DiffusionGemma: 4x faster text-to-image generation

🔑 增強重點摘要

•DiffusionGemma achieves generation speeds exceeding 1,000 tokens per second on a single NVIDIA H100 GPU and over 700 tokens per second on an NVIDIA GeForce RTX 5090, specifically optimized for low-latency, single-user applications.
•The model is built upon the 26B Mixture-of-Experts (MoE) Gemma 4 architecture, but activates only 3.8B parameters during inference, allowing it to fit within 18GB VRAM limits of high-end consumer GPUs when quantized.
•Its speed improvement stems from a discrete text diffusion approach, which generates text by iteratively denoising blocks of 256 tokens in parallel, moving beyond the sequential token-by-token processing of traditional autoregressive models.
•DiffusionGemma is an experimental open model released under an Apache 2.0 license, designed for speed-critical, interactive local workflows such as in-line editing, code infilling, and generating non-linear text structures, with an acknowledged trade-off in overall output quality compared to standard Gemma 4.
•The model is multimodal, capable of processing interleaved text, image (with variable aspect ratio and resolution support), and video inputs to generate text outputs.

🛠️ 技術深入

Architecture Base: Built on the 26B (4B active) Mixture-of-Experts (MoE) Gemma 4 architecture.
Generation Mechanism: Utilizes discrete text diffusion, generating tokens by iteratively denoising blocks of tokens (a 'canvas') in parallel.
Input Modalities: Natively accepts text, images (with variable aspect ratio and resolution support), and video inputs, generating text output. Audio input is not supported.
Model Structure: Employs an Encoder-Decoder Architecture; an autoregressive encoder processes and caches prompt context, while a denoising decoder applies bi-directional attention over the generation canvas.
Efficiency: Leverages a sparse MoE design (8 active experts out of 128 total), activating approximately 3.8B parameters during inference for a low memory footprint.
Hardware Accessibility: When quantized, it fits within 18GB VRAM limits, making it deployable on high-end dedicated consumer GPUs like NVIDIA RTX 4090 or 5090.
Parallel Generation: Generates 256 tokens in parallel with each forward pass, allowing every token to attend to all others.
Self-Correction: Features intelligent self-correction, iteratively refining its own output by evaluating the entire text block at once to fix mistakes.
Reasoning Capabilities: Includes a configurable 'Thinking Mode' with built-in reasoning channels for step-by-step processing before generating a final answer.
Sampling Details: Uses diffusion sampling with Entropy-Bounded Denoising and Adaptive Stopping, with a maximum of 48 denoising steps and a linear temperature decay schedule.
Self-Conditioning: Employs self-conditioning to stabilize the denoising loop and accelerate convergence by feeding back the full softmax distribution from the previous step.
Hardware Optimization: Supports native NVIDIA NVFP4 (4-bit floating-point) format on Blackwell GPUs for accelerated compute throughput.
Inference Speed: Achieves over 1,000 tokens per second on a single NVIDIA H100 (FP8) and 1,288 tokens per second on H200, optimized for small batch size inference.
Adaptive Inference: Simpler prompts and structured tasks (e.g., code) require fewer denoising steps, enabling dynamic tokens-per-second speeds based on task complexity.

🔮 前景展望AI analysis grounded in cited sources

DiffusionGemma will accelerate the development of real-time, interactive AI applications on local hardware.

Its focus on high-speed, low-latency text generation and accessible hardware footprint (18GB VRAM) directly addresses bottlenecks for local inference and interactive workflows like in-line editing and code infilling.

The discrete text diffusion paradigm will become a standard for future on-device generative AI.

By enabling parallel block generation and efficient utilization of GPU compute, this approach offers a blueprint for high-tier AI execution offline, especially for speed-critical tasks.

There will be a growing emphasis on optimizing generative models for specific use cases, even if it means trading off some general quality.

DeepMind explicitly states DiffusionGemma prioritizes speed for interactive workflows, recommending standard Gemma 4 for maximum quality, indicating a strategic segmentation of models based on application needs.

⏳ 時間線

2022-05

Google Brain (later DeepMind) introduces Imagen, a series of text-to-image models.

2023-04

Google Brain merges with DeepMind.

2024-08

Imagen 3 is released, offering improved detail and lighting in generated images.

2025-05-20

Imagen 4, Google DeepMind's latest text-to-image model, is released at Google I/O 2025.

2026-04-03

Google DeepMind releases the Gemma 4 family of open-weight models.

2026-06-10

DeepMind introduces DiffusionGemma, an experimental open model for exceptionally fast text generation.

DiffusionGemma：文字生成圖像速度提升 4 倍

⚡ 30-Second TL;DR

關鍵要點

🧠 深度解析

🔑 增強重點摘要

🛠️ 技術深入

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線

📎 來源 (11)

👉相關動態

阿里發布 Qwen-Image-3.0 圖像生成模型