FLUX image models compared: real photo tests
Can a model keep the same person, follow a pose and make hands touch the right objects? We tested FLUX 3 Image, FLUX.2 Pro, Max, Flex and both Klein sizes on harder photo briefs, alongside our original teapot comparison.
By PhotoGenerAI. Published . Images generated October 14 and 20, 2026.
Harder photo tests: identity, poses and hands
A simple teapot leaves a lot open. These briefs make the models keep track of specific people, clothing and physical contact. Each model gets the same task twice. Switch trials to inspect every result.
A third hand is easier to miss than a bad portrait
Both images look plausible at a glance. Count the woman's hands. These are the second planned trials of the same two-reference brief, with the same inputs and prompt. The full set, including the first trials, follows below.

The woman has three hands: one at the mug and two at the table.

Four hands, one shared mug, and both free hands resting separately on the table.
The shared mug exposed an extra hand in Klein 4B and an extra mug in Flex. Pro and Max kept recognizable faces, but the woman's free hand did not rest separately on the table in either trial. FLUX 3 met our five checks in its one returned image; its first request failed at the API. Klein 9B's second image also had the requested hand arrangement, though its downturned faces made identity harder to judge.
The courier separated physical contact from instruction following. Klein 4B and Flex each produced a disconnected umbrella in trial 1. FLUX 3 used the correct own-left umbrella and own-right handlebar assignment in trial 1, then reversed it in trial 2. The other five models missed that assignment in both trials.
The pose transfer did not expose a clear gap. All eight returned images followed the main pose and clothing brief. Profile faces left identity uncertain; the other four trials returned API errors. These are specific cases to inspect, not evidence that one model always wins.
Two people, one shared mug
Combine two portraits without swapping faces or clothes, then make both people touch the same mug. Each free hand must stay visible on the table.
Task prompt and checks
Create a realistic candid waist-up photograph of the two adults from image 1 and image 2 sitting together at an outdoor cafe table. Put the man from image 1 on the left side of the picture and the woman from image 2 on the right. Preserve each person's own face, hairstyle, skin tone and clothing: the man wears his gray T-shirt, the woman wears her blue denim jacket over her cream T-shirt. The man is passing one white ceramic mug to the woman. Each person touches this same mug with one hand, while each person's other hand rests separately on the table. Show both complete faces and all four hands. There must be exactly two people and one mug. Use natural daylight, realistic skin texture and believable contact shadows.
- Exactly two distinct people, man on picture-left and woman on picture-right
- Both reference faces and hairstyles remain recognizable
- Each person keeps their own shirt and jacket
- Exactly one white mug is touched by one hand from each person
- Both free hands rest separately on the table, with all four hands visible and plausible
Trial 1 of 2. Select an image for the original file.

Klein 4B
$0.016 recordedThe two main people keep their clothes and share one mug, with their free hands on the table. An extra background person breaks the exactly-two-people instruction. Facial preservation is uncertain.
Our five checks
- MissedExactly two distinct people, man on picture-left and woman on picture-right
- UnclearBoth reference faces and hairstyles remain recognizable
- MetEach person keeps their own shirt and jacket
- MetExactly one white mug is touched by one hand from each person
- MetBoth free hands rest separately on the table, with all four hands visible and plausible

Klein 9B
$0.019 recordedOne mug and four main-subject hands are visible, with both free hands on the table. A partly visible background person breaks the count. Downturned faces leave identity uncertain.
Our five checks
- MissedExactly two distinct people, man on picture-left and woman on picture-right
- UnclearBoth reference faces and hairstyles remain recognizable
- MetEach person keeps their own shirt and jacket
- MetExactly one white mug is touched by one hand from each person
- MetBoth free hands rest separately on the table, with all four hands visible and plausible

Pro
$0.060 recordedBoth faces and outfits remain recognizable, with one shared mug. The woman's other hand stays near the mug instead of resting separately on the table.
Our five checks
- MetExactly two distinct people, man on picture-left and woman on picture-right
- MetBoth reference faces and hairstyles remain recognizable
- MetEach person keeps their own shirt and jacket
- MetExactly one white mug is touched by one hand from each person
- MissedBoth free hands rest separately on the table, with all four hands visible and plausible

Max
$0.130 recordedBoth people keep recognizable faces and their own clothes. They share one mug, but the woman uses both hands on the mug instead of placing her free hand separately on the table.
Our five checks
- MetExactly two distinct people, man on picture-left and woman on picture-right
- MetBoth reference faces and hairstyles remain recognizable
- MetEach person keeps their own shirt and jacket
- MetExactly one white mug is touched by one hand from each person
- MissedBoth free hands rest separately on the table, with all four hands visible and plausible
The provider returned an error. No image was received for this trial.
Flex
HTTP 503 from the generation API. No image was returned, so we make no visual judgment. No charge receipt was returned.
The provider returned an error. No image was received for this trial.
FLUX 3 Image
HTTP 503 from the generation API. No image was returned, so we make no visual judgment. No charge receipt was returned.
One person holding two objects
Generate a courier whose own left hand grips an umbrella and whose own right hand holds a bicycle handlebar. A floating umbrella does not meet the brief.
Task prompt and checks
A realistic full-body street photograph of one adult bicycle courier standing still beside a bicycle outside a railway station after rain. The courier wears a teal rain jacket, charcoal trousers, brown boots, a navy bicycle helmet and an orange shoulder bag. The courier's own left hand grips the handle of one open transparent umbrella above their head. Their own right hand holds the bicycle handlebar. Both complete hands must be visible, physically attached to the correct arms and clearly in contact with their assigned object. The courier is standing on the pavement, not riding. Show exactly one person, one umbrella and one bicycle. Natural overcast daylight, realistic skin and wet pavement reflections, no logos or written text.
- Exactly one person, one umbrella and one bicycle
- Own left hand visibly grips the umbrella handle
- Own right hand visibly holds the bicycle handlebar
- Both complete hands are visible, attached and anatomically plausible
- Teal jacket, charcoal trousers, brown boots, navy helmet and orange bag are present
Trial 1 of 2. Select an image for the original file.

Klein 4B
$0.014 recordedThe raised hand holds a detached curved handle while the umbrella floats above it. The umbrella is on the courier’s own right and the bicycle on their own left, reversing the assigned roles.
Our five checks
- MetExactly one person, one umbrella and one bicycle
- MissedOwn left hand visibly grips the umbrella handle
- MissedOwn right hand visibly holds the bicycle handlebar
- MetBoth complete hands are visible, attached and anatomically plausible
- MetTeal jacket, charcoal trousers, brown boots, navy helmet and orange bag are present

Klein 9B
$0.015 recordedThe umbrella grip connects to its shaft, but the courier’s own right hand holds it and their own left holds the bike. Extra people appear in the background. The requested clothing colors are present.
Our five checks
- MissedExactly one person, one umbrella and one bicycle
- MissedOwn left hand visibly grips the umbrella handle
- MissedOwn right hand visibly holds the bicycle handlebar
- MetBoth complete hands are visible, attached and anatomically plausible
- MetTeal jacket, charcoal trousers, brown boots, navy helmet and orange bag are present

Pro
$0.030 recordedThe umbrella and bicycle have believable hand contact and the clothing matches. The courier’s own right hand holds the umbrella and their own left holds the bike, reversing the brief.
Our five checks
- MetExactly one person, one umbrella and one bicycle
- MissedOwn left hand visibly grips the umbrella handle
- MissedOwn right hand visibly holds the bicycle handlebar
- MetBoth complete hands are visible, attached and anatomically plausible
- MetTeal jacket, charcoal trousers, brown boots, navy helmet and orange bag are present

Max
$0.070 recordedThe umbrella is held in the courier’s own right hand. Their own left rests on the bicycle saddle instead of the handlebar. Several background people also break the one-person instruction.
Our five checks
- MissedExactly one person, one umbrella and one bicycle
- MissedOwn left hand visibly grips the umbrella handle
- MissedOwn right hand visibly holds the bicycle handlebar
- MetBoth complete hands are visible, attached and anatomically plausible
- MetTeal jacket, charcoal trousers, brown boots, navy helmet and orange bag are present

Flex
$0.050 recordedThe raised hand is at the umbrella edge, disconnected from the central handle and shaft. That is the courier’s own right hand; their own left holds the bicycle, reversing the requested assignment.
Our five checks
- MetExactly one person, one umbrella and one bicycle
- MissedOwn left hand visibly grips the umbrella handle
- MissedOwn right hand visibly holds the bicycle handlebar
- MetBoth complete hands are visible, attached and anatomically plausible
- MetTeal jacket, charcoal trousers, brown boots, navy helmet and orange bag are present

FLUX 3 Image
$0.048 recordedThe courier’s own left hand crosses their chest to grip the connected umbrella handle. Their own right hand holds the bicycle handlebar. Both hands and all five clothing colors are visible.
Our five checks
- MetExactly one person, one umbrella and one bicycle
- MetOwn left hand visibly grips the umbrella handle
- MetOwn right hand visibly holds the bicycle handlebar
- MetBoth complete hands are visible, attached and anatomically plausible
- MetTeal jacket, charcoal trousers, brown boots, navy helmet and orange bag are present
Same woman, new pose
Keep the woman and her clothes from the first reference, take only the pose from the sketch, and photograph her at a coastal lookout.
Task prompt and checks
Create a realistic full-body photograph of the woman in image 1 at a coastal lookout. Keep her face, short curly black hair, blue denim jacket, cream T-shirt and charcoal trousers from image 1. Use only the body pose from the pencil sketch in image 2: turn her body and face toward the right side of the picture, lean forward slightly, place one foot on a rock with that knee bent, raise one hand to shade her eyes and let the other arm hang down. Show one woman, both hands and both feet. Keep her recognizable as the woman in image 1. The finished image must look like a photograph, with natural skin texture and believable daylight, not a pencil drawing.
- Identity and short curly hair remain recognizable
- Denim jacket, cream shirt and charcoal trousers stay with the woman
- Body and face turn toward picture-right
- One foot is on a rock, one hand shades the eyes and the other arm hangs down
- One photographic person, with both hands and both feet visible and plausible
Trial 1 of 2. Select an image for the original file.

Klein 4B
$0.016 recordedThe rock, raised foot and shading hand follow the sketch. Her outfit stays intact. The strong profile leaves facial identity uncertain against the frontal reference.
Our five checks
- UnclearIdentity and short curly hair remain recognizable
- MetDenim jacket, cream shirt and charcoal trousers stay with the woman
- MetBody and face turn toward picture-right
- MetOne foot is on a rock, one hand shades the eyes and the other arm hangs down
- MetOne photographic person, with both hands and both feet visible and plausible

Klein 9B
$0.019 recordedThe pose, coastline and three clothing colors are present. Both hands and feet are visible. The profile and fuller hair make exact identity preservation hard to judge.
Our five checks
- UnclearIdentity and short curly hair remain recognizable
- MetDenim jacket, cream shirt and charcoal trousers stay with the woman
- MetBody and face turn toward picture-right
- MetOne foot is on a rock, one hand shades the eyes and the other arm hangs down
- MetOne photographic person, with both hands and both feet visible and plausible

Pro
$0.060 recordedThe woman follows the requested pose and keeps the denim-and-cream outfit. Footwear changes, which this brief leaves open. The face is in profile, so identity remains uncertain.
Our five checks
- UnclearIdentity and short curly hair remain recognizable
- MetDenim jacket, cream shirt and charcoal trousers stay with the woman
- MetBody and face turn toward picture-right
- MetOne foot is on a rock, one hand shades the eyes and the other arm hangs down
- MetOne photographic person, with both hands and both feet visible and plausible
The provider returned an error. No image was received for this trial.
Max
HTTP 500 from the generation API. No image was returned, so we make no visual judgment. No charge receipt was returned.

Flex
$0.150 recordedThe requested body direction, hand positions and raised foot are present. The outfit colors remain. A profile face limits comparison with the original frontal portrait.
Our five checks
- UnclearIdentity and short curly hair remain recognizable
- MetDenim jacket, cream shirt and charcoal trousers stay with the woman
- MetBody and face turn toward picture-right
- MetOne foot is on a rock, one hand shades the eyes and the other arm hangs down
- MetOne photographic person, with both hands and both feet visible and plausible
The provider returned an error. No image was received for this trial.
FLUX 3 Image
HTTP 500 from the generation API. No image was returned, so we make no visual judgment. No charge receipt was returned.
How we tested the harder tasks
On October 20, 2026, we ran three fixed prompts twice in each of the six models through Vercel AI Gateway. The 36 planned trials returned 30 images. Every image and failed trial is available above. We retried the first Max pose request once after HTTP 500; it also failed. Both attempts are in the record. We made no extra attempts to replace a visual result. Recorded charges for returned images total $1.592. Failed requests without a charge receipt are excluded from that amount.
The two editing tasks use the same two references in the same order for every model. These are our existing AI-created fictional adults and a pose sketch. Each input passes through the product's JPEG conversion at a maximum long edge of 960 pixels. The courier task has no image inputs.
We used the product's photo-generation prompt wrapper and native model defaults. FLUX.2 requests 768 × 1024 or 1024 × 768; FLUX 3 uses its 1K tier with the same requested aspect ratio. Exact returned sizes, seeds when reported, full submitted prompts, input hashes, output hashes, model IDs, charges and receipts are in the downloadable record.
Our checks were written before generation. Identity judgments are visual, not biometric measurements. A profile view can leave identity uncertain. Two trials show individual failures and variation; they cannot establish a stable success rate or an overall model ranking. Provider charges are for these samples, not PhotoGenerAI credit prices.
We offer these models in PhotoGenerAI. Our observations are editorial judgments; the originals let you check them yourself.
Download the complete test recordThe baseline: one prompt in six models
Select any image to inspect the full-size WebP.

FLUX 3 Image
$0.048 in this testA broad, rounded pot with a low circular pedestal. The large soft highlight and generous side handle give this result a different silhouette from the taller Klein pots.

FLUX.2 Pro
$0.030 in this testA compact pot with both a side handle and an extra overhead handle. The glossy surface has large, dark reflections. The prompt did not ask for the second handle.

FLUX.2 Max
$0.070 in this testA taller, straighter body with broad vertical reflections. The pot sits further back in the frame, with more space above it and a long shadow to the right.

FLUX.2 Flex
$0.050 in this testA round body, a pronounced lid rim and a high curved spout. The blue is vivid, with a purple cast near the base and a large bright reflection on the left.

FLUX.2 Klein 4B
$0.014 in this testA tall, tapered pot on a small circular pedestal. A rectangular highlight and visible glaze texture define the surface. The pot occupies much more of the frame than the pedestal.

FLUX.2 Klein 9B
$0.015 in this testAn upright pot in a lighter blue, on a square pedestal that reaches the bottom edge. The prompt left the pedestal shape open, so this changes the composition without contradicting it.
Our take on this set
All six images include a blue pot, a curved handle and a peach pedestal, with no visible text. Shape and framing vary enough to matter for a product shot. We prefer the broad FLUX 3 silhouette for this composition. For a single-handle design, we would retry the Pro result with a more specific instruction. Klein 4B is the least expensive sample here and still gives a clear ceramic finish.
That is an editorial preference from one result per model. This set helps you choose a starting point for a similar prompt; it cannot establish a general quality or reliability ranking.
Can Pro change the color and keep the scene?
We used the blue Pro output as the input for one FLUX.2 Pro edit. The teapot becomes dark forest green, including both handles. The peach pedestal, cream background and broad reflections remain recognizable. The framing changes with the landscape output, so compare the object and scene rather than expecting a pixel-aligned recolor.


Edit prompt
Change the ceramic teapot to deep forest green. Keep the peach pedestal and the cream background.
Input resized to a maximum of 960 pixels on its long edge. Output 1024 × 768. Recorded edit cost: $0.045.
What the baseline requests cost
These are the actual USD charges recorded by Vercel AI Gateway for our October 14 requests. Klein 4B was $0.014 and Max was $0.070 for this square text-generation sample. Compare the compositions above before deciding which result suits your task. This test supplied no reference photo and did not measure product-identity preservation.
The amounts document this test. PhotoGenerAI uses purchased credits; the generator shows its own credit total before you generate. Current provider prices and charges for other sizes or image inputs can differ.
| Model / task | Recorded USD |
|---|---|
| FLUX 3 Image | $0.048 |
| FLUX.2 Pro | $0.030 |
| FLUX.2 Max | $0.070 |
| FLUX.2 Flex | $0.050 |
| FLUX.2 Klein 4B | $0.014 |
| FLUX.2 Klein 9B | $0.015 |
| FLUX.2 Pro / color edit | $0.045 |
How we made the teapot baseline
PhotoGenerAI generated these demonstration images on , through Vercel AI Gateway. We also offer these models in our product. The visual judgments on this page are our own.
Download the test recordThe record includes exact model IDs, generation receipts, file hashes, prompts and costs.
We used one shared English prompt and one text output per model. FLUX 3 requested its 1K resolution tier; FLUX.2 requested explicit 1024 × 1024 dimensions. The final seven requests returned without parameter warnings. The model IDs are bfl/ followed by the names in the filenames, such as bfl/flux-2-pro.
The Pro edit is a separate task with an image input and a 1024 × 768 output. We encoded the saved assets as WebP at quality 86. The full-size downloads are those WebP files, with no additional crop in this review.
The teapot baseline used one image per model, without a shared seed or repeated trials. It does not measure speed, success rates, portrait quality, typography or editing performance across the six models. The prompt explicitly asks for no text, so it cannot test Flex's typography capabilities.
For the model families and available controls, see Black Forest Labs' FLUX.2 documentation and FLUX 3 Image documentation. The images and costs above come from our test record.
Try your own product prompt
Specify the handle, pedestal shape and framing to narrow the variations you see here.


