Judge Decoding: Faster Speculative Sampling Requires Going Beyond Model Alignment

The performance of large language models (LLMs) is closely linked to their underlying size, leading to ever-growing networks and hence slower inference. Speculative decoding has been proposed as a technique to accelerate autoregressive generation, leveraging a fast draft model to propose candidate tokens, which are then verified in parallel based on their likelihood under the target model. While this approach guarantees to reproduce the target output, it incurs a substantial penalty: many high-quality draft tokens are rejected, even when they represent objectively valid continuations. Indeed, we show that even powerful draft models such as GPT-4o, as well as human text cannot achieve high acceptance rates under the standard verification scheme. This severely limits the speedup potential of current speculative decoding methods, as an early rejection becomes overwhelmingly likely when solely relying on alignment of draft and target. We thus ask the following question: Can we adapt verification to recognize correct, but non-aligned replies? To this end, we draw inspiration from the LLM-as-a-judge framework, which demonstrated that LLMs are able to rate answers in a versatile way. We carefully design a dataset to elicit the same capability in the target model by training a compact module on top of the embeddings to produce ``judgements" of the current continuation. We showcase our strategy on the Llama-3.1 family, where our 8b/405B-Judge achieves a speedup of 9x over Llama-405B, while maintaining its quality on a large range of benchmarks. These benefits remain present even in optimized inference frameworks, where our method reaches up to 141 tokens/s for 8B/70B-Judge and 129 tokens/s for 8B/405B on 2 and 8 H100s respectively.

Imagine Flash: Accelerating Emu Diffusion Models with Backward Distillation

Diffusion models are a powerful generative framework, but come with expensive inference. Existing acceleration methods often compromise image quality or fail under complex conditioning when operating in an extremely low-step regime. In this work, we propose a novel distillation framework tailored to enable high-fidelity, diverse sample generation using just one to three steps. Our approach comprises three key components: (i) Backward Distillation, which mitigates training-inference discrepancies by calibrating the student on its own backward trajectory; (ii) Shifted Reconstruction Loss that dynamically adapts knowledge transfer based on the current time step; and (iii) Noise Correction, an inference-time technique that enhances sample quality by addressing singularities in noise prediction. Through extensive experiments, we demonstrate that our method outperforms existing competitors in quantitative metrics and human evaluations. Remarkably, it achieves performance comparable to the teacher model using only three denoising steps, enabling efficient high-quality generation.

Discovering Class-Specific GAN Controls for Semantic Image Synthesis

Prior work has extensively studied the latent space structure of GANs for unconditional image synthesis, enabling global editing of generated images by the unsupervised discovery of interpretable latent directions. However, the discovery of latent directions for conditional GANs for semantic image synthesis (SIS) has remained unexplored. In this work, we specifically focus on addressing this gap. We propose a novel optimization method for finding spatially disentangled class-specific directions in the latent space of pretrained SIS models. We show that the latent directions found by our method can effectively control the local appearance of semantic classes, e.g., changing their internal structure, texture or color independently from each other. Visual inspection and quantitative evaluation of the discovered GAN controls on various datasets demonstrate that our method discovers a diverse set of unique and semantically meaningful latent directions for class-specific edits.

You Only Need Adversarial Supervision for Semantic Image Synthesis

Despite their recent successes, GAN models for semantic image synthesis still suffer from poor image quality when trained with only adversarial supervision. Historically, additionally employing the VGG-based perceptual loss has helped to overcome this issue, significantly improving the synthesis quality, but at the same time limiting the progress of GAN models for semantic image synthesis. In this work, we propose a novel, simplified GAN model, which needs only adversarial supervision to achieve high quality results. We re-design the discriminator as a semantic segmentation network, directly using the given semantic label maps as the ground truth for training. By providing stronger supervision to the discriminator as well as to the generator through spatially- and semantically-aware discriminator feedback, we are able to synthesize images of higher fidelity with better alignment to their input label maps, making the use of the perceptual loss superfluous. Moreover, we enable high-quality multi-modal image synthesis through global and local sampling of a 3D noise tensor injected into the generator, which allows complete or partial image change. We show that images synthesized by our model are more diverse and follow the color and texture distributions of real images more closely. We achieve an average improvement of $6$ FID and $5$ mIoU points over the state of the art across different datasets using only adversarial supervision.

A U-Net Based Discriminator for Generative Adversarial Networks

Among the major remaining challenges for generative adversarial networks (GANs) is the capacity to synthesize globally and locally coherent images with object shapes and textures indistinguishable from real images. To target this issue we propose an alternative U-Net based discriminator architecture, borrowing the insights from the segmentation literature. The proposed U-Net based architecture allows to provide detailed per-pixel feedback to the generator while maintaining the global coherence of synthesized images, by providing the global image feedback as well. Empowered by the per-pixel response of the discriminator, we further propose a per-pixel consistency regularization technique based on the CutMix data augmentation, encouraging the U-Net discriminator to focus more on semantic and structural changes between real and fake images. This improves the U-Net discriminator training, further enhancing the quality of generated samples. The novel discriminator improves over the state of the art in terms of the standard distribution and image quality metrics, enabling the generator to synthesize images with varying structure, appearance and levels of detail, maintaining global and local realism. Compared to the BigGAN baseline, we achieve an average improvement of 2.7 FID points across FFHQ, CelebA, and the newly introduced COCO-Animals dataset.

Generalized Zero- and Few-Shot Learning via Aligned Variational Autoencoders

Many approaches in generalized zero-shot learning rely on cross-modal mapping between the image feature space and the class embedding space. As labeled images are rare, one direction is to augment the dataset by generating either images or image features. However, the former misses fine-grained details and the latter requires learning a mapping associated with class embeddings. In this work, we take feature generation one step further and propose a model where a shared latent space of image features and class embeddings is learned by modality-specific aligned variational autoencoders. This leaves us with the required discriminative information about the image and classes in the latent features, on which we train a softmax classifier. The key to our approach is that we align the distributions learned from images and from side-information to construct latent features that contain the essential multi-modal information associated with unseen classes. We evaluate our learned latent features on several benchmark datasets, i.e. CUB, SUN, AWA1 and AWA2, and establish a new state-of-the-art on generalized zero-shot as well as on few-shot learning. Moreover, our results on ImageNet with various zero-shot splits show that our latent features generalize well in large-scale settings.