Meta's Muse Image Beats Google but Trails OpenAI in AI Image Generation

Muse Image is Meta's first in-house image generation model, and it is not trying to be a better Midjourney. Instead of directly mapping prompts to images, Muse Image operates as an agent: it invokes search and coding tools to improve accuracy, self-refines its own generations, and improves through scaling test-time compute. That framing matters. This is less a diffusion model with a pretty UI and more a reasoning system that happens to output pixels.

Originally codenamed Mango, the model marks the second major release from Meta Superintelligence Labs, led by Alexandr Wang, who oversaw the April unveiling of Muse Spark. Meta has previously used third-party AI models like Midjourney and Black Forest Labs to power image and video generation features in its Meta AI app and site. The company now plans to use its new model to reduce reliance on those third-party technologies.

Where the benchmarks land

Muse Image holds the No. 2 spot on Arena for text-to-image, single-image editing, and multi-image editing as measured by human preference Elo rankings. Internal benchmark tests show Muse Image trailing OpenAI's GPT Image 2 model but beating Google's Nano Banana 2 in tasks like editing both single and multiple images. For video, Muse Video ranks No. 3 in human-preference Elo for text-to-video.

The agentic architecture under the hood

The core design choice here is that image generation is treated as a reasoning problem, not a lookup. The model gets two main tools during training:

• Code execution: During reinforcement learning, Muse Image learns to write and execute code that produces accurate plots and QR codes, and conditions on rendered figures to improve the accuracy of generated images.
• Web search: Muse Image learns to search the web to ground generated images in factual and real-time information and visual references. Enabling search improves factual accuracy on knowledge-intensive prompts, particularly those involving current events and real-world facts.

On top of tools, the model learned something the team did not explicitly design. Muse Image reflects on and improves upon its own work within its chain of thought. This self-refining behavior can take different forms: a local edit to the current image draft when a small detail is off, a new image generation from scratch when larger parts are wrong, or a different tactic like tool use for more factually accurate generation. Meta did not design this behavior. Instead, it emerged during RL training simply because self-refinement produced better images and therefore higher reward.