Published: September 14, 2026
Image Model
Muse Image is Meta Superintelligence Labs' first image model — text-to-image and image-to-image (editing) with multi-reference composition.
Note: Muse's agentic mode (web search and code execution for exact text, charts and QR codes) is disabled in Artlist. Here, Muse runs as a standard single-pass Text-to-Image / Image-to-Image model.
Edits stay scoped to what was asked and hold across many turns. At launch it ranked No. 2 on Arena for text-to-image and for both single- and multi-image editing.
Key Features
Editing & Composition
- Scoped Edits: Changes only what the instruction names; subject, lighting and backdrop stay put
- Multi-Turn Stability: Assets survive consecutive edits without visible degradation
- Reference Composition: Up to 10 references — subject, product, environment, palette — composed, not averaged
Technical Capabilities
Details |
|
|---|---|
| Quality / Resolution | Fixed ~2K set by aspect ratio (16:9 → 2048×1152); no quality tiers |
| Aspect Ratios | 21:9, 16:9, 4:3, 3:2, 1:1, 2:3, 3:4, 9:16, 9:21 |
| Input Modalities | Text-to-Image, Image Editing |
| Reference Images | 1–10 per edit, no mask input |
| Output Formats | WebP (default), PNG, JPEG |
Limitations
- Agentic capabilities (web search, code-assisted rendering) are disabled in Artlist — accuracy of rendered text, charts and QR codes isn't guaranteed
- No resolution or quality control: ~2K, set by aspect ratio. No 4K path, no cheap draft tier
- No mask input — region-specific edits depend entirely on naming the region in the prompt
Prompting Tips
- Name what must stay, not just what changes: "recolour the label only, keep the backdrop"
- Tag references by role — subject, product, environment, palette — don't just dump ten images
- Set the aspect ratio up front; it's the only size control, and re-cropping means regenerating