fal on September 1, 2026 launched developer and enterprise access to Muse Image, the agentic image generation and editing model from Meta Superintelligence Labs, delivered through Meta Model API on fal’s platform. fal said the model plans each request, calls tools, and reviews its own output before returning it, and fal’s model page lists requests at $0.01 per image.
Muse Image is the first image generation model from Meta Superintelligence Labs. Most image models map a prompt straight to pixels in a single pass; fal said that approach works for simple prompts but can break down on complex briefs, pushing production teams to stitch together several models and run rounds of manual cleanup. fal also said frontier pricing has made the highest-volume workloads, including ad variant generation, catalog imagery, and per-user personalization, economically prohibitive at the scale where they matter most.
Tool Use and Self-Refinement
According to fal’s announcement, Muse Image uses a planner-plus-diffuser architecture with tool calls and refinement inside a single chain of thought. The model searches the web for real references, which fal said measurably improves factual accuracy on knowledge-intensive prompts: accurate logos, real places, working charts, and scannable QR codes. It writes and runs code for precise visual elements, and it checks and corrects outputs before returning them, a verification step fal said improves accuracy on multi-part briefs.
Meta’s July 7, 2026 technical post describes the same agentic design from the lab’s perspective: the model invokes search and coding tools to improve accuracy, self-refines its own generations, and improves through scaling test-time compute. During reinforcement learning, Muse Image learned to write and execute code that produces accurate plots and QR codes and to condition on rendered figures. Its self-refinement can take the form of a local edit when a small detail is off, a fresh generation when larger parts are wrong, or a tool call for more factual grounding. Meta said this behavior emerged during training because self-refinement produced better images and therefore higher reward, without being deliberately designed in.
Meta also reported that Muse Image improves the longer it reasons at inference time: with more test-time compute, the model reasons more, uses more tool calls, and applies more self-refinement steps, with human-preference Elo scores rising in an approximately log-linear relationship. That compute spans text tokens for reasoning and visual tokens for generation, with quality a function of the combined total. Meta found that best-of-N sampling, where the model generates several images and keeps the best, improves quality early but saturates quickly, while spending the same compute on deliberate reasoning scales considerably better.
Generation, Editing, and Composition
fal said one Muse Image model covers three workflows that production teams usually source from separate vendors. The first combines generation with precise editing: a user generates a design, then changes one element while the rest of the image stays unchanged, in a single call. The second is conversational multi-turn refinement, building one asset across many turns without quality drift. The third is reference-driven composition, in which a team feeds a brand kit and sample assets and generates new on-brand work matching style and subject across people, products, and environments.
Muse Image also shares tools and plans with Muse Spark, Meta’s assistant model, so a single request can return animated GIFs, sites with embedded images, and interactive visual output rather than a flat file, according to fal.
Arena Rankings and Availability
fal said Muse Image ranks in the top five on Arena across text-to-image, single-image editing, and multi-image editing. When Meta introduced the model, it reported that Muse Image held the No. 2 spot on Arena in all three categories under human-preference Elo rankings as of July 5, 2026. Meta also said Muse Video, a companion model built on the same pretraining base, ranked No. 3 in human-preference Elo for text-to-video as of that date.
The model is available on fal.ai through an interactive playground and the API. fal lists two endpoints, meta/muse-image/text-to-image and meta/muse-image/edit, both marked for commercial use and priced at $0.01 per image. The text-to-image page highlights faithful instruction-following and accurate rendering of fine details like text, plots, and QR codes; the editing page describes precise edits that change only what the user asks, hold coherence across turns, and compose from multiple reference images.
Muse Image first reached consumers on July 7, 2026 through the Meta AI app and meta.ai, Instagram Stories in the US, and WhatsApp in limited countries. Images created by Muse Image in the Meta AI app and on meta.ai carry Content Seal, Meta’s invisible watermarking system, which Meta said stays intact through cropping, compression, resizing, and screenshots; Meta is previewing a detection tool that checks whether an image carries the watermark.
fal describes itself as a generative media platform offering image, video, and audio models through a unified API, with on-demand serverless GPUs and dedicated compute clusters, and says more than 2.5 million developers use it.
