FLUX 3 Klein Base LoRA Review: Character and Outpaint LoRAs Tested

The gallery shows the six published examples from the FLUX 3 Klein Base LoRA model page: three text-to-image (two with a character LoRA, one with none), one edit with no LoRA and two outpainting results. They show clothed adults only.

A character LoRA at 0.5 and 0.6. A fictional fennec-eared woman by a cabin stove, and the same character beside a bearded man in a sauna lounge. The LoRA's trigger opens the prompt and a full scene caption follows.

No LoRA, for comparison. A man in a snowy woodshed doorway at 2:3, with cold steel-blue light on one side and warm wood tones on the other.

Edit, two sources. A cream cardigan from a product photo put on the woman from a portrait, her face, freckles and red hair kept.

Outpainting at 1.5. A square photo extended to 16:9 and 9:16 by a Base-trained outpainting LoRA, using the instruction from the file's own card word for word.

Open with the LoRA's trigger phrase or instruction exactly as its model card gives it, then describe the scene in full sentences. Keep every person an adult.

1. Character LoRA. loras with the character file at scale 0.6, aspect_ratio 3:4.

fennec girl. Cinematic photograph, 50mm lens. An adult woman in her late twenties with large fennec fox ears and long auburn hair, in a cream off-shoulder knit sweater dress, sits on a reindeer hide beside a black wood stove in a timber cabin at blue dusk. Warm firelight on her face, cold blue light behind her. Soft film grain, realistic skin texture.

2. Edit LoRA, the card's own instruction. One source image with grey margins, outpainting LoRA at 1.5, aspect_ratio 16:9.

Outpaint the image: replace the solid gray areas with a seamless continuation of the scene, keeping the existing picture unchanged.

3. Edit, no LoRA, two sources.

Dress the woman from image 1 in the cream cable-knit cardigan from image 2, worn open over a black knit top. Keep her face, freckles and red hair. Seat her sideways on a pine bench in a timber sauna lounge, glancing back at the camera. Cold blue daylight, warm amber stove glow, cinematic photograph, 50mm lens.

Weak: a LoRA link and a one-line prompt. Strong: trigger first, a full caption after it, and a second run at scale 0 on the same seed.

SpicyAPI adds no content filter, and the page label is softened. We tested on 16 October 2026 with adult subjects only, over five intensity levels (Mild, Moderate, Strong, Very strong, Maximum; the test prompts are not published), mostly with no LoRA loaded, 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 the Moderate edit was delivered with no LoRA loaded.
• Strong: the text request was toned down on this endpoint (a FLUX.2 Klein LoRA was loaded, which changed nothing), while a Strong edit with no LoRA loaded was delivered once. On the plain Base endpoint, which runs the same weights, Strong was toned down in two female-subject text runs and two edits.
• Very strong and Maximum: not run on this endpoint. On the plain Base endpoint they were toned down in the runs we made.
• Refused: none.

We have not tested a purpose-trained unfiltered LoRA for FLUX 3 Klein, so we cannot say whether one lifts these limits. For stronger levels without such a file, the distilled FLUX 3 Klein LoRA toned down less in our tests.

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, each a direct public link to the .safetensors file. We ran two of the endpoint's published cases and both worked: the character LoRA at 0.6 turned the cabin scene's subject into the fennec-eared character, and the outpainting LoRA at 1.5 extended the source image while leaving the original picture unchanged.

Other results on 16 October: a four-step distillation LoRA at 1 changed contrast and sharpness clearly; a tarot-card LoRA in diffusers format changed almost nothing; a camera-angle LoRA in diffusers format showed no visible difference; and LoRAs for FLUX.2 Klein 9B, including a pixel-art file, left the image identical to the control. Black Forest Labs says a LoRA trained on Klein Base also loads into the distilled Klein, at a scale near 0.75. A fourth LoRA or a scale above 4 is rejected when you submit, a repository page link fails with a "direct download link" error, and a file above 2 GB fails; none of these is charged.

• The quick build with the same LoRA slot? FLUX 3 Klein LoRA is distilled, quicker, and toned down less in our tests.
• No LoRA needed? FLUX 3 Klein Base is the same model without the LoRA slot.
• The distilled build without LoRAs? FLUX 3 Klein takes up to ten reference images.
• Previous generation LoRAs? FLUX.2 Klein Base 9B LoRA takes FLUX.2 Klein files.