We simplify diffusion models by defining a design space and improving FID scores.
problem Complexity in diffusion-based generative models.
method Defined a design space, improved sampling and training processes, and pre-conditioned score networks.
result Improved FID scores of 1.79 for CIFAR-10 and 1.97 for unconditional settings.
SciRE-Solver accelerates DMs sampling by recursively calculating the score function derivative.
problem Slow iterative process of diffusion models due to estimating the score function derivative.
method Recursive Difference (RD) method combined with truncated Taylor expansion of score-integrand.
result SciRE-Solver achieves state-of-the-art FIDs with significantly fewer score function evaluations.
Efficiently distills pretrained text-to-image models without real data, improving FID and CLIP scores.
problem Slow iterative refinement process of diffusion-based text-to-image models.
method Guided Score identity Distillation with Long and Short Classifier-Free Guidance.
result Achieves state-of-the-art FID performance with competitive CLIP score.
Recent work has increased the performance of Generative Adversarial Networks (GANs) by enforcing a consistency cost on the discriminator. We improve on this technique in several ways. We first show that consistency regularization can introduce artifacts into the GAN samples and explain how to fix this issue. We then pr…
SiD distills pretrained diffusion models into a fast one-step generator.
problem Efficiently distilling pretrained diffusion models into a fast generator.
method Reformulates forward diffusion processes as semi-implicit distributions and uses three score-related identities to create a loss mechanism.
result Achieves high FID performance and significantly reduces generation time.
Improved image compression with diffusion models outperforming state-of-the-art methods.
problem Difficulties in replicating text-to-image success in image compression.
method Two-stage approach combining autoencoder targeting MSE followed by score-based decoder.
result Significantly improved perceptual quality at a given bit-rate, measured by FID score.
High-quality image synthesis with diffusion models, achieving state-of-the-art FID score.
problem Generating high-quality images from latent variables.
method Training diffusion probabilistic models with a weighted variational bound, inspired by denoising score matching and Langevin dynamics.
result State-of-the-art FID score of 3.17 on CIFAR10 dataset.
Improved continuous-time consistency models for large-scale image generation.
problem Training instability and discretization errors in existing diffusion models.
method Unified theoretical framework, improved diffusion process, and network architecture.
result Trained continuous-time CMs at 1.5B parameters, achieving state-of-the-art FID scores.
CTM improves diffusion model sampling quality with efficient ODE traversal.
problem Lack of natural trade-off between sample quality and speed in consistency models.
method CTM trains a neural network to output scores and traverse ODE trajectories efficiently.
result CTM achieves state-of-the-art FIDs and improves sample quality with increased computational budget.
Recent work (Pennington et al, 2017) suggests that controlling the entire distribution of Jacobian singular values is an important design consideration in deep learning. Motivated by this, we study the distribution of singular values of the Jacobian of the generator in Generative Adversarial Networks (GANs). We find th…
Improves image quality in generative models by estimating pixel-wise aleatoric uncertainty.
problem Lack of quantitative assessment of image quality in diffusion models.
method Estimate pixel-wise aleatoric uncertainty during sampling phase using a perturbation scheme designed for diffusion models.
result Uncertainty-guided sampling leads to better sample generation quality as shown by FID scores.
Enhances VAEs for sharper image synthesis.
problem Blurriness in generated images from VAEs.
method Integrates a downscaled version of the original image into the VAE framework and uses it as input to the decoder.
result Improves FID score in image synthesis while maintaining similar log-likelihood performance.
This paper optimizes diffusion schedules for better sampling from data distributions.
problem Choosing an optimal discretization schedule for denoising diffusion models.
method Adaptive algorithm that selects an optimal schedule based on a work cost measure.
result The learned schedule recovers and outperforms manually tuned schedules.
The paper interprets diffusion models as gradient descent and proposes a new sampler.
problem Improving the efficiency and quality of diffusion models.
method Interprets diffusion models as gradient descent and proposes a new sampler.
result The new sampler achieves state-of-the-art FID scores and generates high quality samples.
Bayesian Power Steering fine-tunes large diffusion models for domain adaptation.
problem Fine-tuning pre-trained diffusion models for tasks in a smaller probability space.
method Bayesian framework with a novel network structure (Bayesian Power Steering).
result Bayesian Power Steering achieves an FID score of 10.49 on the COCO17 dataset.
We introduce a new local sparse attention layer that preserves two-dimensional geometry and locality. We show that by just replacing the dense attention layer of SAGAN with our construction, we obtain very significant FID, Inception score and pure visual improvements. FID score is improved from 18.65 to 15.94 on Im…
LSGM trains SGMs in latent space for faster sampling.
problem Efficiently generating high-quality samples from complex distributions.
method LSGM trains SGMs in latent space using a variational autoencoder framework, introducing new score-matching objectives and parameterizations.
result LSGM achieves state-of-the-art FID score of 2.10 on CIFAR-10 and outperforms previous SGMs in sampling time.
Improved image generation quality using closed-form discriminator guidance in diffusion models.
problem Enhancing the quality of images generated by diffusion models.
method Theoretical framework to analyze GAN discriminator's effect on Langevin sampling, proposing IPM-GAN optimization as smoothed score-matching.
result Closed-form kernel-based discriminator guidance improves metrics like CLIP-FID and KID.
Generative Adversarial Networks (GANs) are known to be difficult to train, despite considerable research effort. Several regularization techniques for stabilizing training have been proposed, but they introduce non-trivial computational overheads and interact poorly with existing techniques like spectral normalization.…
Diffusion models outperform GANs in image synthesis quality.
problem Improving image synthesis quality using generative models.
method Developed and optimized diffusion models, and introduced classifier guidance for conditional synthesis.
result Achieved superior image synthesis quality with diffusion models, matching or exceeding GANs in FID scores.
Improved sampling quality with RBM-Flow and D-Flow models.
problem Efficient sampling of complex data distributions using invertible flows.
method Implement RBM-Flow and D-Flow models with discrete latent variables.
result Significant improvement in sampling quality over baseline models.
Diffusion models optimize objectives similar to ELBO with Gaussian noise augmentation.
problem Optimizing diffusion models for high perceptual quality.
method Showed diffusion objectives are weighted ELBOs over noise levels, with Gaussian noise augmentation.
result Diffusion objectives equate to ELBO with Gaussian noise augmentation under monotonic weighting.
W-Flow generates images in one step, faster and better than multi-step methods.
problem Efficiently generating images from a simple reference distribution to a target data distribution.
method W-Flow uses Wasserstein gradient flows to transform the reference distribution to the target distribution in a single step, trained with Sinkhorn divergence.
result W-Flow achieves state-of-the-art results in ImageNet 256imes256 generation with improved mode coverage and faster sampling. FastFID efficiently trains generative models with FID loss.
problem Efficiently training generative models with FID loss.
method Introduces FastFID to train generative models with FID as a loss function.
result Improves FID for GANs using FID as an additional loss.
Training generative adversarial networks requires balancing of delicate adversarial dynamics. Even with careful tuning, training may diverge or end up in a bad equilibrium with dropped modes. In this work, we improve CS-GAN with natural gradient-based latent optimisation and show that it improves adversarial dynamics b…
Recent advances in generative modeling have led to an increased interest in the study of statistical divergences as means of model comparison. Commonly used evaluation methods, such as the Frechet Inception Distance (FID), correlate well with the perceived quality of samples and are sensitive to mode dropping. However,…
Generative Fractional Diffusion Models improve image diversity and quality.
problem Diffusion models struggle with diversity, mode-collapse, and slow convergence.
method Replaces light-tailed BM with fractional Brownian motion (fBM) and its Markov approximation (MA-fBM).
result GFDM achieves greater diversity and quality in image generation.
AdvAs improves GAN training by penalizing the generator based on discriminator gradients.
problem Mismatch between theoretical analysis and practice in GAN training.
method AdvAs is a penalty imposed on the generator based on the discriminator's gradient norm.
result AdvAs reduces the mismatch between theory and practice and improves GAN training.
Despite recent progress in generative image modeling, successfully generating high-resolution, diverse samples from complex datasets such as ImageNet remains an elusive goal. To this end, we train Generative Adversarial Networks at the largest scale yet attempted, and study the instabilities specific to such scale. We …
Deep generative models (DGMs) of images are now sufficiently mature that they produce nearly photorealistic samples and obtain scores similar to the data distribution on heuristics such as Frechet Inception Distance (FID). These results, especially on large-scale datasets such as ImageNet, suggest that DGMs are learnin…
Unified and extended precision-recall metrics for generative models.
problem Evaluation of generative models using scalar metrics like FID and IS.
method Unified various precision-recall approaches under a common framework.
result Recovered entire precision-recall curves and exposed metric pitfalls.
ECS evaluates synthetic CXR images' distributional fidelity.
problem Evaluating synthetic CXR images' distributional fidelity under privacy constraints.
method Characteristic function transforms of feature embeddings.
result ECS uncovers clinically relevant distributional discrepancies.
Researchers compare two GAN models for lip-synchronization tasks.
problem Improving the quality of generated lip-sync videos.
method Reimplemented and trained two GAN models (LipGAN and L1WGAN-GP) on the GRID dataset.
result L1WGAN-GP outperformed LipGAN in generating more realistic lip-sync videos.
Paper bridges VAEs and KDEs for more flexible posterior estimation.
problem Limitations of Gaussian latent space in VAEs and challenges in KL-divergence estimation.
method Approximate posterior with KDEs and derive upper bound of KL-divergence in ELBO.
result Epanechnikov kernel minimizes KL-divergence upper bound asymptotically.
FID misaligns with human judgment due to reliance on ImageNet classes.
problem FID metric's reliance on ImageNet classes causes discrepancies with human evaluation.
method Investigated and visualized the feature space of FID and its relation to ImageNet classes.
result Aligning histograms of Top-N ImageNet classifications can reduce FID without improving quality. uTSGAN improves on TSGAN for generating time series data.
problem Challenges in generating time-dependent data.
method Unified training of independent networks in TSGAN.
result uTSGAN outperforms TSGAN in 80% of benchmark datasets.
Variational Autoencoders (VAE) are probabilistic deep generative models underpinned by elegant theory, stable training processes, and meaningful manifold representations. However, they produce blurry images due to a lack of explicit emphasis over high-frequency textural details of the images, and the difficulty to dire…
C-t3VAE improves class representation in long-tailed generative models.
problem Latent geometric bias in VAEs under class imbalance.
method Per-class Student's t-distribution priors, closed-form objective, equal-weight latent mixture.
result Consistently lower FID scores and better class-balanced generation for severely imbalanced datasets.
Generative Adversarial Networks (GANs) have shown remarkable results in modeling complex distributions, but their evaluation remains an unsettled issue. Evaluations are essential for: (i) relative assessment of different models and (ii) monitoring the progress of a single model throughout training. The latter cannot be…
This study examines ensembling of diffusion models for improved generative quality.
problem Improving generative quality with ensembling of score-based diffusion models.
method Investigated ensembling of scores from multiple diffusion models on image and tabular data.
result Ensembling scores generally improves model likelihood and score-matching loss but not perceptual quality metrics.
Improved image quality in diffusion models by limiting guidance to a specific noise level range.
problem Improving image quality in diffusion models with guidance.
method Restricting guidance to a specific noise level range.
result Significantly improved FID score from 1.81 to 1.40 in ImageNet-512.
Maximum likelihood estimation of energy-based models is a challenging problem due to the intractability of the log-likelihood gradient. In this work, we propose learning both the energy function and an amortized approximate sampling mechanism using a neural generator network, which provides an efficient approximation o…
Hyperbolic GANs improve image generation metrics.
problem Improving image generation quality in neural networks.
method Integrating hyperbolic layers into GAN architectures.
result Hyperbolic GANs achieve better metrics than Euclidean counterparts.
We study the impact of input noise dimension on GAN performance.
problem Challenges in GANs like mode collapse and unstable training.
method Quantitative analysis using FID and IS on different image datasets.
result The optimal input noise dimension varies by dataset and architecture.
This study analyzes bias in face generation models and proposes mitigation techniques.
problem Bias in face generation models that may misrepresent certain social groups.
method Three approaches: generators, attribute modifier, and post-processing bias mitigators.
result Generators suffer from bias across all social groups, while post-processing mitigators show different feature concentration.
A new GAN framework teaches models to avoid undesirable data.
problem Improving GANs' ability to learn from and avoid negative samples.
method Introducing 'negative samples' to GANs, inspired by Rumi's philosophy.
result The Rumi Framework accelerates GAN learning and improves generalization.
TVM improves generative modeling by matching terminal velocities.
problem Creating high-fidelity one- and few-step generative models.
method TVM generalizes flow matching, modeling transitions between diffusion timesteps and regularizing terminal behavior.
result TVM achieves state-of-the-art FID scores with minimal architectural changes and fused attention kernel.
Within a broad class of generative adversarial networks, we show that discriminator optimization process increases a lower bound of the dual cost function for the Wasserstein distance between the target distribution p and the generator distribution pG. It implies that the trained discriminator can approximate opti…