Skip to main content
Back

Google Improves Diffusion Models with Noise Optimization

LEARN MORE

January 21, 2025 3 min read

Hey,

Welcome to AlphaSignal – the most read newsletter by AI developers. 


We bring you the top 1% of news, papers, models, and repos, all summarized to keep you updated on the latest in AI.

IN TODAY'S SIGNAL

Read time: 3 min 58 sec

🎖️ Top News

📌 Crossing Minds

⚡️ Trending Signals

💻 Top Repos

  • DeepSeek-R1: Excels in reasoning, math, and code with scalable RL-based models.

  • Monolith: Build recommendation systems with real-time training and TensorFlow integration.

  • build-your-own-x: Offers tutorials for building technologies to deepen practical understanding.

🧠 Top Lecture

If you're enjoying AlphaSignal please forward this email to a colleague. 

It helps us keep this content free.

TOP NEWS

Generative Model Framework

Google DeepMind proposes a new method to improve diffusion model outputs without extra training

⇧ 1,835 Likes

What's New

Researchers from NYU, MIT, and Google introduce a framework that improves inference-time scaling for diffusion models. Diffusion models generate images, audio, and videos by denoising noisy data. Traditional methods fail to improve performance when you add more denoising steps during inference.

Key Features

  • Search Operations: The system explores different noise candidates to find the best starting point for generating high-quality samples.

  • Fixed Denoising Steps: The number of denoising steps is kept constant at 250 to avoid diminishing returns.

  • Verifiers: The framework uses two verifiers—Inception Score (IS) and Fréchet Inception Distance (FID)—to assess the quality of generated images during the search process.

Noise Optimization
Diffusion models use random noise as a starting point for generating outputs. Some random noises result in better outputs than others. The new approach systematically searches for the best initial noise to improve results during inference.

Denoising Limitations
In a typical diffusion process, noise is progressively cleaned up in steps. Increasing the number of denoising steps (or NFEs) improves output quality, but only up to a point. Eventually, adding more steps doesn’t make much difference.

Search-Based Approach
Instead of just increasing the number of denoising steps, this framework treats the process as a search for the optimal noise. It uses a search-based methodology to find better starting noise, improving the generation quality without retraining the model.

Test Results

  • Random Search: A Random Search algorithm selects the best noise candidates.

  • Improved Sample Quality: The framework consistently produced higher-quality results across multiple benchmarks.

  • Performance Metrics: ImageReward and Verifier Ensemble performed best in quality across all tests, while ImageReward excelled in text-prompt accuracy.

Learn more

Adapt LLMs to Your Domain Efficiently and Iterate Faster

RAGSys is a new supervised retrieval engine designed for LLM domain adaptation.

  • Train models 300x faster than traditional fine-tuning.

  • Improve accuracy by 76% compared to static few-shot examples.

  • Skip prompt engineering with reusable retrieval pipelines.

  • Use your data with any LLM: GPT, Claude, Llama, and others.

Built with cross-encoder retrieval models and RLAIF algorithms, RAGSys reduces compute costs and shortens development cycles.

Start using it now.

Learn more

partner with us →

TRENDING SIGNALS

Generative AI

Tencent unveils Hunyuan3D 2.0, an all in one platform to generate high resolution geometry, texture generation and animation creation

⇧ 1,387 Likes

Software Update

Git introduces Meson build support, simplifying setup with faster, cross-platform compatibility in the new update

⇧ 1,264 Likes

Developer Tool

LangChain updates LangSmith playground, improves prompt workflow and adds side-by-side comparisons

⇧ 363 Likes

Coding Assistant

ByteDance presents coding assistant for macOS with contextual understanding for automated project building 

⇧ 320 Likes

AI Model

Moonshot AI introduces Kimi k1.5, an open source multimodal model achieving SOTA on short-CoT with text-vision reasoning

⇧ 1,654 Likes

TRENDING REPOS

Reasoning Model

DeepSeek-R1

☆ 6,010 Stars

DeepSeek-R1 helps you solve complex reasoning, math, and code tasks using advanced reinforcement learning (RL) techniques. It outperforms OpenAI-o1-mini in benchmarks like MATH-500 (97.3% pass@1) and Codeforces (2029 rating). The models include DeepSeek-R1-Zero and DeepSeek-R1, with 671B parameters and 37B activated. Distilled versions (1.5B–70B) outperform many dense models like Claude 3.5 and OpenAI-o1, offering scalability and open-source access.

Recommendation System

Monolith

☆ 8,014 Stars

Monolith is a deep learning framework designed for large-scale recommendation systems. It features collisionless embedding tables for unique ID representation and real-time training to capture recent trends and user interests. Built on TensorFlow, Monolith supports batch and real-time training and serving. 

Tutorials

build-your-own-x

☆ 326,812 Stars

The repository contains a curated collection of step-by-step guides for building various technologies from scratch, including 3D renderers, neural networks, operating systems, web browsers, programming languages, and more. Each tutorial focuses on replicating a technology to deepen understanding. Topics span multiple languages, frameworks, and tools, covering real-world implementations and concepts.

TOP LECTURE

RAG

Build Long-Context AI Apps with Jamba

⇧ 2,080 Likes

This course explains how to use the Jamba model, a hybrid transformer-Mamba architecture, to handle long contexts efficiently. Developed by AI21 Labs, this model combines Mamba's computational efficiency with the transformer’s attention mechanism to improve output quality.

Topics include:

  • Jamba’s hybrid architecture combining transformers and state-space models (SSMs).

  • Training methods and evaluation metrics for long-context models.

  • Using the AI21 SDK to process documents over 200k tokens, such as annual financial reports.

  • Tool-calling examples, from arithmetic operations to retrieving SEC 10-Q reports.

  • Building a long-context RAG application using AI21’s Conversational RAG tool and integrating Jamba with LangChain.
Learn more

At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.


Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.

Work with us