glu
Studio
Join the waitlist
All comparisons
By scene · 25 models

Best AI image generator for thriller scenes

25 models compared (November 2026)

Best for thriller scenes: Recraft V4.1 Pro (8.2 out of 10), ahead of Qwen Image 3.0 Pro and Luma Uni-1, from our scores for prompt adherence, realism and film look.

We sent the same prompt, word for word, to the 25 best-ranked AI image models and put every result side by side. One image each, no retouching, no picking the best of several. Each card shows the model, its maker and its benchmark score.

The prompt

Wide shot at shoulder height down a cramped submarine control room, the captain centred in the foreground, crew stations receding into darkness behind him. A captain in his fifties with a salt-and-pepper beard grips an overhead pipe, eyes lifted to the hull, listening, a bead of sweat running from his temple into his collar. Beside him a sonar operator in her twenties presses one earphone to her head, frozen mid-breath. Sweat-darkened navy work shirts, valve wheels, gauges with blank dials, chipped grey paint, condensation beading on steel. The only light is caged red emergency fixtures along the ceiling, a hard deep-red top light, plus a dim green glow from a blank sonar screen as a cool rim on her cheek; 10:1. ARRI Alexa Mini LF, Zeiss Supreme Prime 25mm at T1.5, shallow focus, slightly handheld, Vision3 500T emulation with heavy grain in the reds, crimson grade with crushed blacks, a frame from a taut naval thriller. True-to-life skin with sweat sheen, no CGI look. No text, no logos, no watermarks.

Shape
16:9, or the nearest a model offers
Made
Models
25
How we ran it and ranked it

Every model got the prompt exactly as written above. Where a model can rewrite or expand a prompt before it draws, that was switched off, so each one worked from the same words. We made one image per model at 16:9; a model that cannot make 16:9 made its nearest shape, and its picture is trimmed slightly at the edges to fill the frame. Where a model drew black cinema bars into its picture, the bars are cut off for display. Nothing was retouched or chosen from several tries. A model that failed was tried once more; if it failed again, its card says so and why.

The order is the models' benchmark rank, not our opinion: their score in the Artificial Analysis text-to-image arena (human preference Elo, checked 8 November 2026), or, for a model Artificial Analysis does not score, the Arena text-to-image leaderboard (checked 6 November 2026). The two boards use different scales, so each score is placed by the share of its own board it beats. A model sold under two names is one model, shown once under both.

Only models whose pictures may be used commercially are included.

The answer at the top and the notes under What we noticed come from our own scores, not the benchmark. Each picture was scored 1 to 10 on prompt adherence, realism and film look, and the scores were checked by an independent cross-review.

The 25 results

Ranked by benchmark. Select a picture to see it full size, then use the arrows or your arrow keys to compare.

  1. #1

    Nano Banana 2

    · Google· AA Elo 1142
  2. #2

    Muse Image

    · Meta· AA Elo 1128
  3. #3

    Nano Banana Pro

    · Google· AA Elo 1119
  4. #4

    Qwen Image 3.0 Pro

    · Alibaba· AA Elo 1095
  5. #5

    Qwen Image 3.0

    · Alibaba· AA Elo 1073
  6. #6

    FLUX.3 Klein

    · Black Forest Labs· AA Elo 1043
  7. #7

    FIBO 2.0

    · Bria· AA Elo 1039
  8. #8

    Seedream 4.5

    · ByteDance· AA Elo 1030
  9. #9

    FLUX.2 [flex]

    · Black Forest Labs· AA Elo 1028
  10. #10

    Recraft V4.1 Utility

    · Recraft· AA Elo 1028
  11. #11

    Luma Uni-1 Max

    · Luma AI· AA Elo 1026
  12. #12

    Grok Imagine Image

    · xAI· AA Elo 1024
  13. #13

    FLUX.2 [max]

    · Black Forest Labs· AA Elo 1021
  14. #14

    Seedream 5.0 Lite

    · ByteDance· AA Elo 1019
  15. #15

    Recraft V4.1 Utility Pro

    · Recraft· AA Elo 1018
  16. #16

    Wan 2.6

    · Alibaba· AA Elo 1018
  17. #17

    FLUX.2 [pro]

    · Black Forest Labs· AA Elo 1003
  18. #18

    FLUX.2 [dev]

    · Black Forest Labs· AA Elo 1000
  19. #19

    FLUX.2 [dev] Turbo

    · Black Forest Labs· AA Elo 996
  20. #20

    Luma Uni-1

    · Luma AI· AA Elo 994
  21. #21

    Recraft V4.1 Pro

    · Recraft· Arena 1132
  22. #22

    FLUX.2 [dev] Flash

    · Black Forest Labs· AA Elo 989
  23. #23

    HunyuanImage 3.0 Instruct

    · Tencent· AA Elo 988
  24. #24

    Wan 2.7 Image

    · Alibaba· AA Elo 988
  25. #25

    Cosmos 3 Super

    · NVIDIA· AA Elo 987

What we noticed

Our read of these pictures. It's one image per model, so treat it as a snapshot, not a verdict.

  • Strongest: Recraft V4.1 Pro. Weakest: Wan 2.6.

    Recraft V4.1 Pro delivers the most prompt beats in one frame: the captain centred with his eyes lifted, a bead of sweat at his temple, a hard crimson top light, and the operator pressing her headphone at a blank screen, with a strong wide-lens feel. The top lamp is slightly white and no crew is visible behind. Wan 2.6 crops in so tight that the operator is half lost in shadow, and a second pair of arms reaches in from the left edge.

  • Wide framing and the room receding into darkness

    The prompt asks for a wide, cramped control room with stations falling away behind the captain. Seedream 4.5, Nano Banana 2 and Luma Uni-1 show real depth, with figures dimly visible down the corridor. Muse Image, Recraft V4.1 Utility Pro and Wan 2.6 push in to near close-ups, so the room becomes a backdrop. Cosmos 3 Super puts the operator's back and hair in the foreground, hiding the face the scene depends on.

  • Red top light, green rim: who held the colour split

    Luma Uni-1 Max gives the clearest split, with a red wash across the captain's chest and a teal tone on the operator. Nano Banana 2 and Nano Banana Pro keep the caged red lamps and a faint green glow on her face, but the faces are lit neutrally and the crimson grade is weak. FIBO 2.0, Grok Imagine Image and Wan 2.7 Image read as evenly lit grey rooms with a few red lamps, so the crimson grade and crushed blacks are mostly missing. Qwen Image 3.0 goes the other way and is very dark.

  • Blank sonar screen and sweat

    The prompt wants blank dials and a blank sonar screen. Nano Banana Pro, Qwen Image 3.0 Pro and Seedream 5.0 Lite keep the screen as a plain green glow, while Muse Image, FIBO 2.0 and FLUX.3 Klein fill it with radar sweeps or blobs. On sweat, Muse Image soaks both faces in droplets, which tips into rain, and Nano Banana Pro barely shows any.

  • Things nobody asked for

    FLUX.2 [flex], FLUX.2 [dev] Turbo and FLUX.2 [dev] Flash add shoulder patches, epaulettes and name tags, and the lettering on them is unreadable, despite the 'no text' instruction. Grok Imagine Image swaps the earphone for a telephone handset held to her ear, and FLUX.2 [max] gives the captain odd scratch-like marks on his cheek.

See how these models rank across all 10 scenes

Head to heads featuring these models

Two of the models above on the same film scenes, judged pair by pair.

Other scenes

The same 25 models on another film shot.

Join the waitlist

Sign in once. If you don't have an invite yet, that puts you on the waitlist, and we'll email you when your invite is ready. For people 18 or over.

Join the waitlist

Sold under more than one name, shown once: Nano Banana 2 (Nano Banana 2, Gemini 3.1 Flash Image Preview); Nano Banana Pro (Nano Banana Pro, Gemini 3 Pro Image Preview).