Meta officially launched its first self-developed AI image generation model on July 7 Muse Imagewhile also previewing an early version of its video generation model, Muse Video. This marks Meta Superintelligence Labs’ (MSL) second major AI model release since launching its large language model in April this year, signifying Meta’s official break from dependence on third-party AI models and its move toward developing its own AI technology.
Alongside the release of Muse Image, we’re sharing an early preview of Muse Video. It offers competitive performance in prompt adherence, visual fidelity, and temporal consistency.
We’re investing in areas with current performance gaps, such as audio-video synchronization and… https://t.co/kftRoTgq6W pic.twitter.com/iIFeFGLzoE
— AI at Meta (@AIatMeta) July 7, 2026
Muse Image: AI Image Generation Model with Agency Capabilities
Muse Image is developed by Meta Superintelligence Labs under the leadership of Alexandr Wang, with the internal code name “Mango.” Wang just led the release of the large language model Muse Spark this past April, and now Muse Image has become MSL’s second major release. Unlike traditional prompt-to-image models, Muse Image is designed as an AI agent: it invokes search and coding tools to improve accuracy, reflects on its own outputs and refines generation results, and can continuously enhance quality by scaling test-time compute.
According to Meta AI’s official blog technical explanation, Muse Image learned to write and execute code during reinforcement learning, enabling it to generate precise charts and QR codes. It can also search the web to ensure factual accuracy of generated content, particularly excelling at prompts involving real-time information and real-world knowledge.
The most striking feature is Muse Image’s self-improvement capability. When small errors appear in the generated results, it performs local editing and corrections. When larger portions are flawed, it regenerates from scratch. When necessary, it will even switch to using tools to enhance factual accuracy. Meta states that this behavior emerged naturally during reinforcement learning training, rather than being deliberately designed by engineers, because self-improvement produces better images, which in turn earns higher reward scores.
In terms of performance, Muse Image ranks second in Arena across all three categories—text-to-image, single-image editing, and multi-image editing—based on human preference Elo scores. According to Meta’s published internal benchmarks, Muse Image only trails OpenAI’s latest GPT Image 2 model but surpasses competitors like Nano Banana 2, and now Meta AI It can already be used immediately for free, and actual testing shows its Chinese comprehension and generation capabilities are also outstanding.
A particularly eye-catching feature of Muse Image is that users can @-mention public Instagram accounts in Meta AI to directly incorporate that user’s public photos into AI-generated images. For example, you can @-mention a friend’s public account and ask to “turn me and my friend into 16-bit game characters,” and Muse Image will automatically pull that account’s public photos to generate it. This feature has already sparked privacy concerns.
Muse Video: Early Preview, Native Audio Support
Alongside the release of Muse Image, Meta also shared an early preview of Muse Video. Built on the same pretrained foundation as Muse Image, Muse Video offers competitive prompt adherence, visual quality, and temporal consistency, with native audio support. Meta says it is investing in improving areas where current performance still has gaps, such as audio-video synchronization and physically accurate fast motion. Muse Video ranks third in Arena’s text-to-video category.
Alongside the release of Muse Image, we’re sharing an early preview of Muse Video. It offers competitive performance in prompt adherence, visual fidelity, and temporal consistency.
We’re investing in areas with current performance gaps, such as audio-video synchronization and… https://t.co/kftRoTgq6W pic.twitter.com/iIFeFGLzoE
— AI at Meta (@AIatMeta) July 7, 2026
Currently, Muse Video is only available as an early preview and has not been released to the public, but people abroad have already been testing it, and its performance is excellent—it even rivals Seedance 2.0 (sometimes performing better), and is clearly superior to Google’s Veo 3.1.
Seedance 2.0 vs Meta Muse Video.
Both models feature native audio, but Muse Video seems like its currently stuck with a 10-second limit. Meanwhile, Seedance 2.0 can output 15-second cinematic sequences in crisp up to 4K.
Meta is cooking, but Seedance 2.0 might still hold the… https://t.co/gHH3pM4djU pic.twitter.com/e5LRpikLAv
— JSFILMZ (@JSFILMZ0412) July 8, 2026