2018Demystifying MMD GANscited by 1,293
  4. 2025SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Featurescited by 1,188
  5. 2022Scaling Laws for Reward Model Overoptimizationcited by 1,095
  6. 2023One-step Diffusion with Distribution Matching Distillationcited by 1,019
  7. 2019Improved Precision and Recall Metric for Assessing Generative Modelscited by 944
  8. 2024SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformerscited by 737
  9. 2024Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Thinkcited by 703
  10. 2024Autoregressive Image Generation without Vector Quantizationcited by 651
  11. 2025Mean Flows for One-step Generative Modelingcited by 561
  12. 2018Assessing Generative Models via Precision and Recallcited by 406
  13. 2025Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Modelscited by 393
  14. 2025BLIP3-o: A Family of Fully Open Unified Multimodal Models-Architecture, Training and Datasetcited by 376
  15. 2023Rethinking FID: Towards a Better Evaluation Metric for Image Generationcited by 310
  16. 2025Diffusion Transformers with Representation Autoencoderscited by 295
  17. 2024Simplifying, Stabilizing and Scaling Continuous-Time Consistency Modelscited by 283
  18. 2015Generative Moment Matching Networkscited by 257
  19. 2023Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Modelscited by 236
  20. 2025REPA-E: Unlocking VAE for End-to-End Tuning with Latent Diffusion Transformerscited by 225
  21. 2024Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generationcited by 214
  22. 2022The Role of ImageNet Classes in Fr'echet Inception Distancecited by 213
  23. 2023Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion modelscited by 205
  24. 2026Generative Modeling via Driftingcited by 116
  25. 2025Inductive Moment Matchingcited by 113
  26. 2025Improved Mean Flows: On the Challenges of Fastforward Generative Modelscited by 110
  27. 2025DDT: Decoupled Diffusion Transformercited by 109
  28. 2018Generative Modeling using the Sliced Wasserstein Distancecited by 108
  29. 2025Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Thinkcited by 102
  30. 2025PixNerd: Pixel Neural Field Diffusioncited by 82
  31. 2024FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matchingcited by 80
  32. 2024One-Step Diffusion Distillation through Score Implicit Matchingcited by 71
  33. 2026One-step Latent-free Image Generation with Pixel Mean Flowscited by 53
  34. 2017McGan: Mean and Covariance Feature Matching GANcited by 21
  35. 2019Learning Implicit Generative Models by Matching Perceptual Featurescited by 18
  36. 2026Autoregressive Image Generation with Masked Bit Modelingcited by 9
  37. 2023Revisiting the Evaluation of Image Synthesis with GANscited by 7
  38. 2020Image Generation Via Minimizing Fr'echet Distance in Discriminator Feature Spacecited by 6

50 most cited of 51 All 51 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 37.

37 papers.

Sort:

By area:

  1. 2026Improved Baselines with Representation Autoencoderscited by 31
  2. 2026MIND: Monge Inception Distance for Generative Models Evaluationcited by 10
  3. 2026Diffusing in the Right Space: A Systematic Study of Latent Diffusabilitycited by 6
  4. 2026DiffusionBench: On Holistic Evaluation of Diffusion Transformerscited by 5
  5. 2026Three-Body Scattering for Generative Modelingcited by 4
  6. 2026Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generationcited by 4
  7. 2026AdvFD: Boosting Visual Generation via Adversarial Fr'echet Distance Losscited by 4
  8. 2026Normalizing Trajectory Modelscited by 2
  9. 2026Trajectory Forcing: Structure-First Generation with Controllable Semantic Trajectoriescited by 2
  10. 2026Amortized Moment Matching for Visual Generationcited by 2
  11. 2026GenFirst: Generation Before Reconstruction for Stable End-to-End Latent Generative Modelingcited by 2
  12. 2026MIMFlow: Integrating Masked Image Modeling with Normalizing Flows for End-to-End Image Generationcited by 1
25 more
  1. 2026Improving Medical Image Generative Models with Fr'echet Distance Losscited by 1
  2. 2026Beyond Token-Level Cross-Entropy: Fr'echet Distributional Post-Training for Autoregressive Image Generationcited by 1
  3. 2026Unifying Distributional Training for One-Step Visual Generationcited by 1
  4. 2026Improved Distributional Diffusion Modelscited by 1
  5. 2026Diffusion Image Generation with Explicit Modeling of Data Manifold Geometrycited by 0
  6. 2026Concept Guidance: Precise, Training-Free Latent Control for Text-to-Image Generationcited by 0
  7. 2026A Latent Distribution Perspective on Evaluating and Improving Latent Generative Modelscited by 0
  8. 2026DreamStream: Towards Policy-Oriented Generative Simulation for End-to-End Drivingcited by 0
  9. 2026ViRDM: Taming Representation Distribution Matching for Few-Step Causal Video Generationcited by 0
  10. 2026An End-to-End Latent-Rollout Approach for Pushing Few-Step ImageNet-256 Generation to FID 1.11 without Fr'echet Lossescited by 0
  11. 2026RepFlow: Reciprocal Supervision Improves Generation and Representation in Flow Modelscited by 0
  12. 2026Coherence-Aware Distributional Evaluation of Open-Ended Text Generationcited by 0
  13. 2026What Visual Generators Need from Teachers: Rethinking Representation Alignmentcited by 0
  14. 2026From Scores to Samples: Elastic Forcing for Autoregressive Video Generationcited by 0
  15. 2026Improving Generative Model Self-Training with Geometrically Modified Outputscited by 0
  16. 2026HiRAE: Hierarchical Representation Autoencoding with Residual Budgetscited by 0
  17. 2026DMA2: Pixel-space Distribution Matching with Adversarial and Anchor Lossescited by 0
  18. 2026Visualizing Distribution Coverage in Generative Diffusion Modelscited by 0
  19. 2026Rethinking Generative Image Compression at Extremely Low Bitratescited by 0
  20. 2026Sphere Encoder 2cited by 0
  21. 2026Depth as Time in One-Step Generative Modelscited by 0
  22. 2026Representation-Space MMD for Diffusion Language Modelscited by 0
  23. 2026Image Tokenizers Should Be Trained Between Imagescited by 0
  24. 2026Efficient Fr'echet Distributional Decoder Alignment for Latent Generative Modelscited by 0
  25. 2026Are We Overthinking One-Step Generation?cited by 0