Best AI image model for edits, characters and text: a 7-model test

We ran seven AI image models on 18 real jobs, from edits to characters and exact text. GPT Image 2.5 Flare won the edits; Nano Banana Pro won nothing.

We gave seven AI image models the same 18 jobs a Nodaro user runs every day: edits, characters from a reference photo, scenes, exact text and seven built-in presets. GPT Image 2.5 Flare was the best editor. GPT Image 2 was the best at characters and exact text, and the slowest. Nano Banana 2.1 won the scenes and the illustration. Nano Banana Pro, a popular choice for character work, won nothing.

Models: 7
Prompts: 18
Images: 112
Rounds: 3
Images per prompt: 1
Picks: blind

What we found

Edits

GPT Image 2.5 Flare is the most reliable editor.

It won two of the three edits and kept the pose, the framing and the product design on all three.

Characters

GPT Image 2 leads on characters and exact text.

It made the best reference grid, kept both people right on the cast sheet and drew the cleanest infographic. It is also the slowest, with a median of 97 seconds.

Scenes

Nano Banana 2.1 wins scenes and illustration.

It won the rainy street, the watercolor and the labelled cast sheet. It changed the outfit when it put a person in a new scene.

The surprise

Nano Banana Pro won none of 18 prompts.

It broke the grid layout, dropped a name label, changed outfits and returned a nearly empty infographic: 8 weak marks in 11 prompts.

Presets

GPT Image 2 won four of the seven presets.

The character board, the product shot, the claymation scene and the age progression. Nano Banana 2.1 won the expression sheet and the 4:5 studio portrait, and GPT Image 2.5 Sunburst won the 2:3 movie poster.

Budget

Seedream 5 Flash is the best model for its price.

At 11 credits and about 45 seconds an image, it tied the earlier pick on 7 of 18 prompts. It beat none, and it mixed up the two people on the cast sheet.

The scoreboard

Round 1 ran five models on 11 prompts. A pick is the best result of a prompt, and a weak mark is a result that missed the brief.

Model	Picks	Weak marks	Median time	Credits per 2K image
GPT Image 2.5 Flare	4	1	64 s	28
GPT Image 2	3	2	97 s	42
Nano Banana 2.1	3	1	68 s	22
GPT Image 2.5 Sunburst	1	1	59 s	28
Nano Banana Pro	0	8	51 s	50

In round 2, the seven presets, GPT Image 2 won four, Nano Banana 2.1 two and GPT Image 2.5 Sunburst one. Nano Banana Pro won none again. In round 3, neither Seedream model beat an earlier pick.

How we ran it

Every model got the same prompt, aspect ratio and output size, on Nodaro, on 2026-10-08. Each model made one image per prompt.
Round 1 ran five models on 11 prompts in four groups: edits, identity from a reference photo, text to image, and exact text.
Round 2 ran seven of Nodaro's built-in presets on three models.
Round 3 ran all 18 prompts again on Seedream 5 Pro and Seedream 5 Flash, and compared each result with the earlier pick.
Claude, an AI model, made every pick and weak mark blind: it saw the results without the model names.
One image per prompt is a hint, not proof. The pattern across prompts is the finding.
The edits and the identity prompts start from four photos from Unsplash, shared on Wikimedia Commons under CC0: Brunette woman portrait, Gray-haired man portrait, Grado Headphones SR80e and Confident Eye Contact.

Round 1: five models, 11 prompts

Each model made one image per prompt, at the same aspect ratio and size. The pick of each prompt is marked, and so is every weak result. The galleries show the models in the same order every time.

Edits

Each edit starts from a photo and asks for one change. The rest of the photo must stay as it is.

Swap the clothes

The prompt

Change her black top to a tailored emerald-green blazer over a plain white t-shirt. Keep her face, hair, pose, framing, background and lighting exactly the same.

The starting photo: a young woman with long auburn hair in a black top, in front of a green hedge — The starting photo

GPT Image 2.5 Flare's result for “Swap the clothes” — GPT Image 2.5 Flare50 s · 2368×1776Full-size original

GPT Image 2's result for “Swap the clothes” — GPT Image 297 s · 2688×2016Full-size original

Nano Banana 2.1's result for “Swap the clothes” — Nano Banana 2.168 s · 2400×1792Full-size original

GPT Image 2.5 Sunburst's result for “Swap the clothes” — GPT Image 2.5 Sunburst68 s · 2368×1776Full-size original

Nano Banana Pro's result for “Swap the clothes” — Nano Banana Pro51 s · 2400×1792
Moved the camera closer; framing drifted
Full-size original

Replace the background

The prompt

Replace the grey studio background with a bright modern open-plan office, softly out of focus. Keep the man, his face, glasses, polo shirt, pose and lighting unchanged.

The starting photo: a gray-haired man with glasses in a polo shirt, in front of a gray studio background — The starting photo

GPT Image 2.5 Flare's result for “Replace the background” — GPT Image 2.5 Flare126 s · 1776×2368
Kept pose and framing exactly
Full-size original

GPT Image 2's result for “Replace the background” — GPT Image 297 s · 2016×2688
Changed his pose and crop
Full-size original

Nano Banana 2.1's result for “Replace the background” — Nano Banana 2.197 s · 1792×2400Full-size original

GPT Image 2.5 Sunburst's result for “Replace the background” — GPT Image 2.5 Sunburst56 s · 1776×2368Full-size original

Nano Banana Pro's result for “Replace the background” — Nano Banana Pro44 s · 1792×2400
Shrank and moved him in frame
Full-size original

Restage a product photo

The prompt

Remove the hand. Place these exact headphones on a walnut desk next to a closed laptop, in warm morning window light, premium product photography. Keep the headphones' exact design, materials, colours and logo.

The starting photo: a hand holding a pair of headphones — The starting photo

GPT Image 2.5 Flare's result for “Restage a product photo” — GPT Image 2.5 Flare86 s · 2368×1776Full-size original

GPT Image 2's result for “Restage a product photo” — GPT Image 2127 s · 2688×2016Full-size original

Nano Banana 2.1's result for “Restage a product photo” — Nano Banana 2.168 s · 2400×1792
Correct but flatter light
Full-size original

GPT Image 2.5 Sunburst's result for “Restage a product photo” — GPT Image 2.5 Sunburst50 s · 2368×1776Full-size original

Nano Banana Pro's result for “Restage a product photo” — Nano Banana Pro51 s · 2400×1792
Shrank the headphones and merged the cups
Full-size original

Identity from a reference photo

Each prompt starts from a photo of a person, and the person must stay recognizable.

Put the person in a new scene

The prompt

The woman from the reference photo sitting at a small café table on a Paris street, laughing, holding a cup of coffee, natural daylight, candid 35mm photograph. Keep her exact face, skin tone and hair.

The starting photo: a young woman looking straight at the camera — The starting photo

GPT Image 2.5 Flare's result for “Put the person in a new scene” — GPT Image 2.5 Flare51 s · 1776×2368Full-size original

GPT Image 2's result for “Put the person in a new scene” — GPT Image 279 s · 2016×2688
Face, hair and top kept
Full-size original

Nano Banana 2.1's result for “Put the person in a new scene” — Nano Banana 2.168 s · 1792×2400
Changed her top to a black coat; face small in frame
Full-size original

GPT Image 2.5 Sunburst's result for “Put the person in a new scene” — GPT Image 2.5 Sunburst56 s · 1776×2368Full-size original

Nano Banana Pro's result for “Put the person in a new scene” — Nano Banana Pro51 s · 1792×2400
Changed her top
Full-size original

Character reference grid

The prompt

Create a clean character reference grid using the attached photo of the person as the single source of truth for face, hair, eyes, skin tone, build and outfit. Four equal panels arranged in a tidy two-by-two grid on a single seamless neutral light-grey studio background: a tight front-facing portrait close-up, a 3/4-view portrait, a full-body shot from the front in a relaxed neutral stance, and a full-body shot from the back. The SAME person in every panel — identical age, features, hairstyle and outfit from the attached photo. Identical soft, even studio lighting in all four panels with no dramatic shadows. No decorations, no background props, no text, no UI elements — a sterile, production-neutral reference sheet built to be reused as an identity reference in later generations. Photorealistic, sharp focus, accurate natural skin texture, true-to-life color.

GPT Image 2.5 Flare's result for “Character reference grid” — GPT Image 2.5 Flare68 s · 2480×3312Full-size original

GPT Image 2's result for “Character reference grid” — GPT Image 279 s · 2448×3264Full-size original

Nano Banana 2.1's result for “Character reference grid” — Nano Banana 2.1121 s · 3584×4800Full-size original

GPT Image 2.5 Sunburst's result for “Character reference grid” — GPT Image 2.5 Sunburst98 s · 2480×3312Full-size original

Nano Banana Pro's result for “Character reference grid” — Nano Banana Pro68 s · 3584×4800
Broke the layout: no full-body front, cropped back view
Full-size original

Two-person cast sheet

The prompt

Create a clean multi-character cast reference sheet composed as horizontal strips stacked vertically in one image — one strip per character from the attached reference images, in the order they are attached, using exactly as many strips as there are attached characters. Each strip shows that character in four aligned studio panels: tight front portrait close-up · 3/4 portrait · full body front · full body back, all in a relaxed neutral stance. Label each strip on the left edge with the character's name in clean bold lettering: MAYA · DAVID. Every character must match their attached reference exactly — same face, hair, skin tone, build and outfit — with no identity blending between rows. Single seamless neutral light-grey studio background across the whole sheet, identical soft even lighting in every panel, consistent panel sizing and alignment. No decorations, no props, no UI elements beyond the row name labels. A sterile production casting sheet built to be reused as the single identity reference for this cast in later generations. Photorealistic, sharp focus, natural skin texture, true-to-life color.

GPT Image 2.5 Flare's result for “Two-person cast sheet” — GPT Image 2.5 Flare69 s · 2480×3312Full-size original

GPT Image 2's result for “Two-person cast sheet” — GPT Image 2109 s · 2448×3264
Both identities right; labels placed oddly
Full-size original

Nano Banana 2.1's result for “Two-person cast sheet” — Nano Banana 2.198 s · 3584×4800Full-size original

GPT Image 2.5 Sunburst's result for “Two-person cast sheet” — GPT Image 2.5 Sunburst85 s · 2480×3312
Dropped the MAYA label
Full-size original

Nano Banana Pro's result for “Two-person cast sheet” — Nano Banana Pro68 s · 3584×4800
Dropped the MAYA label and changed David's shirt
Full-size original

Text to image

Three prompts with no starting image: a scene, a product and an illustration.

Rainy street at night

The prompt

A rainy Tokyo side street at night, neon signs reflecting on wet asphalt, a lone cyclist with a clear umbrella passing a ramen shop, cinematic photograph, shallow depth of field.

GPT Image 2.5 Flare's result for “Rainy street at night” — GPT Image 2.5 Flare56 s · 2048×1152
Lost the clear umbrella
Full-size original

GPT Image 2's result for “Rainy street at night” — GPT Image 2109 s · 2048×1152
Lost the clear umbrella
Full-size original

Nano Banana 2.1's result for “Rainy street at night” — Nano Banana 2.197 s · 2752×1536Full-size original

GPT Image 2.5 Sunburst's result for “Rainy street at night” — GPT Image 2.5 Sunburst68 s · 2736×1536Full-size original

Nano Banana Pro's result for “Rainy street at night” — Nano Banana Pro44 s · 2752×1536Full-size original

Product photo of a mug

The prompt

Studio product photograph of a matte black ceramic coffee mug with a wisp of steam, on a light oak surface, soft diffused window light, minimal clean background.

GPT Image 2.5 Flare's result for “Product photo of a mug” — GPT Image 2.5 Flare56 s · 2048×2048Full-size original

GPT Image 2's result for “Product photo of a mug” — GPT Image 268 s · 2048×2048Full-size original

Nano Banana 2.1's result for “Product photo of a mug” — Nano Banana 2.168 s · 2048×2048
Usable; odd handle shape
Full-size original

GPT Image 2.5 Sunburst's result for “Product photo of a mug” — GPT Image 2.5 Sunburst57 s · 2048×2048Full-size original

Nano Banana Pro's result for “Product photo of a mug” — Nano Banana Pro45 s · 2048×2048Full-size original

Watercolor picture-book illustration

The prompt

Children's picture-book watercolor illustration of a fox and a hedgehog sharing a red umbrella in an autumn forest, soft washes, warm palette, gentle linework.

GPT Image 2.5 Flare's result for “Watercolor picture-book illustration” — GPT Image 2.5 Flare68 s · 2368×1776Full-size original

GPT Image 2's result for “Watercolor picture-book illustration” — GPT Image 2142 s · 2688×2016Full-size original

Nano Banana 2.1's result for “Watercolor picture-book illustration” — Nano Banana 2.159 s · 2400×1792Full-size original

GPT Image 2.5 Sunburst's result for “Watercolor picture-book illustration” — GPT Image 2.5 Sunburst50 s · 2368×1776Full-size original

Nano Banana Pro's result for “Watercolor picture-book illustration” — Nano Banana Pro50 s · 2400×1792
Close second; taste call
Full-size original

Exact text

Two prompts that spell out every word the image must show.

Concert poster with exact text

The prompt

Concert poster with this exact text, spelled exactly: headline "MIDNIGHT ECHOES", subline "Live at the Blue Room", details "Friday 14 November · 9 PM", footer "Tickets $25". Bold modern typography, deep navy and gold, clean layout.

GPT Image 2.5 Flare's result for “Concert poster with exact text” — GPT Image 2.5 Flare59 s · 1776×2368Full-size original

GPT Image 2's result for “Concert poster with exact text” — GPT Image 285 s · 2016×2688Full-size original

Nano Banana 2.1's result for “Concert poster with exact text” — Nano Banana 2.153 s · 1792×2400Full-size original

GPT Image 2.5 Sunburst's result for “Concert poster with exact text” — GPT Image 2.5 Sunburst60 s · 1776×2368Full-size original

Nano Banana Pro's result for “Concert poster with exact text” — Nano Banana Pro43 s · 1792×2400
Rendered as a skewed photo of a print
Full-size original

Infographic with exact labels

The prompt

Clean flat infographic titled "How Coffee Gets to Your Cup" with four numbered steps, each with a simple icon and one short label: 1 Harvest, 2 Dry, 3 Roast, 4 Brew. Exact spelling, white background, friendly brown palette.

GPT Image 2.5 Flare's result for “Infographic with exact labels” — GPT Image 2.5 Flare1776×2368Full-size original

GPT Image 2's result for “Infographic with exact labels” — GPT Image 280 s · 2016×2688Full-size original

Nano Banana 2.1's result for “Infographic with exact labels” — Nano Banana 2.11792×2400Full-size original

GPT Image 2.5 Sunburst's result for “Infographic with exact labels” — GPT Image 2.5 Sunburst1776×2368Full-size original

Nano Banana Pro's result for “Infographic with exact labels” — Nano Banana Pro1792×2400
Tiny and mostly empty
Full-size original

Round 2: seven built-in presets

Nodaro's presets are ready-made prompts for one job each. We ran seven of them on Nano Banana Pro, Nano Banana 2.1 and GPT Image 2, each with its own prompt and ratio. The Movie Poster preset is 2:3, a ratio GPT Image 2 cannot draw, so it ran on GPT Image 2.5 Sunburst instead.

Character Board

The prompt

Create a single high-resolution, densely packed character reference sheet titled "CHARACTER BOARD" using the attached photo of the person as the single source of truth for face, hair, beard, eyes, skin tone and body proportions. The same person must appear in every panel — same age, same features. Outfit identical to the attached photo across all panels. All on-image labels in ENGLISH. Editorial reference-board layout with a dark near-black background, thin yellow neon accent light, faint film-grain overlay, and production-grade character-design UI. The composition should feel organized but not rigidly locked: allow the main portrait, metadata, and supporting panels to shift position naturally within the board. The panel arrangement can vary from generation to generation while remaining readable, balanced, and premium. Design the board so it works cleanly in different aspect ratios without depending on a fixed left-column structure. Include: A large hero portrait or 3/4-length character image, accompanied by a detailed metadata block: NAME · AGE · HEIGHT · BUILD · HAIR · EYES · FEATURES · OUTFIT (describe the exact clothing from the attached photo) · CHARACTER · MOOD. Also include six content groupings — render ALL six panel headings exactly as written, never merging or omitting a panel: PANEL 01 — VIEWS (5 full-body shots, neutral pose, identical outfit and lighting, neutral tan backdrop): FRONT · 3/4 LEFT · SIDE LEFT · BACK · 3/4 RIGHT. PANEL 02 — EXPRESSIONS (5 tight headshots, same lighting): CALM · LAUGHING · INTENSE · CONTEMPLATIVE · CONFIDENT. PANEL 03 — DETAILS (3 macros): face/eyes close-up · hand close-up (showing ring, watch or cuff) · distinctive outfit detail. PANEL 04 — OUTFIT FLAT-LAYS (5 isolated product shots on dark background): outerwear · top · bottom · footwear · main accessory. PANEL 05 — LIGHTING / MOOD (4 same-pose portraits under different lighting): SOFT WINDOW LIGHT · GOLDEN HOUR · COOL BLUE NIGHT · DRAMATIC RIM-LIGHT. PANEL 06 — COLOR PALETTE: 6 swatches with HEX codes derived from the outfit and skin tones. Bottom caption: "Use this character board as a visual reference for consistent depiction of the character across all generations." Bottom-right tags: STYLE · Modern · Realistic · Cinematic. Style: photorealistic, no illustration. Consistent face across every single panel. 8K, fine grain, cinematic color grading.

GPT Image 2's result for “Character Board” — GPT Image 2111 s · 2048×1152
Densest board, every panel, closest to the photo
Full-size original

Nano Banana 2.1's result for “Character Board” — Nano Banana 2.198 s · 2752×1536Full-size original

Nano Banana Pro's result for “Character Board” — Nano Banana Pro56 s · 2752×1536
Invented a name; layout fine
Full-size original

Character Sheet · Expressions ×6

The prompt

character expression model sheet of a cheerful freckled teen with messy red hair, six head-and-shoulders portraits arranged in two even rows of three, labeled emotions in order — neutral, happy, angry, surprised, sad, determined — strictly consistent face, hairstyle, and features across every panel with only the expression changing, flat even portrait lighting identical in all six panels, evenly spaced on a plain neutral backdrop, clean concept-art model sheet, natural facial detail and matching color palette

GPT Image 2's result for “Character Sheet · Expressions ×6” — GPT Image 280 s · 1448×1086
Photographic, but dropped the emotion labels
Full-size original

Nano Banana 2.1's result for “Character Sheet · Expressions ×6” — Nano Banana 2.149 s · 1200×896
Labelled, one consistent face
Full-size original

Nano Banana Pro's result for “Character Sheet · Expressions ×6” — Nano Banana Pro36 s · 1200×896Full-size original

Product Shot (white bg)

The prompt

e-commerce studio product photograph of a matte-black wireless headphone, on a seamless pure-white background, 100mm lens at f/8 with deep focus, soft even high-key lighting from a large diffused softbox with gentle fill, perfectly centered with generous margins, clean catalog-ready mood with true-to-life color, crisp edges and accurate material texture

GPT Image 2's result for “Product Shot (white bg)” — GPT Image 265 s · 1024×1024Full-size original

Nano Banana 2.1's result for “Product Shot (white bg)” — Nano Banana 2.141 s · 1024×1024Full-size original

Nano Banana Pro's result for “Product Shot (white bg)” — Nano Banana Pro40 s · 1024×1024
Product small; a softbox in frame
Full-size original

Movie Poster

The prompt

theatrical movie poster for 'NEON TIDE', cinematic key art in a gritty sci-fi thriller mood, dramatic vertical hero composition with a commanding central figure, moody atmospheric lighting with strong rim light and volumetric haze, rich filmic teal-and-amber color grade, bold title treatment along the bottom with a short tagline above it and a small billing-block strip, premium high-detail finish, legible correctly-spelled text

Nano Banana 2.1's result for “Movie Poster” — Nano Banana 2.127 s · 848×1264
Garbled small credits at the top
Full-size original

GPT Image 2.5 Sunburst's result for “Movie Poster” — GPT Image 2.5 Sunburst94 s · 1024×1536Full-size original

Nano Banana Pro's result for “Movie Poster” — Nano Banana Pro34 s · 848×1264Full-size original

Claymation Scene

The prompt

a tiny clay fisherman rowing across a stormy plasticine sea, tactile stop-motion claymation scene — real plasticine clay characters and set with visible thumbprints, tool marks, seam lines and a soft matte sheen, miniature studio photography with shallow depth of field and warm practical lighting, handmade Aardman-quality charm. NOT digital CG, NOT a 3D render, NOT 2D illustration — a physically sculpted miniature scene photographed on a real stop-motion stage

GPT Image 2's result for “Claymation Scene” — GPT Image 289 s · 1672×941
Storm, lighthouse, clay all there
Full-size original

Nano Banana 2.1's result for “Claymation Scene” — Nano Banana 2.155 s · 1376×768Full-size original

Nano Banana Pro's result for “Claymation Scene” — Nano Banana Pro36 s · 1376×768
Flat light, no storm
Full-size original

Studio Portrait

The prompt

studio portrait of a confident woman with sleek dark hair, 85mm lens at f/2.8, shallow depth of field, soft octabox key with gentle fill and a subtle hair light, clean seamless gray backdrop, tight head-and-shoulders framing, polished editorial mood with a neutral true-to-life palette, natural skin texture, catchlights and tack-sharp eyes

GPT Image 2's result for “Studio Portrait” — GPT Image 279 s · 1086×1448
Drawn at 3:4 — it has no 4:5
Full-size original

Nano Banana 2.1's result for “Studio Portrait” — Nano Banana 2.127 s · 928×1152
Natural studio light at the preset's 4:5
Full-size original

Nano Banana Pro's result for “Studio Portrait” — Nano Banana Pro35 s · 928×1152Full-size original

Three-Age Triptych

The prompt

Single photorealistic studio composite of the SAME person from the reference image as a three-age triptych, shoulder-to-shoulder left to right: around age 10, around age 40, and around age 75. Editorial portrait photography, 85mm lens at f/2.8, each figure framed chest-up, facing camera with the same calm expression. CRITICAL identity lock — same eyes, eye color, brow, nose, ears, jawline and facial bone structure across all three, unmistakably the same person; age ONLY via skin texture, fine lines, hairline, hair color (darker in youth, greying with age), and a slight shift in facial fullness. Age-appropriate clothing per stage. Even soft cinematic studio key light with gentle rim separation, clean dark-grey gradient seamless backdrop, refined neutral grade, natural skin tones, tack-sharp facial detail; the three figures fill the frame.

GPT Image 2's result for “Three-Age Triptych” — GPT Image 2107 s · 1086×1448
One person, three ages
Full-size original

Nano Banana 2.1's result for “Three-Age Triptych” — Nano Banana 2.1201 s · 896×1200Full-size original

Nano Banana Pro's result for “Three-Age Triptych” — Nano Banana Pro40 s · 896×1200Full-size original

Round 3: Seedream 5 Pro and Seedream 5 Flash

We ran every prompt of rounds 1 and 2 again on Seedream 5 Pro and Seedream 5 Flash, and compared each result with that prompt's earlier pick: a tie, close to it, worse, or a fail.

Model	Tie	Close	Worse	Fail	Median time	Credits per image
Seedream 5 Flash	7	6	4	1	45 s	11
Seedream 5 Pro	5	3	9	1	100 s	20–66

Neither model beat an earlier pick on any prompt. Seedream 5 Flash came closest, at a fraction of the price: a good draft model, not a quality default.

Swap the clothes

The same prompt as Swap the clothes.

Seedream 5 Pro's result for “Swap the clothes” — Seedream 5 Pro90 s · 2304×1728
Identity, hair and pose kept; the jacket swap is as clean as Flare's.
Full-size original

Seedream 5 Flash's result for “Swap the clothes” — Seedream 5 Flash52 s · 2368×1776
Identity, hair and pose kept; jacket swap as clean as Flare's.
Full-size original

Character Board

The same prompt as Character Board.

Seedream 5 Pro's result for “Character Board” — Seedream 5 Pro191 s · 1312×736
Strong look, but the small text is gibberish and the outfit varies between panels.
Full-size original

Seedream 5 Flash's result for “Character Board” — Seedream 5 Flash45 s · 1424×800
Panels exist, but the labels are garbled and several panels are wrong.
Full-size original

Character Sheet · Expressions ×6

The same prompt as Character Sheet · Expressions ×6.

Seedream 5 Pro's result for “Character Sheet · Expressions ×6” — Seedream 5 Pro97 s · 1184×896
Six clear expressions with correct labels, but drawn semi-illustrated, not photographic.
Full-size original

Seedream 5 Flash's result for “Character Sheet · Expressions ×6” — Seedream 5 Flash27 s · 1152×864
Six clear, labelled expressions, consistent face; slightly illustrated rather than photographic.
Full-size original

Product Shot (white bg)

The same prompt as Product Shot (white bg).

Seedream 5 Pro's result for “Product Shot (white bg)” — Seedream 5 Pro79 s · 1024×1024
Product small in frame; GPT Image 2 fills the frame with a crisper hero shot.
Full-size original

Seedream 5 Flash's result for “Product Shot (white bg)” — Seedream 5 Flash33 s · 1024×1024
Clean product, smaller in frame than GPT Image 2's hero shot.
Full-size original

Movie Poster

The same prompt as Movie Poster.

Seedream 5 Pro's result for “Movie Poster” — Seedream 5 Pro128 s · 832×1248
Strong poster and readable title; on a par with Sunburst.
Full-size original

Seedream 5 Flash's result for “Movie Poster” — Seedream 5 Flash38 s · 832×1248
A strong poster with a bold, legible title; on a par with Sunburst.
Full-size original

Claymation Scene

The same prompt as Claymation Scene.

Seedream 5 Pro's result for “Claymation Scene” — Seedream 5 Pro85 s · 1312×736
Good clay texture, but a sparser scene (no lighthouse or storm) than GPT Image 2's.
Full-size original

Seedream 5 Flash's result for “Claymation Scene” — Seedream 5 Flash33 s · 1424×800
Good clay, lighthouse and storm are there; less dramatic light than GPT Image 2.
Full-size original

Studio Portrait

The same prompt as Studio Portrait.

Seedream 5 Pro's result for “Studio Portrait” — Seedream 5 Pro69 s · 896×1184
Polished studio light, on a par with Nano Banana 2.1.
Full-size original

Seedream 5 Flash's result for “Studio Portrait” — Seedream 5 Flash32 s · 864×1152
Natural, polished studio portrait, on a par with Nano Banana 2.1.
Full-size original

Three-Age Triptych

The same prompt as Three-Age Triptych.

Seedream 5 Pro's result for “Three-Age Triptych” — Seedream 5 Pro77 s · 864×1152
The oldest face is clearly a different person (different ethnicity), so the progression breaks.
Full-size original

Seedream 5 Flash's result for “Three-Age Triptych” — Seedream 5 Flash32 s · 864×1152
One recognisable person across the three ages — as good as GPT Image 2, where Pro failed.
Full-size original

What quality costs

Speed: GPT Image 2 had the slowest median in round 1, 97 seconds, against 51 to 68 seconds for the other four.
Ratios: GPT Image 2 draws only auto, 1:1, 16:9, 9:16, 4:3 and 3:4. For 2:3, GPT Image 2.5 Sunburst made the winning movie poster, and for 4:5, Nano Banana 2.1 made the winning studio portrait.
Clothing: the GPT models refuse some clothing, such as swimwear.
Price: at 2K, GPT Image 2 costs 42 credits an image, against 22 for Nano Banana 2.1, 28 for GPT Image 2.5 Flare and Sunburst, and 11 for Seedream 5 Flash.

Try them yourself

Every model in this test runs in Nodaro: on the Generate Image node for a new image, and on the Modify Image node for an edit. To run your own prompt on several models at once, connect one Text node to several image nodes and choose a different model in each. Choosing a model compares every image model in Nodaro.

Frequently asked questions

Which AI image model is best for editing photos?

In our test, GPT Image 2.5 Flare. It won two of the three edits and kept the pose, the framing and the product design on all three. GPT Image 2.5 Sunburst won the third edit, a background swap.

Which AI image model keeps a character consistent?

GPT Image 2 made the best reference grid, kept both people right on the cast sheet, and won the character board and the age progression. Nano Banana 2.1 won the labelled cast sheet and the expression sheet.

Which AI image model writes text best?

GPT Image 2 drew the cleanest infographic and GPT Image 2.5 Flare the best concert poster. Nano Banana Pro rendered the poster as a skewed photo of a print and returned a nearly empty infographic.

Is Nano Banana Pro still worth using?

Not for these jobs. It won none of the 18 prompts and got 8 weak marks in 11 prompts, including a broken grid layout, a dropped name label and changed outfits.

Is Seedream 5 Flash good for its price?

Yes, as a draft model. At 11 credits an image, it tied the earlier pick on 7 of 18 prompts and came close on 6, in about 45 seconds. It beat no earlier pick and mixed up the two people on the cast sheet.

How were the winners picked?

Claude, an AI model, looked at each prompt's results without the model names and chose the pick and any weak results. Each model made one image per prompt, so read a single result as a hint and the pattern across prompts as the finding.

Image Model Bake-off: seven AI image models on 18 real jobs

Frequently asked questions

Related

15 AI image models, one prompt: what each model drew

Choosing a model

GPT Image 2

GPT Image 2.5 Flare

Nano Banana 2.1

On this page