A new benchmark says the hardest part of AI image generation isn't drawing the scene - it's keeping the details straight.
Researchers built TRACE-Bench, a test suite for multi-reference image generation, where a model combines several source images into one output. Instead of grouping tasks by loose categories like subject composition, the benchmark breaks every prompt into four atomic operations - anchoring a subject, disentangling its traits, applying those traits, and composing the final scene. That structure produced roughly 1,600 test cases built from 631 formula templates and about 4,000 reference images spanning art styles and real-world subjects, with complexity scaled across eight difficulty tiers. The team ran nine leading models through it and scored each operation separately instead of judging only the final image.
The per-operation scoring is the useful part: it shows models aren't failing at the hard-sounding step of assembling a coherent scene. They're failing earlier, at disentangling an object's attributes from its source image and applying them correctly - the best model tested still scored just 0.74 on attribute fidelity. That's a diagnostic distinction, not a semantic one, and it tells developers where to actually spend their engineering effort.
Holistic image-quality scores have papered over this for a while; a benchmark that fails models on the right step, rather than the whole picture, is overdue.