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.

This emergent self-correction is a meaningful signal. It mirrors what happened with chain-of-thought reasoning in language models: the model discovered that checking its own work was worth the extra compute. The same dynamic is now showing up in visual generation.

Test-time compute scaling, now for images

One of the more technically interesting claims is that Muse Image exhibits log-linear scaling with test-time compute, the same relationship seen in reasoning-heavy language models. More thinking at inference time means better images. Crucially, the compute spans two very different modalities: text tokens for reasoning and visual tokens for generation, yet quality scales with the combined total.

Meta also found that how you spend that compute budget matters. Best-of-N sampling (generating several images and keeping the best) improves quality early but saturates quickly. Deliberate reasoning scales considerably better, and tool use compounds on top of that.

Social context as a training signal

Muse Image follows instructions faithfully, edits with precision, composes from multiple references, and draws on Instagram for social context. It also integrates with Muse Spark, allowing the two models to share tools and plan jointly for powerful agentic media generation. The Instagram angle is not just a product feature. Meta is the only lab that owns a platform with billions of images annotated by human engagement, giving it a unique signal for what visuals actually resonate.

With Muse Image-powered Meta AI, users can tag public Instagram accounts and incorporate photos from those profiles into generated images. Instagram users can opt out of being tagged or having their images used, but the feature is enabled by default. That default-on opt-out will likely attract scrutiny.

What it is good at, and where the gaps are

The model's strengths cluster around tasks that benefit from reasoning before rendering:

• Precise image editing that changes only what you ask
• Multi-reference composition (combining people, objects, styles from several source images)
• Text rendering inside images: infographics, QR codes, how-to guides, flyers
• Factually grounded images tied to current events via search
• Animated GIFs, websites with embedded images, and interactive visual games, through the Muse Spark and Muse Image integration.

The acknowledged gaps are on the video side. Meta is investing in areas with current performance gaps in Muse Video, such as audio-video synchronization and physically accurate fast motion. On the image side, the model still trails GPT Image 2 in overall quality, and no independent evals exist yet beyond Arena Elo.

Practical use cases

The clearest immediate applications break into three buckets:

1. Creators and consumers: More than 30 new AI-powered effects for Instagram Stories and image generation in direct chats with Meta AI on WhatsApp.
2. Small businesses: Muse Image will power advertiser-specific image generation tools as part of Meta's AI-powered Advantage Plus service, letting brands more easily develop ad creative for their marketing campaigns.
3. Agentic pipelines: Muse Spark and Muse Image integrate to use the combination of code and media generation to create animated GIFs, websites with embedded images, and interactive visual games.

Availability and pricing

Muse Image is available today across the Meta AI app and on meta.ai, Instagram Stories in the US, and WhatsApp in limited countries, and is coming soon to Facebook. Using Meta AI with Muse Image is free for everyday creation. For people who want to create even more, it is available as part of Meta's subscription plans.

Meta is still evaluating whether it will make Muse Image available to outside developers. Developers still do not have access to Muse Spark either, though the company says API access to Spark is coming soon. For now, both models remain consumer-facing products, not developer infrastructure.

The bigger picture: a closed-source media stack

Muse Image is the second piece of a vertically integrated AI stack that Meta is quietly assembling. Muse Spark handles reasoning and language. Muse Image handles visuals. Muse Video is previewed and coming. Muse Image extends this architecture into visual generation, giving Meta a unified AI stack capable of handling text, images, and eventually video.

The industry context here is significant. The product signal is scale: image generation is moving from standalone tools into social editing, messaging, and ad-creative workflows where latency, policy enforcement, and provenance matter. Midjourney, Adobe Firefly, and Stability AI all built standalone tools. Meta is embedding generation directly into surfaces where 3 billion people already spend time.

To help with provenance, Muse Image includes Content Seal, an invisible watermarking system. Images created by Muse Image carry a hidden provenance signal that stays intact even when cropped, compressed, resized, or screenshotted. Meta is also previewing a detection tool to check whether an image carries a Content Seal watermark.

The open-source question looms over all of this. The Llama series built Meta a massive developer ecosystem precisely because the weights were free to download and run. Meta may now have a more credible answer to rival frontier labs on capability, but it is no longer offering the same immediate developer access that helped define its public AI identity. Whether that trade-off pays off depends on how fast the Muse API opens up, and whether the models are good enough to pull developers away from the pipelines they have already built around OpenAI and Stability.