The standard Glif tests

Each test gives every model the same prompt. The Glif team votes blind on which output is better, and an AI judge scores each one. Open one of these 62 tests to see every output side by side, best first. Tests only one model has run are marked Demo.

Image 36

  • n0wnxl0trvh7lbxceuvhe.jpgtovcz3waakpx0ybo8zxus.jpgs58qko64qk26wzbg7utd6.jpgft45e8iq29tzjgqoir8zk.jpg

    Gibberish prompt

    26 models · Winner: Muse Image

    Turns nonsense words into a coherent, striking surreal image instead of rendering the words as text.

    Text-to-image · Reliability · Style

  • 7xkwv5m3a8vrjpqpb2lcw.pngavoqvqyzx1jt9j84nryxf.jpgmbcihzw1yyzv3znlozlut.jpg8x7nzpz8e4q7r7ntwzlmd.jpg

    Long and detailed prompt

    26 models · Winner: GPT Image 2.5

    Adheres to a long and detailed prompt.

    Text-to-image · Prompt adherence · Spatial reasoning

  • u842kgzvctl8h8u4wj00u.jpg6eo9j9l9wegfcei8t4o3j.jpg0hbo91jqfmgwwxpttig0s.jpg1078152yf5to68vby6adr.jpg

    Magic laptop

    26 models · Winner: FLUX.2 Pro

    Creatively merges the ancient/mystical elements with a laptop. Runes are present and the screen shows a correctly themed UI.

    Text-to-image · Style · World knowledge

  • p3oyp75zf7fjp5iyz41oj.jpgvxgx5gv2dajwi3wu6mxt1.jpg6g44op02gykkzcoc1i5bx.png2tsw2w7cwlpct47ram3qt.jpg

    Miniature scene

    26 models · Winner: Nano Banana Pro

    Can do tilt-shift scenes. Keeps proportions correct. Correctly produces physically-sound trebuchet. Creates plausible medieval setting.

    Text-to-image · Style · World knowledge

  • 1xxj61z9sedc6jb4rlbp8.jpga400735jkctcskq6zit3e.pngwn9dy1akvcz1ouqjn1hen.jpgoym06ptb9jiluq6d3r1m9.png

    Portrait realistic

    26 models · Winner: Muse Image

    Human is realistic and has no artifacts. Applies Kodachrome film effect.

    Text-to-image · Style · Reliability

  • 2qk6q7rnlxuyyah7rnmlg.jpgqmis9w7m9za896qdt2c3r.pngkl164omz7jzdiohad9vct.jpgt2wdpd88m4p3o4qyhr7ak.jpg

    Social media selfie

    26 models · Winner: Z-Image Generator

    Has correct camera positioning. Looks realistic.

    Text-to-image · Composition · Reliability

  • jpeblhwdewscjb72lb8qt.jpgoh8sz5zb0ivi2cagrx1ii.jpgg9cmkll4t8jm9ecwfmxv4.pngha72fk7z2y37mvejiyjqu.jpg

    Stacked animals

    26 models · Winner: Gemini 3.1 Flash

    Stacks the animals in the correct order on top of each other. Keeps the 1980s photo look and the German medieval town setting.

    Text-to-image · Spatial reasoning · Prompt adherence · Style

  • fkyb87xygsh8dvdzy9u96.jpgunk4eu5n7qv2y1iabagpd.png5tr9me3lv4nrj8eaw0n7j.jpgyodswjgoeibc0etc7ie85.jpg

    Tall forehead

    26 models · Winner: Gemini 3.1 Flash

    Renders the exaggerated, very tall forehead. No double eyes. Adheres to 1920s black-and-white rubber hose cartoon style.

    Text-to-image · Prompt adherence · Style

  • up39xai6narcio7lynka1.jpgcctq0bkrkuil11ofqbsmm.pngi2u9rrq7h6bzp5f1cay64.jpg81bp93tiwpmk8qkso08l9.jpg

    Trading card

    26 models · Winner: FLUX.2 Flex

    Produces a plausible trading-card layout with legible text. Anime style.

    Text-to-image · Composition · Typography · Style

  • snhax0j4o0n30l9b8htv5.jpgoeilzjxhwg587svxf28ta.jpgh5jm6l7bdh7i5r38msq0u.jpgqy9od7235ynmpbri2x5lt.jpg

    Vague prompt

    26 models · Winner: Gemini 3.1 Flash

    Picks a clear, good-looking image from a one-word abstract prompt instead of noise or a muddle.

    Text-to-image · Reliability · Style

  • mzm0328n87yetkauz01k2.jpgfyl3tf3fr6ypca2lqnaha.jpg923pcskkznvs74j6izple.jpgyv4n8k531aozkyv3xl1ya.png

    Edit parts of image

    23 models · Winner: Krea 3

    Cybernetics should be perfectly integrated into the subject. Likeness should be preserved. Legible text and annotations.

    Image-to-image · Image editing · Consistency · Typography

  • mkqo6jrc1arq2lxjpstf7.jpg6k6l1s6u6nce0z8mr3m7e.jpgr1myjqnudmesn8qfvtcgi.jpgecpvio1fevcs19msj09c2.png

    Product placement

    23 models · Winner: Seedream V4.5

    Preserves the reference can exactly (label, colors, proportions). Composes a believable bench scene in Ghent. The can should be sized proportionally.

    Image-to-image · Reference following · World knowledge

  • zpmjzakcv99j70rf2a4li.pngqcvluujq3hnthx4otmw7s.jpg3d1u8rs727f3pcbawtha9.png8fqsxcm15hjvjr3t1dxjx.jpg

    Cat in space

    22 models · Winner: GPT Image 2.5

    Adheres to all elements in the prompt. Physically-sound space scene.

    Text-to-image · Prompt adherence · World knowledge

  • 6ycxy2cp9kek45dtney7z.jpgr3efc53nopheju7p8xwqq.jpged2nd2r6k8cfdmuj3909t.jpggdkr83szhyjjvbdxvk2hn.png

    Edit from markup (garden)

    22 models · Winner: Phota Image Editor

    Edits only the marked up area. Removes markup. Knows plant names and appearances.

    Surgical image editing · Image editing · World knowledge

  • y4gbs2tftxes0pmg0t2zp.jpgn2zsh5z73iuwef0tm64fd.png9ydjzvpg2zvkqsjv597kt.jpgh8sjyngdz0tu4vvi0b05n.jpg

    CMYK texture

    22 models · Winner: Muse Image

    Designs an interesting trading card and applies realistic CMYK dot pattern texture. Applies the texture with the right scaling.

    Image-to-image · Reference following · Composition

  • ykejfj1ehrk9sh3oy95n9.jpgwfbdwp5v0sgz0mv5maqsl.pngoqrryc8ytssqy2u7y6vo8.jpg9njscnrtc7g4rgeerzc6v.jpg

    Character sheet reference

    21 models · Winner: Muse Image

    Places the exact same character, in the same art style, into a new scene.

    Character consistency · Reference following · Consistency · Spatial reasoning

  • 4o84lm3v5sx67chqm670c.jpg67uray7814k24gd8imn6h.png6kkuz5a71yb7zqpnvqxpl.jpggz11hh2xl1ta07wt3ibrz.jpg

    Magazine cover

    21 models · Winner: Muse Image

    Uses the exact same character and follows the brief. Good typography and layout.

    Character consistency · Consistency · Typography · Composition

  • qunjl8fjzj8nwxxg9lj72.jpgjtd0x8nn75bkwov91fqrj.jpgdij8p8yqlpj9tot0xxy56.jpgf07gef3mvkwsj5oo2ylps.jpg

    Compose image

    20 models · Winner: Phota Image Editor

    Combines the referenced woman, white coat and handbag into one coherent photorealistic fashion shot. Identity and products match the references.

    Character consistency · Reference following · Consistency

  • 2ef602jw8vh52t8egu7in.pnggagjp3v5qnz884zvo3kjx.jpg2a112eyxvss0rxzkh4kxb.jpg89ffm1bfni70dqxv56gdf.jpg

    Style references

    19 models · Winner: GPT Image 2.5

    Draws a cat cyborg in the style of the reference images, combining features from all the reference images.

    Multi-image reference · Reference following · Style

  • ktdyg7t6obqctfm2eqjl3.jpgd4egyqvjueleoa9xgwiyr.png236qy6cv3czhnplgshaik.pngof9au9ectfhac1ic3ku2v.jpg

    Beer closeup condensation

    18 models · Winner: Seedream 5.0 Lite

    Preserves the exact can design from the reference (label, colors, red 'glif' wordmark) at macro distance. Condensation and droplets look physically real; wordmark stays crisp and legible.

    Product photography · Reference following · Consistency

  • m1y6g5pqh5coqrvcnekjl.pngn5aj9jdxcympxfaxqqemd.jpgf7kfubeika6jl8i3t67xc.jpgvu73bxfez6iost0v4xtmp.jpg

    Beer festival

    18 models · Winner: GPT Image 2.5

    Every visible can matches the reference exactly with legible labels. Believable golden-hour festival scene, crowd and stage softly blurred, natural hands and faces.

    Product photography · Reference following · Composition · Consistency

  • ns4ycqqi9f4sns9w14nz8.jpgeshqs9n2shb1kp4yam605.pngkm5w8mg7doj4a2vz7zs6m.jpgk6l7p0kp5b0vo1wmo63sc.jpg

    Beer ad text

    17 models · Winner: Nano Banana 2.1

    On-image text carries the signal: crisp headline and 'glif' wordmark, clean cream-and-red ad layout with negative space. The can itself is preserved exactly as the hero object.

    Brand and text rendering · Typography · Reference following · Composition

  • dlcej9lcy98lzuri9z5f6.jpgjw0pstp8upisslza46081.jpgh41mr59gcwa4x7tn4gjg5.pngq1a4sabpogfycgord0bi4.jpg

    Text rendering 2

    9 models · Winner: FLUX 3 Image

    Has black, blue and yellow tints. 3D object. Correct lettering and good typography. All info is legible and present.

    Text-to-image · Typography · Style

  • tztcddyc3twx7kig5ppvc.jpgj8cojiw4so1aq6qcnxk3j.pngwgc6s9uzsv1btej2arzmi.pngmnp4disio1k4k34evot7m.webp

    Upscale concrete floor

    5 models · Winner: Gemini 3.1 Flash

    Preserves and sharpens the real-world concrete texture without (VAE) artifacts or invented detail.

    Image upscaling · Image editing · Reliability

  • tcn674vqsnaeq1yn3u7az.png91n7hihkv2ttf6qzzw9vp.pngjij9x24i6id86q4k72bi9.jpg0xc1l6j2b1io40h0fmj2s.webp

    Upscale eye

    5 models · Winner: Clarity Upscaler

    Doubles the resolution and keeps the eye's real texture (iris, lashes, skin) without inventing new detail.

    Image upscaling · Image editing

  • 40gu5jcwd0vmze5gycmxz.png0vxbg0hs7v1d31s9g6itu.webpc0hxlonn70nyq8knmkjrl.pngt1vqrg8dz9pq0qvnoeq70.jpg

    Upscale lowpoly

    5 models · Tied first: Clarity Upscaler, P-Image Upscale

    Keeps the low-poly game look; does not hallucinate realistic detail.

    Image upscaling · Image editing · Style

  • ill5azulvfd91nvrrcupw.pngok25jbxienuflb0oydqiq.webpgq7yt6xqrwbju1l1sk583.jpgiv6y2jsqm906k7886fdz8.png

    Upscale painting

    5 models · Winner: Clarity Upscaler

    Preserves the watercolor texture; does not turn the painting photographic.

    Image upscaling · Image editing · Style

  • r95mkygljyijpva24jh21.jpgjikg3iw5q55r7bodrcgd9.pngwshkxa24p980ddywpk2fc.pngs36mjiygz9jxjn6pk8lwd.webp

    Upscale portrait

    5 models · Winner: Gemini 3.1 Flash

    Adds convincing detail without changing the person's identity. Face and anatomy stay correct.

    Image upscaling · Image editing · Consistency

  • vmzajessfvij3983c7e3w.pnggoqaflf0exan4m8jlw3kf.pngxfgsw0bijgj675knq73p5.png

    Remove background

    3 models · Winner: Rembg Background Remover

    Clean matte around the hair/curls. Fully transparent background, no leftover fringes.

    Remove background · Image editing · Reliability

  • uxh1tw24x4d8qy82i9ftf.jpg

    Outpaint asymmetric

    Demo
    1 model

    Extends 1024px to the right, continuing the alley, bricks and lighting consistently. No visible seam.

    Image outpainting · Image editing · Composition

  • 11if4dmymcadltxyt3ax5.jpg

    Outpaint uniform zoomout

    Demo
    1 model

    Zooms out evenly on all four sides. Scene stays coherent; no seams or repetition.

    Image outpainting · Image editing · Composition

  • r764xeq184w7guqkpnang.jpg

    Outpaint vertical extend

    Demo
    1 model

    Extends the photo so that more of the man is visible. Extended part should be consistent with the original photo.

    Image outpainting · Image editing · Composition

  • nistavdk1o5cawpbq8cxr.jpg

    Outpaint widen portrait

    Demo
    1 model

    Extends left and right into a landscape with no visible seams. New content matches lighting and style.

    Image outpainting · Image editing · Composition

  • x0ttewqb6a71mbixlxotl.png

    Segment beer can

    Demo
    1 model

    One mask covers the whole can (body, top rim, bottom) with tight edges and none of the white background. The overlay is drawn cleanly over the original image.

    Image segmentation · Prompt adherence · Image editing

  • kebfsj9fi4wbt69p9tgod.png

    Segment person curly hair

    Demo
    1 model

    The mask covers the whole person including individual curls, with no halo into the background and no eaten hair. Nothing else in the image is masked.

    Image segmentation · Prompt adherence · Image editing

  • 0vuzgkwksgwzsybzal9s6.png

    Text removal beer can

    Demo
    1 model

    Every piece of text is gone: the red 'glif' wordmark, 'Average GLIF PILS', the ring of text at the top, '330 ml' and 'ALC. 5,2% VOL.'. The hop mascot, can shape, colors and lighting are untouched, with clean fills where the text was.

    Text removal · Image editing · Reliability

Video 18

  • Miniature diorama

    35 models · Winner: Gemini Omni Flash 1.1 · Reference to video

    Keeps the tilt-shift miniature look while adding coherent motion. Motions are physically accurate.

    Image-to-video · Style · World knowledge

  • iPhone handheld shot

    33 models · Tied first: Grok Imagine · Reference to video, Hailuo H3 Max Turbo (MiniMax), Seedance 2.0 Mini

    Handheld phone-selfie camera motion. Looks realistic; the character stays consistent.

    Image-to-video · Spatial reasoning · Consistency · Reliability

  • Animate portrait

    31 models · Tied first: Gemini Omni Flash 1.1, Kling V3 Pro · Image to video

    Animates the person while preserving identity. Ends on a questioning expression; motion is coherent.

    Image-to-video · Consistency · Prompt adherence · Reliability

  • Animate radar

    31 models · Winner: FLUX.3

    Animates the radar module with believable mechanical motion: the dish turns smoothly and nothing warps.

    Image-to-video · World knowledge · Reliability

  • Face likeness

    31 models · Winner: Kling O3 Pro · Reference to video

    The person shown at the end is recognizably the same as in the input image. Temporally stable through the bus pass.

    Image-to-video · Consistency · Reliability

  • Comic book history style

    26 models · Winner: Hailuo H3 Max Turbo (MiniMax)

    Produces a 4-panel comic layout telling a history-of-the-world narrative. Text and layout are legible. Style and story are consistent throughout the panels.

    Text-to-video · Composition · Typography · Consistency

  • Action scene

    24 models · Winner: Seedance 2.5

    Smooth camera follows the kickflip. The board and rider move with believable physics.

    Text-to-video · Spatial reasoning · World knowledge · Reliability

  • Frame to frame transition watch

    22 models · Winner: Seedance 2.5

    Interpolates smoothly from the first to the last frame. The legs are spinning correctly. The watch is spinning in the correct direction.

    First-to-last frame · Reference following · Spatial reasoning

  • Reference video: character kitchen

    8 models · Winner: Grok Imagine · Reference to video

    The character keeps its exact design and art style from the reference sheet throughout. The kitchen is isometric, the actions happen in order and the motion is smooth.

    Multi-image reference · Reference following · Consistency · Style

  • Video to video (anime style)

    6 models · Tied first: FLUX.3 · Edit video, Gemini Omni Flash 1.1 · Edit video, Grok Imagine · Edit video, Kling O3 Pro · Edit video

    Turns the man into an anime character while keeping his identity and the original motion. Temporally consistent.

    Video-to-video · Image editing · Style · Consistency

  • Reference video: fashion walk

    5 models · Tied first: Hailuo H3 Max (MiniMax), Seedance 2.0 Reference

    Takes the right element from each reference: her likeness, the white coat and the handbag all match the images and stay consistent for the whole clip. Natural runway walk, no morphing.

    Multi-image reference · Reference following · Consistency · Reliability

  • Extend ghost alley

    2 models · Winner: FLUX.3 · Extend video

    The continuation starts seamlessly from the final frame. The ghost, alley, graffiti and lighting stay consistent, the drift away reads clearly, and any generated audio fits the scene.

    Extend video · Consistency · Reliability · Prompt adherence

  • Keyframes car orbit return

    2 models · Winner: FLUX.3 Keyframes

    The trajectory looks like plausible camera movement. No glitches.

    Keyframes-to-video · Reference following · Spatial reasoning · Reliability

  • Keyframes watch pinned endpoints

    2 models · Winner: FLUX.3 Keyframes

    Two keyframes pinned to frame 0 and frame 120 (start and end of a 5s clip at 24fps). Interpolates smoothly, legs spinning, correct rotation direction. Comparable to the first-to-last-frame watch test on other models.

    Keyframes-to-video · Reference following · Spatial reasoning

  • Audio driven shepherd

    Demo
    1 model

    The clip is driven by the provided audio: lips match the words, delivery feels natural, and the shepherd keeps his identity and river setting.

    Audio-to-video · Reference following · Consistency · Reliability

  • Motion control character

    Demo
    1 model

    Transfers the reference video's motion onto the character while preserving the character's identity.

    Motion control · Reference following · Consistency

  • Subtitles glass

    Demo
    1 model

    Dynamic preset: words animate in sync with speech, karaoke-style highlighting. Glass / blurred-backdrop style. Subtitles match the audio.

    Video subtitling · Typography · Reliability

  • Subtitles simple

    Demo
    1 model

    Basic preset: one static caption block per phrase. Minimal, clean style. Subtitles match the audio.

    Video subtitling · Typography · Reliability

Audio 8