Head to head: FLUX 2 vs OpenAI: GPT-5.4 Image 2
FLUX 2 vs OpenAI: GPT-5.4 Image 2
By Ryan Merket · Published
A head-to-head across text-heavy posters, laboratory scenes, and spatial compositions tests how each image model handles exact instructions as well as visual polish. The comparison turns on a recurring tradeoff: fidelity to the requested format versus richer detail and presentation.
The aggregate scores are **60.5 for FLUX 2** and **64.2 for OpenAI: GPT-5.4 Image 2**. Despite that numerical gap, the saved result is a tie: FLUX 2 won three tasks, OpenAI won four, and one was tied. FLUX 2’s wins came on the neon isotope poster, the negation prompt, and the cryo sample sprint. Across those tasks, it was particularly effective when the brief depended on a wide frame, a clean specified layout, or a clearly readable action. OpenAI took the multi-line text, failed experiment dread, spatial layout, and bound materials bench set tasks. Its wins show up in crisp text and detailed object presentation; the results were less uniform when the requested framing or composition was at odds with its chosen presentation. The perspective-and-scale task ended level, underscoring why the aggregate lead does not settle this matchup. **Final call: Too close to call (not statistical).** This assessment applies to the saved task sample, not to image-model capability in general.
Neon isotope poster
A crisp graphic poster mounted outside a university radiochemistry suite, designed in bold modern typography and easy to read from a distance, with the exact text: "ISOTOPE NIGHT" on the top line, "LAB 3E" beneath it, and "08 NOV 21:40" on the bottom line, fluorescent lime and white lettering on a matte charcoal background with a small magenta trefoil icon, clean centered layout, realistic indoor hallway lighting, straight-on composition, 16:9
OpenAI: GPT-6 Luna: Model A closely follows the requested 16:9 straight-on hallway poster, with the exact three lines clearly centered and a small magenta trefoil. Model B is bold and polished, but its near-square framing, oversized middle line, and large icon interrupt the requested clean three-line layout. (Second judge pass, order swapped — scores are the average of both: Model B has exceptionally bold, distance-readable typography, exact text, a prominent magenta trefoil, and clear radiochemistry-suite context. | Anthropic: Claude Fable 5.1: Model A delivers the exact text in the correct…
Legible multi-line text
A minimalist event poster with three lines of crisp, correctly-spelled text stacked and centered: 'NIGHT MARKET' large on top, 'Fridays · 6–11pm' in the middle, 'Riverside Pier 4' at the bottom, on a deep navy background, clean sans-serif, subtle grain.
OpenAI: GPT-6 Luna: Model B presents all three lines crisply and correctly, with stronger visual hierarchy and a more polished, subtly grained poster aesthetic. Model A is also clear and faithful, but its layout feels less balanced and distinctive. (Second judge pass, order swapped — scores are the average of both: Model B delivers the requested centered three-line text with crisp, correct wording, a deep navy grainy background, and a polished, spacious poster composition. Model A also renders the wording accuratel | Anthropic: Claude Fable 5.1: Both render all three lines correctly spelled,…
Failed experiment dread
A film-noir black-and-white close portrait of a lone biochemist in a midnight laboratory, staring at a cracked petri dish in her gloved hand with unmistakable dawning dread and disbelief, brows tightened unevenly, lower lip slightly parted, eyes glossy and fixed, shoulders rigid, a single hard sidelight from a desk lamp carving deep shadows across her face, blurred glassware and rain-streaked windows behind her, intimate 85mm composition
OpenAI: GPT-6 Luna: Model A more clearly conveys dawning dread and shows an unmistakably shattered petri dish, with the lamp-lit, rain-streaked laboratory reinforcing the noir setting. Model B has a tighter portrait and strong lighting, but the expression reads more subdued and the dish’s damage is less immediately clear. (Second judge pass, order swapped — scores are the average of both: Model A makes the cracked dish and dawning dread especially unmistakable, with a visibly startled expression that closely matche | Anthropic: Claude Fable 5.1: Model B nails the film-noir brief with a hard d…
Negation
A cozy reading nook with an armchair, a stack of books, and a mug of tea by a window — with absolutely NO plants, NO lamps, and NO artwork or picture frames anywhere in the frame. Warm afternoon light, 16:9.
OpenAI: GPT-6 Luna: Model A matches the requested 16:9 framing and includes the cozy chair, books, tea, window, and warm light without any forbidden plants, lamps, or artwork. Model B is especially inviting, but its square framing misses the requested aspect ratio; neither image visibly violates the exclusions. (Second judge pass, order swapped — scores are the average of both: Model B creates a polished, cozy nook with all the requested objects and none of the forbidden items, but it misses the requested 16:9 form | Anthropic: Claude Fable 5.1: Both Model A and Model B correctly omit plants,…
Spatial layout
A clean isometric illustration of a bedroom: a bed against the LEFT wall, a round rug centered on the floor, a desk under the WINDOW on the back wall, and a floor lamp in the FRONT-RIGHT corner. Flat-vector style, consistent perspective.
OpenAI: GPT-6 Luna: Model A follows the requested flat-vector style and places the bed, centered rug, desk beneath the back-wall window, and floor lamp in the specified positions. Model B is more richly detailed, but its rug is partly tucked under the bed and its rendered look is less faithful to the requested flat-vector style. (Second judge pass, order swapped — scores are the average of both: Model A better matches the requested flat-vector style while keeping the bed, centered rug, desk beneath the back-wall wi | Anthropic: Claude Fable 5.1: Both place the bed against the left wall and th…
Bound materials bench set
A meticulously arranged product-style still life on a materials science lab bench: a matte cobalt-blue ceramic cube, a translucent amber glass spiral tube, a brushed silver titanium hexagonal plate, a deep crimson velvet drawstring pouch, and an opaque mint-green silicone gasket, each object clearly separated so its own color and material are unmistakable and do not transfer to the others, neutral gray background, soft diffused studio lighting, ultra-clean sharp focus, 16:9
OpenAI: GPT-6 Luna: Model B depicts all five requested objects with especially clear material cues and a convincing amber spiral tube, while Model A has a stray extra-looking metal slab beneath the tube and a less distinct tube form. Model A is more minimal and matches the wide framing, but Model B’s stronger object detail and arrangement outweigh its distracting lab props and square crop. (Second judge pass, order swapped — scores are the average of both: Model B presents all five requested materials clearly and d | Anthropic: Claude Fable 5.1: Model A correctly delivers 16:9 and a neutral g…
Cryo sample sprint
A high-speed editorial photograph of a glaciologist in a polar research laboratory sprinting across a wet steel floor while catching a sliding cryogenic sample canister labeled Q-17, one boot splashing through meltwater, coat tails whipping sideways, loose frost crystals and droplets frozen in the air, motion blur trailing from the outstretched arm but the face and canister tack sharp, blue emergency strip lights reflecting off stainless benches, dramatic low-angle composition, cinematic dynamic energy, 16:9
OpenAI: GPT-6 Luna: Model A better matches the requested wide, low-angle sprint with a sharp face and legible Q-17 canister, dramatic splash, and visible motion blur. Model B has compelling close-up energy, but its square framing, blurred and partly cropped canister, and less readable running action weaken adherence. (Second judge pass, order swapped — scores are the average of both: Model A better captures the requested wide, low-angle sprint with coat tails whipping, airborne droplets, blue strip lighting, and a | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 low-angle la…
Perspective & scale
A photorealistic one-point-perspective shot down a long empty library aisle, shelves receding to a single vanishing point, books getting consistently smaller with distance and nothing warped or wrongly sized, warm overhead light, 16:9.
OpenAI: GPT-6 Luna: Model A matches the requested 16:9 framing and presents a long, empty aisle with a clear central vanishing point and warm overhead lighting. Model B is sharper and richly detailed, but its square framing misses the requested aspect ratio despite its strong perspective. (Second judge pass, order swapped — scores are the average of both: Model B has the clearer, more convincing one-point perspective, with crisp shelves and books that recede consistently, though its square format misses the request | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 frame with…
Matchup powered by OpenRouter.