FLUX 3 Klein Review: Unfiltered Images and Multi-Reference Edits
Pick FLUX 3 Klein for cheap, quick unfiltered stills and edits with up to ten reference images; look elsewhere if top-level requests must land every time.
FLUX 3 Klein is Black Forest Labs' 5.6B open-weights image model from October 2026, in its step-distilled build. SpicyAPI serves it as text-to-image and as an edit endpoint, at 1k or 2k, with no content filter on the prompt, the source images or the result. We tested both endpoints on 16 and 22 October 2026 with a fixed set of intensity levels and checked every image by eye.
Cinematic 35mm photograph in a Sicilian lemon grove at noon. An adult Italian woman about thirty with sun-streaked brown hair leans back against a weathered wooden picking ladder, a woven basket of lemons on her hip, wearing an ochre linen slip dress with one thin strap slipping off her shoulder. Dappled light through glossy leaves, warm skin, a slow sultry glance at the camera. Kodak Portra colour, soft grain.
Intimate photograph on the whitewashed steps of a small hilltop chapel at blue hour. An adult couple in their thirties, a Spanish woman in an off-the-shoulder cream cotton dress and a Lebanese man in a loose sand-coloured linen shirt, sit close with their foreheads touching, his hand resting on her bare shoulder, eyes nearly closed. A single warm lamp above the chapel door, indigo sky, soft film grain, 85mm lens, shallow depth of field.
Wide cinematic photograph of a sleepy Greek island afternoon. An adult Black man about thirty-eight dozes in a faded rope hammock strung under an old fig tree, white linen shirt unbuttoned to the waist, one arm behind his head, half-open eyes catching the camera with a lazy smouldering look. A cobalt-blue painted wall behind him, broad leaves throwing patterned shade. 50mm lens, golden late-afternoon light, natural skin texture.
Dress the woman from <ref_image_1> in the cobalt silk halter dress from <ref_image_2>. Keep her face, freckles and copper hair exactly. Move the light to a low golden late-afternoon sun; she leans back against the ochre wall, glancing over her bare shoulder with a slow, smouldering look while the sea breeze lifts her hair.
Keep his face, beard and the stone quay. Replace the knitted sweater with an open white linen shirt, make his hair wet and swept back as if he has just climbed out of the sea, and relight the whole scene to a low amber sunset. He looks straight into the camera with a steady, intense gaze.
Keep her face, the potter's wheel and the shelves of painted plates. Change her apron and work shirt to an off-the-shoulder terracotta cotton blouse, let her dark hair fall loose over one shoulder, and relight the workshop to deep blue hour with one warm brass lantern on the bench. She looks up from the clay with a soft, knowing smile.
It is the quick, low-cost member of the FLUX 3 Klein family. In our runs it delivered the first three intensity levels as written on both text-to-image and edit. At Very strong and Maximum it sometimes toned requests down, as the Base build does, because that is how the model itself behaves. Turning prompt expansion on made it more willing in our tests, and it is off by default.
Pick Klein when you want many drafts fast, when you need to edit with several reference images, or when you want to keep unfiltered work cheap. Look elsewhere if you need exact counts, a negative prompt or the top two levels on every run. We did not run a blind quality test for this review, so no quality score is quoted here; the score card on this page shows the live Spicy Index and Freedom Score.
Content limits: SpicyAPI adds no filter of its own, and the page label is Uncensored. Where results were toned down, it was the model, not a block: nothing was refused in our runs.
Text-to-image:prompt (up to 5,000 characters), aspect_ratio (15 options, or custom width and height up to 1024), resolution (1k or 2k), output_format (JPEG, PNG or WebP), enable_prompt_expansion (default off) and seed.
Edit: the same, plus image_urls with 1 to 10 reference images that the prompt can name one by one as <ref_image_1>, <ref_image_2> and so on.
Sizes: 1k is about one megapixel (880 x 1184 at 3:4); 2k is about four (1776 x 2368 at 3:4).
Speed: most 1k text-to-image runs finished in seconds of model time; edits and 2k take longer.
Price: the model page always shows the current price per image for each tier.
良かった点と足りなかった点
良かった点
Low price per image, which makes batches of drafts and seed comparisons affordable.
Up to ten reference images in one edit. A four-image group composite kept each of the four people recognisable.
Edits keep the person. The published edit examples swap clothing, light and place while the face, hair and freckles stay put.
Two sizes, fifteen shapes. From 3:1 panoramas to 1:3 phone strips at 1k or 2k.
Same weights on both routes. We ran the same prompts and seeds on two routes and got the same kind of result, so behaviour does not depend on which route serves a request.
足りなかった点
The top two levels are not guaranteed. Maximum prompts were sometimes toned down, and the result changed from seed to seed.
Prompt expansion changes the outcome. With it on, our Very strong text-to-image runs came back as asked three times out of three; with it off, at most one of three did. It also rewrites your wording, so use it on purpose.
No negative prompt. Describe what you want to see.
Custom sizes are capped at 1024 per side, so go through the 2k tier for large images.
Anatomy and proportions can drift in the most demanding scenes; render several seeds when a scene matters.
What the images look like
The gallery shows the published examples from the FLUX 3 Klein model page: three text-to-image and three edits. They show clothed adults only.
Warm light, long lens. A Sicilian lemon grove at noon with dappled light through the leaves, a woven basket and film grain, in a 35mm photographic look.
Two people on chapel steps. A couple sit close at blue hour on whitewashed steps; both faces and the cream and sand clothing are kept consistent.
Wide frame. A sleepy island afternoon under a fig tree in 16:9, with the hammock and the rope texture sharp across the frame.
Edits from references. The dress swap takes a dress from a second image and puts it on the woman from the first, keeping her face and freckles; the quay and workshop edits change clothing and relight the scene while keeping the face and the place.
Prompts that worked
Write a full scene caption in plain sentences: framing, subject with an age, clothing, place, light, lens. For edits, name each source (<ref_image_1>) and say what must stay. Keep every person an adult.
Wide cinematic photograph of a sleepy Greek island afternoon. An adult man about thirty-eight dozes in a faded rope hammock strung under an old fig tree, white linen shirt, one arm behind his head, dappled light on whitewashed walls behind him. 35mm lens, warm colour, fine grain.
2. Text-to-image, portrait.aspect_ratio 3:4.
text
Cinematic 35mm photograph in a lemon grove at noon. An adult woman about thirty with sun-streaked brown hair leans against a weathered wooden ladder, a woven basket of lemons on her hip, an ochre linen dress, dappled light through glossy leaves, Kodak Portra colour.
3. Edit, new clothing and light. Two references in image_urls, aspect_ratio left unset.
text
Dress the woman from <ref_image_1> in the cobalt silk dress from <ref_image_2>. Keep her face, freckles and copper hair exactly. Move the light to a low golden late-afternoon sun against an ochre wall.
Weak: "make it look better". Strong: a full caption with clothing, place and light, and a short list of what must not change.
How far it goes
SpicyAPI adds no content filter, and the page label is Uncensored. We tested in October 2026 with adult subjects only, over five intensity levels (Mild, Moderate, Strong, Very strong, Maximum; the test prompts are not published), and read every image. Results are reported as delivered, toned down or refused.
Mild to Strong: delivered as written on text-to-image and edit.
Very strong and Maximum: a mix. Edits delivered at Very strong; text-to-image at Very strong came back as asked in at most one run of three with expansion off, and the Maximum requests were toned down.
Prompt expansion on: three of three Very strong text-to-image runs came back as asked.
Refused: none, on either endpoint. Safety checking of the request is switched off.
Adults only, in every prompt and every source photo. Never upload a real person's likeness without consent, never depict anyone who could read as a minor, and keep your outputs private unless you have the right to share them.
Alternatives
Stricter instruction following?FLUX 3 Klein Base is the undistilled build: slower and costlier, and in our tests it settled on tamer results more often.
Yes. SpicyAPI adds no filter on prompts, source images or results, and nothing was refused in our October 2026 tests. The model itself sometimes tones down Very strong and Maximum requests.
Up to ten in one edit request. Name them in the prompt as <ref_image_1>, <ref_image_2> and so on.
Only if you want the model to enrich a short prompt. It rewrites your wording, and in our tests it made Very strong requests more likely to come back as asked. It is off by default.
Klein is the quick step-distilled build; Base is the undistilled build with full sampling, slower and costlier. In our tests Klein delivered Very strong requests more often.
1k (about one megapixel) and 2k (about four), in 15 aspect ratios, or a custom width and height up to 1024.
