Head to head: FLUX.1 [dev] with LoRAs vs Rundiffusion Photo Flux

This is a close matchup between a more geometry-disciplined base model setup and a more dependable photo-tuned finisher. The edge goes to the model that misses fewer prompt specifics in real-world scenes, but not by enough to call it a rout.

On aggregate, **Rundiffusion Photo Flux takes this one, 61.2 to 57.8**, with a **77% confidence lean**. That’s a real win, but not a blowout: the task count is only **4 wins to 2, with 2 ties**, and several categories were close enough to expose each model’s distinct strengths rather than establish total dominance. The pattern is pretty clear. **Rundiffusion wins where prompt fidelity has to survive photographic realism.** It was better on **hands & anatomy**, where FLUX.1’s bracelet-tying scene looked polished but not truly in-progress; better on the **rainy repair kiosk**, where it nailed more of the walnut-counter, cobalt-console, mixed-lighting brief; better on **legible multi-line text**, where neither model was perfect but FLUX.1 drifted harder in typography; and better on **attribute binding**, where FLUX.1 simply broke the object relationships more severely. In plain English: B was more trustworthy when the prompt asked for specific things in specific places inside a believable photo. FLUX.1 [dev] with LoRAs, though, was not outclassed. It won **perspective & scale** outright with cleaner one-point geometry, stronger symmetry, and more convincing long-aisle recession. It also took **exactly eleven gadgets**, less because it fully solved the brief than because it produced the cleaner, more coherent top-down product image while B wandered into more object confusion. And the two ties tell an important story: on **server aisle perspective** and **watchmaker hands**, the judges split because each model exposed the other’s weakness. FLUX.1 often looked more composed and architecturally disciplined; Rundiffusion often felt more naturalistic and task-plausible. That’s the verdict in one sentence: **FLUX.1 is the better framer, Rundiffusion is the better follower.** If your prompts depend on strict perspective, layout discipline, or polished visual structure, FLUX.1 with LoRAs still has a case. But across this set, Rundiffusion was more likely to preserve the actual assignment instead of substituting a nice-looking approximation. **Final call: Rundiffusion Photo Flux wins on points. It’s a lean, not a landslide, but it was the more reliable model where adherence, anatomy, text, and object binding actually mattered.**

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.

FLUX.1 [dev] with LoRAs:

Rundiffusion Photo Flux:

Model B better matches the prompt by showing a more believable friendship bracelet in progress, both hands clearly visible, and a more plausible finger grip with stronger natural-photo composition. Model A is attractive and sharp, but the hand anatomy and threading action feel less convincing, and it reads more like holding a looped knot than actively tying a bracelet. (Second judge pass, order swapped — scores are the average of both: Model B better matches the prompt with a more believable candid photo of someone actively tying a friendship bracelet, both hands visible, natural daylight, and mostly convincing finger anatomy and grip. Model A is visually clean, but the hand anatomy and pose feel more artificial, and the thread arrangement reads less plausibly as an in-progress friendship bracelet.)

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.

FLUX.1 [dev] with LoRAs:

Rundiffusion Photo Flux:

Model A better matches the requested one-point perspective with a cleaner central vanishing point, more consistent shelf scaling, and a stronger long-aisle feel under warm overhead lighting. Model B is also good and photorealistic, but the aisle appears slightly less precise in perspective and scale consistency, with a softer, less striking composition. (Second judge pass, order swapped — scores are the average of both: Model A better matches the requested one-point perspective, with cleaner symmetry, a stronger central vanishing point, and more consistent shelf/book scaling down the aisle. Model B is still good and photorealistic, but its perspective feels slightly less precise and the aisle appears a bit less rigorously centered and scaled.)

Server aisle perspective

Ultra-detailed architectural visualization, 16:9: inside a narrow underground micro-data-center built beneath a coastal tram station, viewed from one end of a long service aisle with a single strong vanishing point; two rows of server racks recede accurately into the distance, overhead cable trays, floor tiles, and yellow safety lines all aligning consistently, while scale reads correctly from a full-size wheeled tool cart in the foreground to compact sensor boxes mounted farther away; include a glass inspection window at the far wall showing a maintenance drone outside for scale reference, neutral industrial lighting, crisp geometry, no warped shelves or mismatched proportions.

FLUX.1 [dev] with LoRAs:

Rundiffusion Photo Flux:

Model B better matches the requested underground micro-data-center aisle with a stronger single vanishing point, more convincing rack recession, and a clearer industrial service-space feel. Model A is clean and attractive, but the bright ocean-like opening replaces the specified far-wall inspection window and drone, making it less faithful to the prompt despite its polished look. (Second judge pass, order swapped — scores are the average of both: Model A better matches the requested architectural-visualization aisle perspective with a stronger vanishing point, cleaner geometric alignment, and a more convincing full-size wheeled tool cart in the foreground. Model B is solid and detailed, but it misses key prompt specifics with a person at the far wall instead of a maintenance drone outside the inspection window, and its foreground cart reads less like the specified tool cart.)

Rainy repair kiosk

Photorealistic editorial-style 16:9 image of a tiny late-night electronics repair kiosk under a translucent bus-stop canopy during a drizzle, a technician in a charcoal fleece replacing the battery in a scratched cobalt-blue handheld game console on a walnut counter; physically convincing mixed lighting from a cool white LED desk lamp, a warm sodium streetlight reflected in wet pavement, and faint neon from a pharmacy sign, with believable shadows and reflections on brushed aluminum tools, matte black rubber cable insulation, varnished wood grain, damp skin, clear raindrops on plastic, and slightly wrinkled fabric sleeves; shot at eye level with a 50mm lens, natural color, real-world materials only.

FLUX.1 [dev] with LoRAs:

Rundiffusion Photo Flux:

Model B adheres more closely to the prompt with a tiny late-night repair-kiosk feel, a cobalt-blue handheld game console on a walnut counter, and more believable mixed warm/cool rainy-night lighting. Model A is attractive and technically solid, but it reads more like someone using a handheld under shelter than actively replacing a battery, and it misses several specified material/details cues such as tools and the translucent bus-stop canopy context. (Second judge pass, order swapped — scores are the average of both: Model B better matches the prompt’s tiny rainy repair-kiosk mood, eye-level editorial framing, walnut counter, and cobalt-blue handheld console under mixed lighting, though it underplays the visible repair action and tool detail. Model A is attractive and technically solid, but it reads more like someone using a clamshell handheld than replacing a battery, and it misses key prompt specifics such as the scratched cobalt-blue console and believable repair setup.)

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.

FLUX.1 [dev] with LoRAs:

Rundiffusion Photo Flux:

Model B better matches the requested minimalist poster style with a clean sans-serif look, centered stacked layout, and stronger legibility overall. Model A has good balance and a deep navy background, but it uses a serif typeface and renders the middle line incorrectly with a '>' instead of the requested centered dot, making it less faithful to the prompt. (Second judge pass, order swapped — scores are the average of both: Model B adheres more closely to the requested minimalist poster: the text is centered, highly legible, and mostly correct, with a clean sans-serif feel and balanced hierarchy, though it uses a '>' instead of the requested dot separator and adds a divider line. Model A has a pleasing navy palette but deviates more in typography by using a serif style, smaller lower lines, and the same incorrect separator, making it less aligned with the prompt and less effective for crisp multi-line text.)

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.1 [dev] with LoRAs:

Rundiffusion Photo Flux:

Model B follows the requested attribute bindings much more closely: the matte green cube is left of the glossy red sphere, the yellow duck is on top of the cube, and the small blue cylinder is behind/partly occluded between them. Model A is more visually balanced, but it introduces a blue cube instead of a blue cylinder and places the duck on the wrong object, which breaks the prompt more severely. (Second judge pass, order swapped — scores are the average of both: Model B follows the key bindings better: the matte green cube is on the left, the glossy red sphere is to its right, and the yellow duck is on top of the cube, though the blue cylinder is awkwardly inserted between cube and sphere rather than clearly behind both. Model A breaks the object bindings more severely by placing the duck on a blue cube instead of the green cube and introducing an extra cube, even though its lighting and rendering are clean.)

Watchmaker hands

Naturalistic magazine-style portrait of a young wearable-tech watchmaker seated at a small bench, clearly visible from head to waist, assembling a custom smartwatch with both hands in frame; her anatomy must be correct and relaxed, with natural shoulders, forearms, wrists, and two believable hands each showing five well-formed fingers in plausible positions as she holds a tiny screwdriver and the watch body; include a linen apron, a loupe pushed up on her forehead, and trays of micro-components around her, soft north-window daylight with subtle fill, shallow depth of field, realistic skin texture and proportions, 16:9.

FLUX.1 [dev] with LoRAs:

Rundiffusion Photo Flux:

Model B adheres more closely to the prompt by showing both hands clearly engaged with a tiny tool and watch body in a believable, relaxed pose, with a stronger head-to-waist portrait feel. Model A is attractive and technically polished, but it misses key prompt details like the loupe and smartwatch-specific context, and the large microscope dominates the composition more than the watchmaking action. (Second judge pass, order swapped — scores are the average of both: Model A adheres more closely to the prompt with a clearer head-to-waist portrait, visible smartwatch assembly context, apron, trays of components, and stronger magazine-style composition; Model B is naturalistic but misses key details like the loupe and reads more like generic bench work, with less convincing hand/task clarity. Model B’s hands are reasonably believable, but Model A delivers the fuller scene and more polished overall image despite also lacking the forehead loupe and introducing a slightly distracting wristwatch.)

Exactly eleven gadgets

Clean top-down product photo on a pale graphite tabletop, 16:9, showing EXACTLY 11 distinct retro-futurist desk gadgets, all fully visible and individually countable with small gaps between them: 1 transparent amber pager, 1 mint-green foldable keyboard, 1 silver voice recorder, 1 cobalt wireless mouse, 1 ivory e-ink notepad, 1 smoked-glass smart ring case, 1 brass USB hub shaped like a hexagon, 1 coral portable SSD, 1 black stylus, 1 teal mini projector, and 1 white clip-on webcam; soft diffused studio lighting, realistic shadows, no duplicates, no extra accessories, no partial objects, no text labels.

FLUX.1 [dev] with LoRAs:

Rundiffusion Photo Flux:

Model A has a cleaner top-down product-photo layout and stronger visual polish, but it misses the requirement of exactly 11 distinct gadgets and several specified identities/colors are off. Model B also fails prompt adherence with multiple wrong objects and extras, and its composition is less consistent and less product-photo-like overall. (Second judge pass, order swapped — scores are the average of both: Model A is much closer to the requested clean top-down product layout with 11 clearly separated gadgets and stronger overall polish, though it still misses several specified identities and includes some text-like markings. Model B also has 11 items, but multiple objects are the wrong type or ambiguous, including a phone-like object and a tiny dongle-like extra instead of the requested gadgets, making adherence notably weaker.)

Matchup powered by OpenRouter.