The paper introduces Muse Image and Muse Video, the inaugural media generation models from Meta Superintelligence Labs. Muse Image distinguishes itself by transitioning from simple prompt-to-image mapping to an agentic architecture that utilizes reasoning, tool use, and self-refinement to enhance generation quality.

Core Methodology and Agentic Capabilities

Unlike traditional generative models, Muse Image functions as an agent that executes a chain of thought to solve complex visual tasks. Its core technical pillars include:

• Tool-Augmented Reasoning: The model is equipped with specific tools to bridge gaps in knowledge and visual accuracy.
  • Coding: During reinforcement learning (RL), the model learns to write and execute code. This allows it to generate accurate mathematical figures (e.g., $S = \frac{n(n + 1)}{2}$), render QR codes, and condition visual outputs on these generated figures for improved precision.
  • Search: The model performs web searches to ground generations in factual, real-time data and to acquire relevant visual references for knowledge-intensive prompts.
• Emergent Self-Refinement: Through RL training, the model developed the capability to reflect on its own drafts. This process is not hard-coded; rather, it emerged because iterative refinement consistently yielded higher rewards. The model performs local edits to resolve minor errors, restarts generations if major flaws exist, or pivots to tool use when factual verification is required.
• Test-Time Compute Scaling: The model demonstrates a log-linear scaling relationship between compute spent at inference time and human-preference Elo scores. Crucially, the model utilizes a combined compute budget for both textual reasoning (chain of thought) and visual token generation. Performance is optimized by prioritizing "deliberate reasoning" over simple Best-of-N (BoN) sampling, as reasoning scales more effectively and avoids early saturation.

Functional Features

• Integration and Composition: Muse Image supports seamless integration with Muse Spark, enabling the joint planning and execution of complex media tasks such as creating interactive games, websites with embedded images, or animated GIFs. It also enables multi-reference composition, allowing users to interleave text and multiple images to define styles, characters, and environments.
• Precision Editing: The model maintains high coherence across iterative editing turns, allowing users to perform complex, multi-step refinements while preserving the target composition.
• Provenance: To address AI safety and transparency, Muse Image incorporates Content Seal, an invisible watermarking system. This provenance signal remains embedded in the file even after cropping, resizing, or compression, and is supported by a detection tool for verification.

Muse Video

Built on the same pretraining foundation as Muse Image, Muse Video is an early-stage model focused on competitive prompt adherence, visual fidelity, and temporal consistency. It supports native audio integration and is currently undergoing improvements in physical motion accuracy and audio-visual synchronization. At the time of the paper, it holds the No. 3 ranking for text-to-video on the Arena human-preference leaderboard, while Muse Image ranks No. 2 across text-to-image and editing benchmarks.