GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 827 -5.3%
QWEN-38X 824 —
CL-OP55X 820 —
GPT-6A 820 —
GROK-46H 820 -5.2%
GLM-5 784 -8.4%
KIMI-K3X 742 -8.4%
CL-FAB5H 742 -5.7%
CL-OP5H 718 -6%
CL-OP5X 708 -18.2%
CL-OP46H 696 -6.2%
CL-OP47H 688 -6.1%
GEM-38FH 677 +0.1%
GEM-37FH 655 -24.3%
GPT-56S 619 —
GPT-55H 580 —
CL-OP47 579 -0.7%
INKL 531 —
GEM-31P 512 —
GEM-3P 498 —
CL-OP46 496 —
CL-OP48 489 -0.2%
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 827 -5.3%
QWEN-38X 824 —
CL-OP55X 820 —
GPT-6A 820 —
GROK-46H 820 -5.2%
GLM-5 784 -8.4%
KIMI-K3X 742 -8.4%
CL-FAB5H 742 -5.7%
CL-OP5H 718 -6%
CL-OP5X 708 -18.2%
CL-OP46H 696 -6.2%
CL-OP47H 688 -6.1%
GEM-38FH 677 +0.1%
GEM-37FH 655 -24.3%
GPT-56S 619 —
GPT-55H 580 —
CL-OP47 579 -0.7%
INKL 531 —
GEM-31P 512 —
GEM-3P 498 —
CL-OP46 496 —
CL-OP48 489 -0.2%
← Back to feed

Luma uni-1.1-max Enters Arena Image Leaderboards at #1 Human Preference Elo — $0.10 Per Image

Luma Labs’ uni-1.1-max entered Arena’s Text-to-Image and Image Edit leaderboards on May 5, 2026, landing at #1 on Human Preference Elo across overall, style and editing, and reference-based generation categories. The Arena Leaderboard Changelog confirms both uni-1.1 and uni-1.1-max were added simultaneously.

The model positions Luma as a top-3 lab on Arena’s image benchmarks — the first time the company has placed at that tier across both generation and editing tasks.

Architecture

Uni-1.1 takes a different approach from most image generation models. Reasoning and image generation run inside the same model architecture rather than in separate systems stitched together at inference time. Luma calls it a unified understanding and generation model: the same weights that process reasoning and visual understanding also drive generation.

The practical result, per Luma’s benchmarks: tighter adherence to multi-constraint prompts, cleaner reference image grounding, and editing that responds to intent rather than to prompt syntax.

On RISEBench — a benchmark evaluating Reasoning-Informed Visual Editing across Temporal, Causal, Spatial, and Logical categories — uni-1.1 achieves state-of-the-art results, leading on both overall reasoning and spatial logic. The RISEBench result is notable because spatial reasoning in image editing (accurate shadows, correct perspective, physically plausible object layouts) is a consistent failure mode for generation models that don’t integrate visual understanding tightly with the generation process.

Capabilities

Both uni-1.1 and uni-1.1-max support the same feature set through a unified API endpoint:

  • Text rendering — readable text on signs, labels, and surfaces
  • Spatial reasoning — accurate shadows, perspective, and object physics
  • Reference-guided generation — up to 9 reference images for text-to-image, 8 for editing
  • Multi-panel output — storyboards and sequential frames with consistent style
  • Web search grounding — searches for visual references before generating when enabled
  • Cultural styles — manga, ukiyo-e, film noir, and other visual traditions

Image editing uses a source parameter — the model modifies the source image based on the prompt while preserving unmentioned parts. Style transfer, background replacement, and object swaps are the primary use cases.

Pricing

ModelPer image (text-to-image, 2K)
uni-1.1$0.0404
uni-1.1-max$0.1000

Luma claims the models run at less than half the price and latency of comparable models. Generation time is approximately 31 seconds per image at the uni-1.1 tier. Provisioned throughput is available for production workloads requiring guaranteed capacity.

Context

Uni-1 launched in March 2026. The 1.1 release is the first update, adding improvements to output quality without changing the API surface — the same prompts, parameters, and aspect ratio options work identically on both model versions. Developers switch by changing only the model field.

Arena added both uni-1.1 variants on May 5, the same day gpt-image-2 (medium) received an updated score reflecting performance across Arena’s full user base at scale — Arena’s standard methodology for score stabilisation as sample sizes grow.

The image generation arena now has three distinct competitive tiers: proprietary frontier models (OpenAI, Google), specialised generation labs (Luma, Stability), and open-weight providers. Luma’s #1 Human Preference Elo position makes it the top-ranked non-frontier-lab entrant on the leaderboard.