The current path of generative AI relies heavily on Latent Diffusion Models (LDMs) To manage the computational cost of high-resolution synthesis. By compressing data into a lower-dimensional latent space, models can scale effectively. However, a fundamental trade-off remains: lower information density makes learning the latent information easier but sacrifices the quality of the reconstruction, while higher density enables near-perfect reconstruction but requires greater modeling ability.
Google researchers introduced DeepMind Uniform Latency (UL)a framework designed to navigate this trade-off systematically. The framework jointly organizes latent representations with prior diffusion and decodes them via a diffusion model.
Architecture: Three pillars of unified latency
the Uniform latency (UL) framework is based on three specific technical components:
- Invariant Gaussian noise coding: Unlike standard variable autoencoders (VAEs) that learn the distribution of encryption, UL uses deterministic encryption E which predicts z as a single latentCleans. This latent is then perturbed to a final signal-to-noise ratio (log-SNR) of 5(0)=5.
- Pre-alignment: The previous propagation model agrees with this minimum noise level. This alignment of the Kullback-Leibler (KL) term in the lower bound of evidence (ELBO) allows minimization of the simple mean square error (MSE) over noise levels.
- ELBO reweighted decoder: The decoder uses a sigmoid weight loss, which provides an interpretable bound on the latent bit rate while allowing the model to prioritize different noise levels.
The training process has two stages
The UL framework is implemented in two distinct phases to improve the quality of latent learning and generation.
The first stage: shared latent learning
In the first stage, encryption and pre-propagation (P), and diffusion decoding (D) They are trained jointly. The goal is to find out the underlying elements that are simultaneously encoded, organized, and modeled. The encoder output noise is directly related to the previous minimum noise level, providing a tight upper bound on the latent bit rate.
Stage 2: Scaling the basic model
The research team found that a previous trainer who was only trained on ELBO loss in the first stage does not produce perfect samples because it weights low-frequency content and high-frequency content equally. Thus, in the second stage, the encoder and decoder are frozen. The new “base model” is then trained on the latent items using sigmoid weighting, significantly improving performance. This stage allows for larger model sizes and batch sizes.
Technical performance and SOTA standards
Uniform latent powers show high efficiency in the relationship between training computability (FLOPs) and generation quality.
| metric | Data set | a result | indication |
| FID | ImageNet-512 | 1.4 | It outperforms models trained on the latent components of stable propagation for a given computational budget. |
| Cardiovascular disease | Kinetics-600 | 1.3 | He appoints a new one State of the art technology (SOTA) To generate video. |
| PSNR | ImageNet-512 | Up to 30.1 | Maintains high reconstruction accuracy even at higher pressure levels. |
In ImageNet-512, UL outperforms previous methods, including DiT and EDM2 variants, in terms of training cost versus FID generation. In video tasks using the Kinetics-600, the small UL model achieved 1.7 FVD, while the medium variant reached a SOTA 1.3 FVD.
Key takeaways
- Integrated Deployment Framework: UL is a framework that jointly optimizes the pre-diffusion encoder and the diffusion decoder, ensuring that the latent representations are encoded, regularized and modeled simultaneously to generate high efficiency.
- Static binding noise information: By using a deterministic encoder that adds a fixed amount of Gaussian noise (specifically at a log SNR of (0)=5) and relating it to the previous minimum noise level, the model provides a tight and interpretable upper bound on the latent bit rate.
- Two-stage training strategy: The process involves an initial joint training phase for the autoencoder and its pre-training, followed by a second phase where the encoder and decoder are frozen and a larger “base model” is trained on the latent elements to maximize sample quality.
- State-of-the-art performance: The framework generated a new state-of-the-art Fréchet Video Distance (FVD) (SOTA) of 1.3 on Kinetics-600 and achieved a competitive Fréchet Inception Distance (FID) of 1.4 on ImageNet-512 while requiring fewer training FLOPs than standard latent diffusion baselines.
verify paper. Also, feel free to follow us on twitter Don’t forget to join us 120k+ ml SubReddit And subscribe to Our newsletter. I am waiting! Are you on telegram? Now you can join us on Telegram too.
