We’re excited to launch Muse Picture and preview Muse Video, the primary media technology fashions developed by Meta Superintelligence Labs.
Muse Picture is our most superior picture technology mannequin but: it follows directions faithfully, edits with precision, composes from a number of references, and attracts on Instagram for social context. It additionally brings agentic device use capabilities and integrates with Muse Spark. Muse Video, constructed on the identical pretraining base, delivers distinctive visible constancy with native audio help.
Muse Picture is out there at present throughout the Meta AI app and on meta.ai, Instagram Tales within the US, and WhatsApp in restricted nations, and is coming quickly to Fb. Muse Video is coming quickly to creators and Meta AI.
Muse Picture: Agentic Picture Era
As a substitute of immediately mapping prompts to pictures, Muse Picture operates as an agent: it invokes search and coding instruments to enhance accuracy, self-refines its personal generations, and improves via scaling test-time compute. Muse Picture additionally integrates with Muse Spark, permitting the 2 fashions to share instruments and plan collectively for highly effective agentic media technology.
Instrument Use
We offer Muse Picture with entry to instruments to boost its agentic capabilities.
Coding. Throughout reinforcement studying, Muse Picture learns to jot down and execute code that produces correct plots and QR codes, and situation on rendered figures to enhance the accuracy of generated pictures. Muse Spark and Muse Picture additionally combine to make use of the mix of code and media technology to create animated GIFs, web sites with embedded pictures, and interactive visible video games.
Search. Muse Picture learns to look the online to floor generated pictures in factual and real-time info and visible references. Enabling search improves factual accuracy on knowledge-intensive prompts, significantly these involving present occasions and real-world info.
Muse Picture improves with search device use. Win price from inner ablation.
Self-Refinement
Muse Picture displays on and improves upon its personal work inside its chain of thought. This self-refining conduct can take completely different types: a neighborhood edit to the present picture draft when a small element is off, a brand new picture technology from scratch when bigger components are fallacious, or a unique tactic like device use for extra factually correct technology. We didn’t design this conduct. As a substitute, it emerged throughout RL coaching just because self-refinement produced higher pictures and subsequently increased reward.
Muse Picture improves with emergent self-refinement. Win price from inner ablation.
Self-refinement
Discovering reference pictures
Self-refinement
Assembling the unfold
I am placing collectively the shiny journal web page, weaving the proof steps, headline, and vogue portrait right into a cohesive format whereas checking typography, spacing, and the gold-accented particulars for a sophisticated editorial really feel.
Self-refinement
Refining the method
I am correcting the method to incorporate the lacking division slash, updating the format so the expression reads S = n(n + 1) / 2, and double-checking the encompassing textual content for readability and consistency.
Self-refinement
Checking the picture
I am reviewing the generated journal picture to substantiate the macro element, shiny format, and correct proof earlier than sharing it, and I am able to tweak the typography or swap the portrait if you would like. What would you want to regulate subsequent?
Take a look at-Time Compute Scaling
Like language fashions, Muse Picture improves the extra it thinks at inference time. With extra test-time compute, the mannequin causes extra, makes use of extra device calls, and makes use of extra self-refinement steps to enhance its generations. Growing reasoning power (and thus test-time compute) improves human-preference Elo scores and exhibits an roughly log-linear scaling relationship. Notably, this compute spans two very completely different sorts of labor — textual content tokens for reasoning, visible tokens for technology — but high quality is a operate of the mixed complete compute.
We discover that utilizing the token price range judiciously issues simply as a lot for efficient test-time scaling. Finest-of-N (BoN), the place the mannequin generates a number of pictures and retains one of the best, improves high quality early however saturates shortly. Spending that very same compute on deliberate reasoning scales significantly higher. Reasoning and gear use compound when mixed. Instruments let the mannequin attain past what it already is aware of, whether or not by trying to find references it lacks or writing code to get exact particulars proper, filling gaps that reasoning alone can’t.
Muse Picture improves with scaling test-time compute. Elo from inner ablation.
Picture Modifying
Muse Picture edits pictures with precision, altering precisely what the consumer asks for. It could actually observe a wide range of directions as our examples present.
Muse Picture maintains coherence throughout enhancing turns, supporting iterative refinement and open-ended brainstorming towards a goal outcome.
Multi-Reference Picture Composition
Muse Picture can compose components from many enter reference pictures within the immediate, together with individuals, objects, clothes, types, and environments. It helps interleaving textual content and pictures inline in prompts for complicated picture compositions.
Picture Benchmarks
Muse Picture holds the No. 2 spot on Area for text-to-image, single-image enhancing, and multi-image enhancing as measured by human choice Elo rankings on the time of writing.
Area Elo rankings as of July 5, 2026.
Area Elo rankings as of July 5, 2026.
Area Elo rankings as of July 5, 2026.
Previewing Muse Video
Alongside the discharge of Muse Picture, we’re sharing an early preview of Muse Video. It affords aggressive efficiency in immediate adherence, visible constancy, and temporal consistency. We’re investing in areas with present efficiency gaps, resembling audio-video synchronization and bodily correct quick movement. Muse Video is coming quickly to creators and in Meta AI.
On Area, Muse Video ranks No. 3 in human-preference Elo for text-to-video on the time of writing.
Area Elo rankings as of July 5, 2026
Content material Seal
To assist individuals confirm whether or not a picture is AI-generated, Muse Picture contains Content material Seal, our invisible watermarking system. Pictures created by Muse Picture within the Meta AI app and on meta.ai carry a hidden provenance sign that stays intact — even when cropped, compressed, resized, or screenshotted. We plan to increase Content material Seal to video quickly. We’re previewing a detection tool that permits you to verify whether or not a picture carries a Content material Seal watermark, offering an preliminary method that can assist you higher perceive if a picture was made with Meta AI.
Muse Picture in Meta Merchandise
Muse Picture connects deeply with the Meta ecosystem. Mixed with social instruments in Meta AI, customers can create pictures with pals and reimagine their Instagram pictures. Our ongoing investments in picture and video technology will additional allow creators and companies to generate dynamic content material throughout Meta merchandise.
Advertising belongings for small companies like @averyandme
Pictures generated in Meta AI with @-mention of public Instagram accounts.
Personalised presets immediately in Instagram
Discover further assets
Written by:
Meta Superintelligence Labs
Our latest updates delivered to your inbox
Subscribe to our publication to maintain up with Meta AI information, occasions, analysis breakthroughs, and extra.