Summary for Hands & Anatomy

GPT Image 2.6 Sunburst is the official category leader at 1350 Elo. Muse Image follows at 1310 Elo, while FLUX 3 Image and Nano Banana 2.1 share third place at 1251 Elo.

Leading models

Ordered by category Elo—not by the accompanying 1–10 averages.

Position Model Category Elo Judge average /10 Main takeaway
1 GPT Image 2.6 Sunburst 1350 8.0 Strongest broad combination of anatomy, contact, and instruction following
2 Muse Image 1310 7.5 Excellent photographic texture and strong results at much lower recorded latency
=3 FLUX 3 Image 1251 7.3 Particularly useful for explicit poses, connected groups, and matched mirror views
=3 Nano Banana 2.1 1251 7.6 Strong hand close-ups and natural material rendering
5 GPT Image 2.5 Sunburst 1215 7.3 Polished execution; especially compelling drawing example
6 Qwen-Image-3.0-Pro 1174 7.1 Excellent atmosphere, with standout drawing and high-five examples
7 MiniMax-H3-Image 1161 7.0 Notable strengths in running, mirror correspondence, and literal group composition
8 GPT Image 2.6 Flare 1142 6.3 Strong finish, but conspicuous finger-count failures in some close-ups

The leader's 40-point advantage over second place is meaningful but modest: approximately a 56% expected comparison win rate, not overwhelming dominance. The 13-point difference between sixth and seventh is firmly in the close range.

What matters most

  • Surface realism is ahead of structural reliability. Sharp pores, nails, fabric, and lighting frequently coexist with extra digits, confused grips, or incorrect limb connections.
  • No leader wins every scenario. The category winner scores only 6/10 on both the heart-shaped hands and the runner.
  • Relational instructions are major separators. Correct participant counts, neighboring handholds, actual pencil contact, and the specified mirror orientation matter as much as attractive anatomy.
  • Newer or Quality labels are not guarantees. Performance remains strongly dependent on the requested interaction.

Quick choices

  • Best overall starting point: GPT Image 2.6 Sunburst.
  • Best speed-quality balance among the leaders: Muse Image.
  • Promising for literal complex poses and connected arrangements: FLUX 3 Image.
  • Promising for detailed hand close-ups: Nano Banana 2.1.
  • For a specialized task: inspect the matching examples below rather than assuming the category leader is always the best fit.

General Analysis & Useful Insights

1. The leaders succeed in different ways

GPT Image 2.6 Sunburst combines 9.1/10 prompt adherence, 8.5 technical quality, and 8.0 realism/coherence. Its strongest images establish convincing finger lengths, contact pressure, and hand-to-object relationships. The close-up handshake and two-handed laptop typing, both 9/10 overall, illustrate that advantage.

Muse Image has slightly higher adherence at 9.3/10 and particularly strong natural surface variation. Its older hands forming a heart and forest runner both score 9/10. However, clear digit visibility does not rescue the stiff proportions in its hand holding an apple, which scores 6/10.

The joint-third models offer different trade-offs:

  • Nano Banana 2.1 records 7.8/10 realism/coherence, with convincing nails, skin variation, and contact shadows in its strongest close-ups.
  • FLUX 3 Image records 9.0/10 adherence, but lower realism/coherence at 7.0/10. Its open, inspectable compositions can be excellent, while crowded contact regions remain vulnerable.

Their identical 1251 Elo ratings should therefore remain a tie, despite different judge averages and visual profiles.

2. Detail is not anatomy

Several models generate persuasive local textures without maintaining the larger structure.

DALL-E 3, at 708 Elo, averages 7.2/10 artistic merit but only 3.8/10 realism/coherence. Its close-up handshake has dramatic lighting and dense texture, yet the evaluation identifies an apparent extra digit and unresolved hand interaction.

Midjourney v7, at 805 Elo, similarly averages 7.6/10 technical quality against 5.9 realism/coherence. Its drawing-hand image scores 3/10 overall despite strong focus and reflective materials, because the grip appears to contain an extra digit.

A particularly clear contrast appears within the leading model family. GPT Image 2.6 Flare has 8.2/10 technical quality, close to the category leader's 8.5, but only 6.3 realism/coherence. Its apple-holding hand scores 3/10 because of apparent six-digit anatomy.

Practical lesson: inspect joints, digit ownership, and contact before judging pores or sharpness.

3. Separate three kinds of failure

Structural impossibility: Extra limbs, merged fingers, or incompatible wrist connections are the most serious defects. The MiniMax Image-01 complex yoga image scores 2/10 overall and 1/10 realism/coherence because multiple limb attachments are incoherent.

Visibility failure: A concealed thumb, blurred fingertip, or cropped foot can prevent verification without proving incorrect anatomy. Even the strongest close-up handshakes retain some normal grip occlusion.

Instruction failure: A physically plausible scene can still depict the wrong action. Examples include one hand resting on the trackpad instead of both hands typing, a hoop replacing joined hands, or a conventional mirror portrait reversing the specified views.

These distinctions matter when deciding whether an image needs regeneration, reframing, or a more explicit prompt.

4. Interaction geometry is a recurring bottleneck

The difficult region is usually where objects or bodies meet:

  • In high-fives, compressed fingers and separated palm planes often produce a static palm press or glancing touch rather than impact.
  • In pencil drawing, some tips hover, look metallic, or fail to connect convincingly with the marks.
  • In laptop typing, plausible hands can sit above warped key rows and inconsistent hardware.
  • In five-person hand circles, attractive bodies surround grips whose ownership becomes difficult to trace.

This explains why an attractive scene can receive only a moderate overall score. The peripheral rendering is often better than the requested interaction.

5. Dynamic anatomy rewards restraint

The instruction to make limbs well-defined often produces over-carved muscles, inflated shoulders, rigid ankles, and exaggerated airborne strides.

The stronger running examples instead combine readable joints with believable arm-leg opposition, appropriate stride length, and coherent grounding. MiniMax-H3-Image earns 10/10 realism/coherence for its runner, while the category leader's runner scores 5/10 on that criterion because of sculpted anatomy and leap-like extension.

Likewise, an authentic yoga posture is not automatically the requested posture. Many complex yoga images show valid bent-knee balances while missing the explicit all-limbs-extended constraint.

6. Improvements are substantial, but uneven

Within this category:

These gains do not imply universal prompt-level superiority. An earlier model can still produce an excellent individual image. Nor should stylization itself be treated as failure when photography was not requested; physical coherence remains the relevant anatomical test within the chosen style.

Best Models by Use Case

The recommendations below combine category Elo with specific prompt evidence. They are not separate subtask Elo rankings. Individual 1–10 results identify useful examples, not guaranteed repeatable winners.

Close-up hand interaction

Two people shaking hands

Start with GPT Image 2.6 Sunburst, Muse Image, or joint-third Nano Banana 2.1. Each produces a 9/10 example.

The category leader's handshake has particularly convincing joint placement, overlap, and pressure. The Muse Image handshake adds character through denim, leather, jewelry, and differentiated textures.

Watch out: none establishes fully independent visibility of every digit. If counting is essential, request a more open grip and a camera angle that exposes the far thumb.

A single hand holding an object

A hand holding a red apple

The leading shortlist is GPT Image 2.6 Sunburst, Nano Banana 2.1, and FLUX 3 Image. All have 9/10 examples with clear five-digit arrangements.

The FLUX 3 Image apple study is especially informative: it receives 10/10 for adherence and detail execution, with separated digits and coherent support. The Nano Banana 2.1 example combines clear finger positions with natural surface variation.

For a lower-ranked specialist worth testing, Ideogram V2, 960 Elo, also produces a 9/10 apple image. That isolated success does not move it above the category leaders.

Complex poses with extended limbs

A full-body complex yoga pose

FLUX 3 Image is the strongest high-ranked fit in the supplied examples. Its wide-straddle handstand scores 9/10 overall, 10 adherence, and 9 realism/coherence. Straight limbs and an open silhouette make the anatomy inspectable.

Recraft V4.1, 1109 Elo, and Nano Banana 2, 1051 Elo, also produce 9/10 examples: an extended outdoor balance and a raised-leg side plank.

Prompting tip: name a compatible pose explicitly. A straight-arm straddle handstand is less ambiguous than asking for complex yoga, which often elicits a bent-knee dancer pose.

Dynamic two-person contact

Two people high-fiving

GPT Image 2.6 Sunburst is the clearest starting point. Its high-five scores 9/10 overall and 10 adherence, combining believable hands with expressions and airborne particles that support the impact moment.

Qwen-Image-3.0-Pro supplies a 9/10 café high-five with especially cohesive setting and anatomy. Ideogram 4.0 (Quality), 1053 Elo, offers a 9/10 hand-focused alternative.

Watch out: motion streaks do not establish contact. Check palm alignment and finger boundaries first.

Drawing hands and precise tool contact

A hand sketching with a pencil

GPT Image 2.5 Sunburst, Qwen-Image-3.0-Pro, and Nano Banana 2 each deliver 9/10 examples.

The Qwen botanical study is particularly well matched to the brief: 10/10 adherence, a credible grip, and a clear graphite-to-paper junction. The GPT Image 2.5 Sunburst landscape balances hand focus with an accomplished sketch.

Choose this kind of composition when both tool mechanics and the surrounding artistic process matter. Avoid prompts that encourage an extreme macro crop so narrow that supporting fingers or the contact point fall out of focus.

Two-handed typing

A person typing on a laptop

GPT Image 2.6 Sunburst provides the strongest high-ranked default. Its typing scene scores 9/10, with coherent hardware and differentiated finger bends.

Other 9/10 examples come from GPT Image 2, 1120 Elo; GPT Image 1.5, 1087 Elo; and FLUX.2 Max, 1025 Elo. The FLUX.2 Max close-up is especially convincing in wrist continuity, relaxed thumbs, and separate key contacts.

Quality check: verify that both hands are on keys—not one on the trackpad—and inspect the keyboard as a physical object, not just as background texture.

Symmetrical hand gestures

Two hands forming a heart

Muse Image is the highest-rated model with a 9/10 result here. Its older-hand heart preserves natural wrinkles, pigmentation, and credible contacts.

GPT Image 2.5 Flare, 1137 Elo, and Grok Imagine 2.0, 1134 Elo, also produce 9/10 examples: a balanced contrasting-hand heart and a crisp symmetrical heart. Their category ratings are effectively close.

Important distinction: different hand sizes can be realistic. The relevant checks are plausible joints, readable tucked fingers, and a clean heart opening—not mechanically identical hands.

Exactly five people holding hands

Five people joining hands in a circle

For a literal neighboring-hand chain, FLUX 3 Image is the clearest high-ranked example. Its five-person circle scores 9/10 overall and 10 adherence, with all five connections traceable.

GPT Image 2.6 Sunburst also scores 9/10 with a highly convincing five-person overhead huddle, but the hands form a central stack. That is less literal if the intended output requires an open ring.

MiniMax-H3-Image produces an 8/10 literal circle, with strong composition but less refined fingers.

Prompting tip: specify exactly five people, each holding the hands of their immediate neighbors, with an empty center. Explicitly exclude a central team stack if that distinction matters.

Full-body running and sports anatomy

A runner mid-stride

Muse Image is the highest-category-rated model with a 9/10 running example. Its forest runner balances exertion, anatomical clarity, and environmental motion.

MiniMax-H3-Image also scores 9/10 overall, with 10 realism/coherence for its track runner. Seedream 5.5, 1128 Elo, and FLUX.2 Max provide further 9/10 examples with natural profile running and a convincing sprint phase.

Avoid: exaggerated bodybuilding language when the goal is realistic running. Request natural muscle tension, complete head-to-toe framing, and a sharp athlete against restrained background motion.

Front/back reflection consistency

A person with their front and reflected back visible

Strong literal examples come from FLUX 3 Image, MiniMax-H3-Image, and GPT Image 1.5, each scoring 9/10.

The FLUX 3 Image mirror scene clearly matches sleeves, shirt hem, hairstyle, trousers, and footwear. The MiniMax-H3-Image portrait maintains dress straps, hairstyle, and arm positions across both views.

GPT Image 2.6 Sunburst and Muse Image provide strong 8/10 correctly oriented alternatives.

Critical distinction: a direct back with a reflected front can be physically correct but still be wrong for this prompt. Specify that the person faces the camera, with the mirror behind them showing their back. Then compare clothing, hair, accessories, and pose—not just the presence of two views.

Choosing for iteration speed

These are arithmetic means of the 10 recorded generation times for each listed model, not controlled benchmarks or cost estimates.

Model Category Elo Mean recorded time
GPT Image 2.6 Sunburst 1350 136.6 s
Muse Image 1310 13.4 s
FLUX 3 Image 1251 39.5 s
Nano Banana 2.1 1251 23.1 s
MiniMax-H3-Image 1161 13.7 s
Nano Banana 2 Lite 935 3.0 s

Practical recommendation: use Muse Image for fast, high-quality iteration, and test GPT Image 2.6 Sunburst when broad category performance matters more than latency. Nano Banana 2 Lite is useful for rapid drafts, but its lower category rating warrants stricter anatomical review.

Before final delivery, check counts, limb continuity, contact geometry, required visibility, and cross-view consistency. Those checks catch more consequential failures than judging sharpness alone.