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
Eight image prompts put FLUX 2 and OpenAI’s GPT-5.4 Image 2 through tests of scene detail, composition, anatomy, counting, and palette control. The results expose a recurring trade-off between faithful framing and convincing, well-integrated detail.
OpenAI wins decisively: it leads 64.8 to 57.6, takes six of eight tasks, and earns a limited confidence verdict. FLUX 2 wins none; the remaining two tasks are ties. This is a consistent edge, not a single spectacular image carrying the score. GPT-5.4 Image 2 was stronger when the prompt demanded several elements working together. In the rainy kiosk scene, it integrated the vendor, street signs, scooter, bus, and puddle reflections into a more coherent night view. It also made the bracelet-making action clearer and rendered exactly seven distinct cups, where FLUX’s cups were marred by a bar-like artifact. FLUX’s clearest advantage was framing. It more often honored the requested 16:9 format, and its compositions were stronger in the tied alley and tram-stop comparisons. That matters: GPT-5.4 repeatedly drifted toward square images, including in the bookstore, campsite, and stoop tasks. Its wins came despite that miss, not because framing was unimportant. The overall distinction is execution: GPT-5.4 more reliably delivered legible details, plausible depth, and a convincing finished scene; FLUX more reliably respected the requested canvas and, in some cases, the intended composition. **Verdict: OpenAI’s GPT-5.4 Image 2 is the clear winner across these tests, while FLUX 2 remains the better bet when wide-format adherence is the priority.**
Rain kiosk reflections
A photoreal urban night scene centered on a glass-walled newspaper kiosk at the corner of Vesper Lane and 9th, after heavy rain: the kiosk windows reveal hanging magazines and a vendor inside, while also reflecting a red scooter, a striped awning across the street, and the blurred headlights of bus 218; puddles on the asphalt mirror the kiosk and neon pharmacy sign, and a chrome parking meter and a cracked convex security mirror add secondary reflections, with all glass transparency, mirror geometry, water reflections, and slight refraction rendered consistently from a three-quarter street-level viewpoint, 16:9.
OpenAI: GPT-6 Luna: Model B more fully integrates the street signs, vendor, red scooter, bus, parking meter, awning, and vivid puddle reflections into a cohesive rainy-night scene. Model A better matches the requested 16:9 format and clearly shows the cracked mirror, but its signage is less legible and several reflections feel less integrated. (Second judge pass, order swapped — scores are the average of both: Model B more fully captures the three-quarter street view and the requested details, especially the vendor | Anthropic: Claude Fable 5.1: Model B nails nearly every requested element—le…
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 has a clear centered vanishing point, but Model B has more convincing depth, consistent scale recession, and photorealistic detail. Model B is nearly square rather than 16:9, which slightly reduces its prompt adherence. (Second judge pass, order swapped — scores are the average of both: Model B presents a crisp, convincing one-point-perspective aisle with books and shelves receding naturally under warm overhead light, though its square format misses | Anthropic: Claude Fable 5.1: Model A delivers a true 16:9 one-point perspect…
Hands & anatomy
A natural candid photo of a person tying a friendship bracelet, both hands clearly visible with anatomically correct fingers and a plausible grip on the threads, soft daylight, shallow depth of field, 16:9.
OpenAI: GPT-6 Luna: Model B more clearly shows both hands plausibly working the threads on a friendship bracelet, with soft daylight and appealing shallow focus. Model A matches the requested wide framing better, but the bracelet-making action is less clear and the hands feel more awkwardly posed. (Second judge pass, order swapped — scores are the average of both: Model B more clearly shows the person actively tying a bracelet, with both hands visibly and plausibly manipulating the threads in a natural, softly lit | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 frame with b…
Exact counting
A clean overhead flat lay on white marble of EXACTLY seven identical espresso cups arranged in a single straight row, each clearly separate and individually countable, even soft lighting, 16:9.
OpenAI: GPT-6 Luna: Model B shows exactly seven clearly separate, individually countable cups, while Model A’s cups appear joined by a continuous bar-like artifact. Model A better matches the requested 16:9 format, but Model B’s distinct cups and clean overhead arrangement make it the stronger overall match. (Second judge pass, order swapped — scores are the average of both: Model B shows seven clear, attractive cups, but the square format misses the requested 16:9 aspect ratio. Model A matches the wide format and | Anthropic: Claude Fable 5.1: Both Model A and Model B render exactly seven id…
Restricted palette
A flat-vector illustration of a mountain campsite at dusk using ONLY four colors — cream, burnt orange, teal, and charcoal — with no gradients and no stray off-palette colors, balanced composition, 16:9.
OpenAI: GPT-6 Luna: Model B uses a cleaner, flatter four-color palette with no apparent gradients, while Model A has noticeable tonal variation and a gradient-like sky. Model A fits the requested 16:9 panorama and has a more balanced campsite, but Model B’s strong graphic style makes it the slightly better match overall despite its square format. (Second judge pass, order swapped — scores are the average of both: Model A has the requested 16:9 framing and a balanced, polished campsite scene, while Model B is square | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 frame and a…
Haint Blue alley palette
A meticulously color-restricted illustration of a rainy service alley behind a noodle shop, shown in graphic poster style, using only the named palette Haint Blue, Oxblood, Bone, and Tarnished Brass with no stray colors whatsoever: a delivery cyclist in a short cape pauses beside stacked crates, a blinking vacancy sign reflects in puddles, and sagging fire escapes frame the scene; balanced side-on composition, crisp edges, and controlled flat lighting that makes palette accuracy unmistakable, 16:9.
OpenAI: GPT-6 Luna: Model A follows the requested 16:9, side-on composition more closely and clearly shows the paused caped cyclist, crates, vacancy sign, puddles, and fire escapes. Model B makes the noodle-shop setting especially clear and is richly rendered, but its nearly square, deep-perspective composition departs from the requested framing and the cyclist appears to be riding onward. (Second judge pass, order swapped — scores are the average of both: Model B more clearly depicts the noodle shop, rain, cape-we | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 format, a s…
Golden-hour tram stop
A cinematic urban street scene at golden hour: a dented teal tram marked route 47 glides past the narrow stop outside 'Marek & Daughters Key Cutting' on Brzozowa Street, with a violin case leaning against the bench and three long shadows from pedestrians stretching diagonally across warm cobblestones; low sun skims the façades, windows flare amber, airborne dust glows, and the whole image should feel wistful and late-summer, photographed in realistic editorial style with a slightly low angle, layered depth, and 16:9 composition.
OpenAI: GPT-6 Luna: Model A better captures the requested 16:9, wistful golden-hour scene, with a dented teal route 47 tram, readable shop and street signs, and long diagonal pedestrian shadows. Model B has strong detail and includes a case by the bench, but its square framing and more frontal tram composition depart from the prompt. (Second judge pass, order swapped — scores are the average of both: Model B more precisely captures the dented teal route 47 tram, the named shop, the violin case, and three pedestrian | Anthropic: Claude Fable 5.1: Model A nails the 16:9 framing, bench and signa…
Midnight tailor stoop realism
A believable real-world street portrait outside 'Orsini Alterations' at midnight: a tailor in a charcoal wool coat sits on the shop stoop repairing a sequined silver dress under a single cool LED overhang while warm tungsten leaks from the doorway behind him; wet concrete, brushed aluminum shutters, varnished wood doorframe, human skin, frayed denim cuffs, silk thread, and cracked leather shoes must all show physically convincing material response, with soft shadow falloff, subtle bounce light from the street, and photorealistic documentary style in a restrained 16:9 composition.
OpenAI: GPT-6 Luna: Model B feels more like a restrained documentary photograph, with stronger material detail and clearly rendered shop signage; Model A better matches the requested 16:9 framing and shows the scene more fully, but its signage is cropped and the image feels more staged. Model B’s tighter, square composition is a notable departure from the prompt, though its overall realism and finish give it the edge. (Second judge pass, order swapped — scores are the average of both: Model B delivers a convincing, | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 documentary…
Matchup powered by OpenRouter.