FLUX 3 Klein LoRA Review: Your Own LoRAs, No Content Filter

The gallery shows the six published examples from the FLUX 3 Klein LoRA model page: three text-to-image and three edits, with and without the pastel LoRA. They show clothed adults only.

Pastel at 0.5. A woman on the rim of a stone well in a village square, photographed with the palette pulled toward lilac and cream. At this strength it is still a photograph.

Pastel at 0.6, wide. A man lying back in a meadow of wild fennel and poppies at 16:9, soft powder-blue tones.

No LoRA at all. A couple sharing a fig under a vine pergola at dusk, to show the distilled model on its own.

Edits. A straw hat taken from one photo and put on the man in another, restyled as a soft watercolour with the pastel LoRA at 1.2; a single photo restyled as a pastel illustration at 1.2; and an emerald dress swap with no LoRA that keeps the face and the courtyard.

Put the LoRA's trigger word first, then describe the scene in full sentences: subject with an age, clothing, place, light, lens. For edits, name each reference as <ref_image_1> and say what must stay. Keep every person an adult.

1. Text-to-image with a style LoRA. loras with one pastel LoRA at scale 0.5, aspect_ratio 3:4.

pastelstyle, photograph of a woman in her late thirties with fine laugh lines and long wavy auburn hair, sitting on the rim of an old stone well in a sun-bleached village square, in a pale lilac linen sundress, almond blossom drifting around her, soft pastel palette, airy morning light.

2. Edit with two references. Same LoRA at scale 1.2.

pastelstyle, Put the straw panama hat from <ref_image_2> on the man from <ref_image_1>, tilted low over one eye. Keep his face and curls. Turn the market into a soft pastel watercolour scene in early evening light.

3. No LoRA, two people. aspect_ratio 4:5.

Cinematic photograph under a vine-covered pergola at dusk on a Greek island. An adult couple in their thirties, a Greek woman in a rust-red wrap dress and a Moroccan man in an open cream shirt, share a ripe fig across the corner of a long rustic table, both laughing softly. Strings of warm bulbs, terracotta jugs, indigo sky, 50mm lens, film grain.

Weak: a LoRA link with no trigger word and no control run. Strong: trigger first, a full caption after it, and a second run at scale 0 on the same seed.

SpicyAPI adds no content filter. We tested with adult subjects only, over five intensity levels (Mild, Moderate, Strong, Very strong, Maximum; the test prompts are not published), with no LoRA loaded unless stated, and read every image. Results are reported as delivered, toned down or refused. Sample sizes are small, so these are observations, not rates.

• Mild and Moderate: delivered on text-to-image and edit (16 October).
• Strong: delivered on edit and in some text-to-image runs, toned down in others. A male-subject Strong run was toned down on both endpoints.
• Very strong: toned down on text-to-image and on edit (26 October).
• Maximum: of the two request types, one was delivered and one toned down on text-to-image, and at least one was delivered on edit (26 October).
• With the pastel LoRA at 1.0: a Moderate text request came back as a clothed illustration, and the matching edit was delivered.
• Refused: none, on either endpoint.

We have not tested a purpose-trained unfiltered LoRA for FLUX 3 Klein, so we cannot say whether one lifts the Very strong limit. Do not assume a similarly named FLUX.2 file does.

Adults only, in every prompt, source photo and LoRA. 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.

Point loras at up to three {path, scale} entries. We ran two of the endpoint's published cases and both worked: the pastel style LoRA at 0.5 in text-to-image gave the well-side photograph its lilac palette, and the same file at 1.2 in editing restyled a hat swap and a single photo. Start around 0.6 to 1 for photos and stop where faces still read as adults; raise the strength in edits.

Files that did not help in our tests: LoRAs trained for FLUX.2 Klein 9B (loaded, no effect), a tarot-card style LoRA and a camera-angle LoRA in diffusers format (almost no effect). Black Forest Labs says a LoRA trained on Klein Base loads into the distilled Klein without conversion; a scale near 0.75 is their suggestion. Civitai download links worked in our tests, including ones that answer a bare request with a login error, but a repository web page does not.

• Full 50-step sampling and the Base LoRA training base? FLUX 3 Klein Base LoRA runs the undistilled model with up to three LoRAs.
• No LoRA needed? FLUX 3 Klein is the same model with up to ten reference images in an edit.
• Previous generation LoRAs? FLUX.2 Klein 9B LoRA takes FLUX.2 Klein files.
• A different base? FLUX.1 Dev LoRA has a large library of ready-made files.