The current trajectory of generative AI is highly dependent on: Latent Diffusion Model (LDM) Manage the computational cost of high-resolution compositing. Compressing the data into a low-dimensional latent space can effectively extend the model. However, fundamental trade-offs still exist. In other words, lower information density facilitates latent learning, but at the expense of reconstruction quality. On the other hand, high information density allows near-perfect reconstruction but requires greater modeling power.
Google DeepMind introduced by researchers Unified Potential (UL)a framework designed to systematically navigate this tradeoff. This framework jointly normalizes latent representations with a spreading prior and decodes them via a spreading model.
Architecture: Three Pillars of Integrated Potential
of integrated potential (The UL) framework is based on three specific technical components:
- Fixed Gaussian noise encoding: Unlike the standard variational autoencoder (VAE), which learns the encoder distribution, UL uses a deterministic encoder E.𝝷 predict a single potential zclean. This latent is then forward noise processed until the final log signal-to-noise ratio (log-SNR) is λ(0)=5.
- Pre-adjustment: Previous diffusion models are tuned to this minimum noise level. This adjustment reduces the Kullback-Leibler (KL) term in the evidence lower bound (ELBO) to a simple weighted mean squared error (MSE) relative to the noise level.
- Reweighted decoder ELBO: The decoder utilizes a sigmoid weighted loss to provide an interpretable bound on the potential bitrate while allowing the model to prioritize different noise levels.
Two-step training process
The UL framework is implemented in two different stages to optimize both latent learning and generation quality.
Stage 1: Joint latent learning
In the first stage encoder, the pre-spreading (P𝝷), spreading decoder (D𝝷) are jointly trained. The goal is to learn a latent that is simultaneously encoded, regularized, and modeled. The encoder’s output noise is directly linked to the previous minimum noise level, placing a hard upper limit on the potential bitrate.
Stage 2: Scaling the base model
The researchers found that a pretrain trained solely on the stage 1 ELBO loss did not produce optimal samples because it weighted low and high frequency content equally. As a result, in stage 2, the encoder and decoder freeze. A new “base model” is then trained on the latent using sigmoid weighting, which significantly improves performance. This stage allows for larger model and batch sizes.
Technical performance and SOTA benchmarks
Unified Latents demonstrates high efficiency in the relationship between training computation (FLOP) and production quality..
| metric | dataset | result | significance |
| F.I.D. | Image Net-512 | 1.4 | It performs better than models trained with stable diffusion potentials for a given computing budget. |
| FVD | Kinetics-600 | 1.3 | set a new one Cutting Edge (SOTA) For video generation. |
| PSNR | Image Net-512 | Until 30.1 | Maintains high reconstruction fidelity even at higher compression levels. |
On ImageNet-512, UL outperformed previous approaches including DiT and EDM2 variants in terms of training cost for generated FIDs. In a video task using Kinetics-600, the small UL model achieved 1.7 FVD, while the medium model reached SOTA 1.3 FVD.
Important points
- Integrated dissemination framework: UL is a framework that jointly optimizes the encoder, spreading prior, and spreading decoder so that latent representations are simultaneously encoded, normalized, and modeled for highly efficient generation.
- Constraining fixed noise information: By using a deterministic encoder that adds a fixed amount of Gaussian noise (specifically a logarithmic SNR of λ(0)=5) and linking it to a prior minimum noise level, the model provides a hard and interpretable upper bound on the potential bitrate.
- Two-step training strategy: This process includes an initial co-training phase with the autoencoder and earlier, followed by a second phase where the encoder and decoder are frozen and a larger “base model” is trained on the latent to maximize sample quality.
- Cutting-edge performance: The framework establishes a new state-of-the-art (SOTA) Fréchet Video Distance (FVD) of 1.3 on Kinetics-600 and achieves a competitive Fréchet Inception Distance (FID) of 1.4 on ImageNet-512 while requiring fewer training FLOPs than the standard latent diffusion baseline.
Please check paper. Please feel free to follow us too Twitter Don’t forget to join us 120,000+ ML subreddits and subscribe our newsletter. hang on! Are you on telegram? You can now also participate by telegram.
