Head to head: FLUX.2 [klein] 9B Base LoRA vs Fibo Gen 1.5 Text to Image

This one wasn’t a rout, but it also wasn’t especially ambiguous. FLUX.2 [klein] 9B Base LoRA shows flashes of tighter object control, yet Fibo Gen 1.5 Text to Image wins the matchup by being more reliable on the prompts that actually punish weak adherence.

Fibo Gen 1.5 Text to Image takes this head-to-head on both the scoreboard and the substance: 63.0 to 56.1 overall, with an 86% confidence verdict and a 5–2 task lead plus one tie. That’s not a photo finish. It’s a clear win built on broader competence across the harder prompt-following tests. The pattern is straightforward. Fibo was better when the prompt demanded specific subject presence, spatial logic, and scene-level coherence. It won **Macro moth proboscis** because it actually gave us the moth feeding from the flower, while FLUX simply missed the central subject. It won **Reflections & glass** by delivering the apple reflection in both chrome and water, where FLUX looked good but failed the assignment. And in **Golden-hour heron marsh** and **Attribute binding**, Fibo was the model more likely to keep the exact bird, the rusted half-submerged bicycle, and the single duck on the correct object instead of drifting into near-miss territory. FLUX’s case rests on two narrower strengths. It beat Fibo on **Bound colors on field kit**, where its still-life layout was cleaner and its color binding more exact, and on **Exact counting**, though that win comes with an asterisk because both models failed the core requirement of showing exactly seven cups. That result says less about counting excellence than about FLUX being the less wrong image. The **Kingfisher breach sprint** tie reinforces the broader story: FLUX can absolutely produce polished, compelling frames, but it doesn’t consistently convert that polish into decisive prompt obedience. What hurts FLUX most here is not image quality in the abstract; it’s avoidable prompt failure. Missing the moth entirely, adding an extra duck, and falling short on key reflection requirements are the kinds of errors that sink trust fast. Fibo isn’t flawless either—there’s some stylization, and one judge pass even split on the restricted-palette task—but across the full set it was simply the steadier model when details mattered. **Final call: Fibo Gen 1.5 Text to Image is the winner. FLUX.2 [klein] 9B Base LoRA has moments of tighter control, but Fibo is the more dependable generator overall, and the 86% confidence figure matches what the task-by-task results already show.**

Macro moth proboscis

An extreme macro close-up photograph of a comet moth feeding from a single midnight-blue columbine blossom, captured with a true macro lens so the curled proboscis and pollen grains are razor sharp while the background falls into creamy circular bokeh, shallow depth of field, off-center framing, luminous dusk garden light, ultra-detailed natural color.

FLUX.2 [klein] 9B Base LoRA:

Fibo Gen 1.5 Text to Image:

Model A is a beautiful macro flower study, but it misses the central subject entirely because there is no comet moth feeding from the blossom. Model B clearly depicts a moth with a curled proboscis on a midnight-blue columbine in strong macro style, with better prompt adherence and a compelling off-center composition, despite some stylization and anatomical inaccuracy. (Second judge pass, order swapped — scores are the average of both: Model B clearly includes a moth feeding from a midnight-blue columbine with strong macro styling, visible curled proboscis, pollen detail, and pleasing dusk bokeh, though the insect looks somewhat stylized rather than fully photographic. Model A is a beautiful macro flower image, but it omits the comet moth entirely and therefore misses the core subject of the prompt.)

Bound colors on field kit

A meticulous studio still life in hyper-detailed natural-history illustration style showing five expedition objects on dark slate: a matte cobalt-blue ceramic canteen, a glossy saffron-yellow silk map pouch, a brushed copper-orange titanium compass, a deep moss-green wool falconry glove, and a pale ivory bone-handled specimen knife, each object clearly separated so every color stays bound to its own material with no color bleeding, soft neutral side lighting and precise shadows.

FLUX.2 [klein] 9B Base LoRA:

Fibo Gen 1.5 Text to Image:

Model A adheres more closely to the prompt by presenting five clearly separated objects on dark slate with the requested color assignments and minimal color bleeding; its layout also reads more like a meticulous studio still life. Model B is attractive and painterly, but the canteen shape/material are less accurate, the objects are less clearly separated, and the knife handle reads more tan than pale ivory bone. (Second judge pass, order swapped — scores are the average of both: Model A adheres more closely to the prompt by presenting five clearly separated expedition objects on dark slate with cleaner color binding and a more precise natural-history still-life arrangement. Model B is attractive and cohesive, but the canteen shape/material read less convincingly, the glove is bulkier and overlaps the intended illustration-like clarity, and the knife/compass styling is slightly less exact.)

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.

FLUX.2 [klein] 9B Base LoRA:

Fibo Gen 1.5 Text to Image:

Model A is closer to the prompt with an overhead flat lay on white marble and a single straight row of clearly separated identical espresso cups, but it shows only five cups instead of exactly seven. Model B also shows only five cups and they appear empty rather than espresso-filled, though its row is neat; overall Model A adheres better and looks more polished. (Second judge pass, order swapped — scores are the average of both: Model B misses the key requirement of EXACTLY seven cups, showing only five, though the overhead marble flat lay is clean and technically solid. Model A also shows only five cups, but it better conveys espresso cups in a straighter, more deliberate row with stronger visual appeal and clearer separation.)

Reflections & glass

A photorealistic close-up of a chrome teapot on a polished dark marble counter next to a glass of water; the window and a red apple on the counter must be correctly reflected in BOTH the chrome and the water's surface, dramatic side light, 16:9.

FLUX.2 [klein] 9B Base LoRA:

Fibo Gen 1.5 Text to Image:

Model B better satisfies the key reflection requirement: the red apple is clearly reflected in both the chrome teapot and the water, and the lighting/composition feel more dramatic and polished. Model A is attractive and photorealistic, but the apple reflection is not convincingly present in the teapot or water, making it less faithful to the prompt. (Second judge pass, order swapped — scores are the average of both: Model B better matches the prompt with a convincing chrome teapot, polished dark marble, dramatic side light, and visible reflections of both the window and red apple in the chrome and the water. Model A is attractive and well composed, but the water reflection/refraction of the apple is less clearly correct and the overall close-up prompt match is weaker.)

Golden-hour heron marsh

A painterly wildlife scene of a black-crowned night heron standing on a half-submerged rusted bicycle frame in the Narev fen at golden hour, low amber sunlight skimming the reeds and turning the mist peach-gold, long shadows across rippled water, a calm melancholic mood, dragonflies catching rim light, cinematic composition in 16:9.

FLUX.2 [klein] 9B Base LoRA:

Fibo Gen 1.5 Text to Image:

Model B adheres more closely to the prompt by depicting a black-crowned night heron on a clearly rusted, half-submerged bicycle in a misty golden-hour fen with stronger melancholic mood and better dragonfly emphasis. Model A has a more cinematic 16:9 composition and lovely light, but the bird reads more like a generic heron and the bicycle appears less rusted and less convincingly half-submerged. (Second judge pass, order swapped — scores are the average of both: Model B better matches the prompt’s specific black-crowned night heron, rusted bicycle frame, misty marsh mood, and painterly cinematic framing, though the bicycle is more fully visible in Model A. Model A has attractive golden-hour lighting and composition, but it departs more from the prompt’s exact subject and mood, making Model B the stronger overall fit.)

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.

FLUX.2 [klein] 9B Base LoRA:

Fibo Gen 1.5 Text to Image:

Model A better matches the requested mountain campsite at dusk in a 16:9 flat-vector format with a strong balanced composition and an appealing restricted-palette look, though it appears to introduce a few extra shades beyond the strict four colors. Model B is attractive and cleaner in shape language, but it is square rather than 16:9 and reads more like night than dusk, which hurts prompt adherence despite solid palette discipline. (Second judge pass, order swapped — scores are the average of both: Model B adheres much more closely to the prompt with a clear flat-vector dusk campsite, balanced 16:9 composition, and a palette that stays very close to the specified four colors. Model A is visually appealing, but it departs from the dusk requirement, introduces more than four apparent shades/colors, and feels less restricted-palette compliant despite strong composition.)

Attribute binding

A photorealistic tabletop scene: a matte green cube to the LEFT of a glossy red sphere, a small blue cylinder BEHIND both, and a single yellow rubber duck sitting ON TOP of the cube. Colors must stay on the correct objects. Soft studio light, 16:9.

FLUX.2 [klein] 9B Base LoRA:

Fibo Gen 1.5 Text to Image:

Model A gets the main object colors and relative placement mostly right, but it violates the prompt by adding a second yellow rubber duck on top of the red sphere. Model B matches the requested single duck on top of the green cube, keeps the blue cylinder behind and the red sphere to the right, and also has a cleaner, more polished composition and rendering. (Second judge pass, order swapped — scores are the average of both: Model B follows the prompt closely: the matte green cube is left of the glossy red sphere, the blue cylinder sits behind them, and there is a single yellow rubber duck on top of the cube; it is also cleanly lit and visually balanced. Model A has strong rendering and composition, but it violates the prompt by adding a second duck on top of the red sphere, breaking the single-object attribute binding requirement.)

Kingfisher breach sprint

A high-energy realist image of a pied kingfisher exploding upward from a narrow canyon stream with a silver loach in its beak, wings flared asymmetrically, droplets frozen in a curved spray while the water below streaks with motion blur, tense dynamic diagonals, storm-cleared daylight and crisp contrast, shot as if captured at the peak split-second of action in 16:9.

FLUX.2 [klein] 9B Base LoRA:

Fibo Gen 1.5 Text to Image:

Model B better matches the prompt’s peak-action realism with a clearer pied kingfisher pattern, stronger dynamic diagonals, and a more convincing curved spray of frozen droplets in a narrow canyon stream. Model A is attractive and technically solid, but its frontal symmetry and less specific species rendering weaken adherence to the requested asymmetrical, high-energy breach moment. (Second judge pass, order swapped — scores are the average of both: Model A better matches the prompt’s peak-action realism with a pied kingfisher breaching from a narrow canyon stream, a fish in its beak, strong dynamic diagonals, and convincing motion in the water. Model B is striking and detailed, but its centered, more symmetrical pose and highly stylized circular droplet arc feel less faithful to the requested asymmetry and split-second naturalism.)

Matchup powered by OpenRouter.