Ask an AI image model for a person and you usually get one. Ask it for a hand holding a coffee cup, a second person in the background, or the same face twice, and the illusion starts to leak. Anatomy is the last frontier of generated photography: skin and light now read as real, but fingers still fuse, identities still drift, and the defects hide in the few hundred pixels nobody zooms into.
A current image-model leaderboard publishes criterion-level scores that let you separate these problems instead of guessing. Lined up against real per-image prices and generation times, they tell a fairly clear story about which model to reach for — and where no leaderboard can save you.
Seven models, scored where it counts
Scores are Elo-style from blind head-to-head voting; one model (Grok Imagine 2.0) sits at 1,000 as the fixed anchor, so every number reads as better or worse than Grok. Compare within a column only.
| Model | Overall | Film | Realism | Editing | Ref. adherence | Price | Speed |
|---|---|---|---|---|---|---|---|
| Seedream 5.0 Pro | 1,051 | 1,074 | 1,142 | 1,037 | 1,037 | $0.09 | 88.9s |
| GPT Image 2 | 1,047 | 1,058 | 1,058 | 1,045 | 1,035 | $0.057 | 44.8s |
| Nano Banana Pro | 1,008 | 1,059 | 1,184 | 1,035 | 1,002 | $0.134 | 39.9s |
| Grok Imagine 2.0 | 1,000 | 1,000 | 1,000 | 1,000 | 1,000 | $0.06 | 18.3s |
| Nano Banana 2 | 985 | 998 | 1,064 | 1,032 | 991 | $0.101 | 18.1s |
| Qwen Image 3.0 | 966 | 964 | 975 | 992 | 973 | $0.075 | 99.1s |
| Flux.2 Pro | 927 | 978 | 1,022 | 955 | 881 | $0.08 | — |
One honest caveat before any of it is useful: these boards score realism, identity consistency and editing precision. There is no dedicated anatomy test, so nothing here counts fingers for you. Realism is the closest published proxy — and the spread is wide enough to mean something, from Nano Banana Pro's 1,184 down to Qwen's 975. Where confidence intervals overlap, call it a tie.
Reading the table for human subjects
Seedream 5.0 Pro is the safe default for people. It leads Overall and Film, tops creativity, and its 1,037 on reference adherence is the number that decides whether a face you feed it comes back looking like the same person. It also does transparent backgrounds. The costs: a 2K ceiling, ten reference images, and nearly ninety seconds per frame. It also dips on graphic design, so set typography elsewhere.
GPT Image 2 is the repair tool. It leads image editing at 1,045 and quietly leads prompt adherence at 1,041 — which matters precisely when you are specifying which hand holds what. It outputs 4K, accepts sixteen references and a 32,000-character prompt, runs in about forty-five seconds, and at $0.057 it is the cheapest of the seven. Its aesthetic ceiling trails the leaders. The pattern that works: generate elsewhere, fix here.
Nano Banana Pro is the realism leader by a margin. 1,184 on realism, 4K output, fourteen references, under forty seconds. Two brakes: it is the most expensive model here at roughly 2.4 times GPT Image 2, and it scores below the anchor on subjective aesthetics and graphic design. It reads as technically convincing more often than it reads as beautiful.
Nano Banana 2 is the draft model. Eighteen seconds at 4K makes it the fastest route to a composition you like before committing a slower model to the final render. Its realism holds at 1,064 even if overall sits at 985 — treat output as a sketch.
What to avoid for people. Flux.2 Pro's 881 on reference adherence rules it out for recurring characters. Qwen Image 3.0's 915 on prompt adherence means "left hand only, gripping the handle" is likelier to come back wrong, and it is the slowest model on the board. Grok is the anchor: a useful reference point and cheap, but not a recommendation.
The number that matters most and is hardest to read
Reference adherence is the least intuitive column and, for series work, the most important. The test behind it is concrete: one man, three candid street shots — walking, leaning on a wall, crossing a road — same face, hairstyle and outfit every time. The leading model keeps the navy sweater, the jeans, the boots and the backpack across frames; the weakest keeps the sweater but flattens the jeans into plain dark trousers and loses the backpack entirely. Same brief, wildly different survival rate.
If your work involves a recurring character, a brand ambassador or an AI persona, that column outranks the overall leaderboard. Identity drift is what kills series, not lighting.
Why hands still break in 2026
A hand has sixteen independently moving joints, most of them occluded behind each other or behind an object in any given shot. The model infers anatomy it cannot see, from a region that may only be a few hundred pixels wide.
A March 2025 study including researchers at Peking University quantified it: on a 500-prompt human benchmark, close to half of generated images contained some distortion, and their annotated companion dataset found most flawed regions occupied only a small share of the frame — which is exactly why defects survive a quick glance. That work tested an older generation of models, so read it as proof the problem is real and hard to spot, not as a ranking of the seven above.
What has changed is where failures live. Centered portraits and open palms are mostly solved. The errors migrated to harder places: fingers wrapped around a handle, two people holding hands, a hand near a face, a foreshortened arm, and the small faces standing behind your subject.
A two-hour test that beats any leaderboard
Scores narrow the field; your prompts decide it. Pick three candidates, give each identical prompts, references, aspect ratio and resolution, and generate four images per prompt. Retry only on service errors — never because you disliked a result, or you are scoring your own patience.
Six prompts cover most failure modes: a hand gripping an object by its handle; two people holding hands; a hand raised near a face; a left-handed action; a foreshortened arm reaching toward camera; a group of five with faces in the background. Add one identity test: the same reference image across all three models.
Then score blind, at full resolution, because a defect invisible in a feed is obvious at 100%. On hands: finger count, separation, thumb placement, wrist continuity, joint direction, and whether the grip actually touches the object. On faces: gaze, teeth, ears, profile geometry, unintended asymmetry. Count every visible person, including the small ones at the back. Report two numbers per model — pass rate on first outputs, and pass rate across all four. A model that lands it once in four creates more work than its average suggests.
Fixing anatomy without regenerating the person
Prompting is prevention. State how many hands appear and what each one does — "one visible left hand holding a coffee cup by the handle" gives the model far more to work with than "a person with coffee." Keep the first generation simple: crossed arms, overlapping fingers, three props and four people in one prompt is where things fall apart. Frame for the anatomy you care about; a face that occupies forty pixels has forty pixels of detail no matter which model you chose.
When something fails, mask only the defective region and ask for one specific correction — natural thumb placement, not "fix the hand." A narrow mask leaves the model less room to redraw your person. Then compare the repair against the original at full resolution and inspect the boundary: sleeves, jewelry, shadows and contact points are where a local edit quietly stops being local.
Choosing, and staying honest about it
Leaderboards re-rank as new models ship, and the top two here sit four points apart — close enough that one release reorders them. Before standardizing, check the current numbers on the four signals that matter for human subjects: overall, film, image editing, and the realism criterion inside film.
The short version: Seedream 5.0 Pro for people who must stay recognizable, Nano Banana Pro for the most convincing skin, GPT Image 2 for the cheap fast repair, Nano Banana 2 for drafts — and a fixed six-prompt test on your own briefs before any of them gets written into your pipeline.
