How is Base 4B different from the distilled Klein 4B?

Black Forest Labs released both on 15 January 2026. The distilled 4B is trained to finish in a few steps for speed. Base 4B keeps the full, undistilled model: its makers recommend it for fine-tuning and LoRA training and say it gives more varied results. In our tests Base was about four times slower, showed style LoRAs more strongly, and followed bold edit instructions less reliably.

Do LoRAs trained on the distilled 4B work here?

In our tests, yes: a LoRA for the stronger levels and a colored-pencil LoRA, both made for the distilled 4B, applied on Base. LoRAs trained on Base also worked on the distilled build. Files for FLUX.1 or the 9B Klein models load silently and have no effect.

Can I use uncensored LoRAs for NSFW images?

Yes. Weights are not reviewed and results are not filtered. In our tests a matching LoRA brought out the Strong level that the model softened on its own; the Very strong level stayed toned down even then.

Why did my edit come back unchanged?

This undistilled build sometimes returns the source when an instruction asks for one of the strongest edits. Load a LoRA trained for the change, or use FLUX.2 Klein 4B LoRA, which followed such edits in our tests.

Why did my LoRA request fail?

Usually the link needs a login (gated Hugging Face files answer with HTTP 401) or points to a repository page instead of the file. Use a link that downloads the .safetensors file itself. Failed requests are not charged.

Can I stack several LoRAs?

Yes, up to three per request, each with its own scale. A fourth entry is rejected. Loading LoRAs adds no separate charge.

Does a bigger size cost more?

No. Every aspect ratio and both size tiers cost the same per image, in text to image and in edits.

Am I charged for a failed request?

No. Failed requests are not charged.

FLUX.2 Klein Base 4B LoRA API: Uncensored NSFW Images

black-forest-labs/flux-2-klein-base-4b-lora/edit

FLUX.2 Klein Base 4B LoRA image edit on Black Forest Labs' FLUX.2 [klein] Base 4B: change or combine up to three source images with a written instruction and up to three of your own LoRA files.

From $0.024/image

Run model View API

Input

Image editing

Prompt436/5000

Reordering updates prompt references. Removing media also removes its references; review your prompt.

Images2/3

Drop files or media links, paste, or browse · Up to 3 files · Max 10 MB · JPG · PNG · WEBP · GIF

image 1
image 2

Aspect ratio

Resolution

Style LoRAs0/3

Seed

Safety checkerOff

SpicyAPI does not run a safety checker on this model. Your prompt and media go to the model as written, and the model itself decides what comes back.

Checking your account before enabling submission and pricing.

Output

Ready for a new run

Example result

Animate

Edit

Split into layers

{
  "model": "black-forest-labs/flux-2-klein-base-4b-lora/edit",
  "input": {
    "image_urls": [
      "https://cdn.spicyapi.ai/models/examples/inputs/b74f2dfda7dfec36.jpeg",
      "https://cdn.spicyapi.ai/models/examples/inputs/cc20b9c19e58cd5c.jpeg"
    ],
    "aspect_ratio": "3:4",
    "resolution": "2k",
    "seed": 88917,
    "prompt": "Place the woman from image 1 inside the observatory from image 2 at night: the dome slit is open to a full moon and silver light falls across the brass telescope. She leans against the telescope mount in a midnight-blue velvet slip dress with thin straps and a low neckline, her cream cardigan slipped down to her elbows, bare shoulders, braid over one shoulder, looking at the camera with a quiet smile. Keep her face and silver braid."
  }
}

Introduction

FLUX.2 Klein Base 4B LoRA Image edit changes or combines up to three source images with a written instruction on Black Forest Labs' FLUX.2 [klein] Base 4B, the undistilled build, with up to three of your own LoRA files loaded. Keep the shape of your first image with auto, or choose one of nine ratios and a 1k or 2k size tier. One flat price per image, and nothing is filtered.

Choose a mode by starting material

What you have	Mode
one to three pictures to change or combine	FLUX.2 Klein Base 4B LoRA Image edit
a written description only	FLUX.2 Klein Base 4B LoRA Text to image

Key features & highlights

Restyle with the LoRAs you train

Load a Base 4B LoRA and ask for a redraw: a park photo became a fantasy painting at a telescope, and a daylight room became pixel art at night.

New settings for the same people

Move a person into a new place and new clothes while face, hair and glasses stay, from one or two sources.

Up to three sources

Combine a portrait with a location or an outfit. Name the sources as image 1, image 2 and image 3.

A cautious editor

Nothing is filtered, but this build often returns the source unchanged when asked for a large change in wardrobe. A matching LoRA delivered the Strong level in our tests; the Very strong level came back toned down.

Request parameters

Request parameters for FLUX.2 Klein Base 4B LoRA · Image Edit
Parameter	Required	Description
`prompt`	Yes	Say what should change and what must stay the same, up to 5,000 characters. With several sources, refer to them as image 1, image 2 and image 3. at least 1 characters · at most 5000 characters
`image_urls`	Yes	One to three source images, up to 10 MB each, in the order your prompt refers to them. Required. at least 1 items · at most 3 items
`aspect_ratio`	No	Auto (default) keeps the shape of the first source at about one megapixel. Pick one of nine ratios from 21:9 to 9:21 to reframe the result. `auto` · `1:1` · `16:9` · `9:16` · `4:3` · `3:4` · `3:2` · `2:3` · `21:9` · `9:21` · Default: `auto`
`resolution`	No	1k (default) or 2k, which is 2048 pixels on the long side. It only applies when you choose a ratio; with auto the size is about one megapixel. `1k` · `2k` · Default: `1k`
`loras`	No	Up to three {path, scale} entries trained for FLUX.2 [klein] 4B or Base 4B. The path must download the .safetensors file itself; scale runs from 0 to 4 (default 1). Default: `[]` at most 3 items
`seed`	No	A number from 0 to 2147483647. Reuse it to compare edits while you change one detail of the prompt; it does not promise identical pixels. `≥ 0` · `≤ 2147483647`

Pricing

$0.024 / image

Billed per image at one flat price, whatever the number of source images, the aspect ratio or the size tier. Loading LoRAs adds no separate charge.

Use cases

Photo to painting or pixel art

A style LoRA trained on Base 4B plus a short redraw instruction turns photos into matching artwork, keeping the people and the layout recognisable across a whole set.

Scene and wardrobe moves

Put a portrait into a location photo and describe the outfit you want there. Naming each garment and where it sits gives the most faithful result.

Reframing

Pick a ratio different from the source and describe what should fill the new space; the edit recomposes the frame while faces and clothing stay close to the source.

Explore key use-case tutorials

Prompt guide & pro tips

Protect first, change second

Open with what must stay (face, hair, pose), then describe the change in concrete words.

Use a LoRA for big changes

If an edit comes back unchanged, load a LoRA trained for that change or try FLUX.2 Klein 4B LoRA, which followed such instructions more reliably in our tests.

Compare against no LoRA

Run the same prompt and seed without the LoRA. If the two results look the same, the file was made for another base model, such as FLUX.1 or the 9B Klein models: those load without an error and do nothing here.

Use direct, public links

The link must download the .safetensors file itself. A Hugging Face file link or a Civitai download link works; a repository page or a gated file that needs a login fails the request, which is then not charged.

API quick start

# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
  -H "Authorization: Bearer $SPICY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
  -d '{
  "model": "black-forest-labs/flux-2-klein-base-4b-lora/edit",
  "input": {
    "aspect_ratio": "auto",
    "image_urls": [
      "https://example.com/reference.png"
    ],
    "prompt": "A ceramic teapot on a linen cloth, soft window light, muted palette",
    "resolution": "1k"
  }
}'

# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
  -H "Authorization: Bearer $SPICY_API_KEY"

See the API reference for authentication, the complete schema, polling, and error handling.

Missing the model you need?

Tell us what to add and where you would use it. Requests go to the people who decide what gets onboarded next.

LORA IMAGE API

FLUX.2 Klein Base 4B LoRA API at a glance

FLUX.2 [klein] Base 4B with your own LoRAs: the full, undistilled 4B model that LoRAs are trained on, and nothing filtered.

Parameters: 4B; Black Forest Labs' undistilled FLUX.2 [klein] Base 4B, Apache 2.0 open weights.
LoRAs per call: 3; Up to three weight files per image, each with its own scale from 0 to 4.
Typical render: ~20 s; Median text-to-image request time across our tests; the distilled 4B took about 4 seconds.
Source images: 3; Edit or combine up to three pictures in one request.
Largest side: 2048 px; Nine aspect ratios at about one megapixel or 2048 pixels on the long side.
Content filter: None; No filter on prompts, weights or results.

Results

What you can make with FLUX.2 Klein Base 4B LoRA

Style
A painting from a Base-trained LoRA
A fantasy-painting LoRA trained on Base 4B with its trigger first: an observatory stair under a single moonbeam, painted at 16:9 and 2k.
Text to image with a fantasy-painting LoRA at 1, 16:9 at 2k.
Restyle
A park photo moved into a painting
An edit with the same painting LoRA: a couple photographed in a sunny park now stands at the eyepiece of a brass telescope under the moon, with their faces kept.
Image edit with a fantasy-painting LoRA at 1, 3:2.
Base model
Photographs without any LoRA
No LoRA loaded: moonlight from the dome slit and a green banker's lamp over hand-drawn star charts, a 2k portrait from a few sentences.
Text to image without a LoRA, 3:4 at 2k.

More clips made with FLUX.2 Klein Base 4B LoRA

Image edit with two sources: a portrait and an observatory, 3:4 at 2k.
Text to image with a colored-pencil LoRA at 1, 4:3.
Image edit with one source, aspect ratio auto.
Text to image without a LoRA, 2:3.
Image edit with a pixel-art LoRA at 1, aspect ratio auto.

Content test

Is FLUX.2 Klein Base 4B LoRA uncensored? What we tested

We ran each level on our live setup and checked every clip frame by frame. A pass means the content was actually on screen, not just that the request went through.

Content levels tested on FLUX.2 Klein Base 4B LoRA, by mode
Level	Text to image	Image editing
MildThe lightest prompts in the test	RendersTested Oct 9, 2026	RendersTested Oct 9, 2026
ModerateWhere filtered models start to refuse	RendersTested Oct 9, 2026	Not tested yet
StrongThe middle of the scale	RendersTested Oct 9, 2026	RendersTested Oct 9, 2026
Very strongWhere most models start to hold back	Not tested yet	Not tested yet
MaximumThe top of the scale	RendersTested Oct 9, 2026	RendersTested Oct 9, 2026

Renders: on screen as written. Softened: generated, but milder than what was written. Refused: rejected. Not tested yet: we have not run that level on that mode.

How each mode behaves

All modes: Use LoRAs trained for FLUX.2 [klein] Base 4B or the distilled Klein 4B. A LoRA made for FLUX.1 or for the 9B Klein models loads without an error and changes nothing, and one community file in our tests turned the whole frame into noise without an error. A link that needs a login or points to a web page fails the request. Up to three LoRAs, scale 0 to 4; at 4 the image turned to noise.
All modes: Nothing is reviewed or refused, but the model is conservative: without a matching LoRA it often tones the Strong level down, and the Very strong level came back toned down even with one loaded. The Maximum level rendered.
Image editing: The strongest edits often come back as the unchanged source. In our tests most Strong-level edit instructions without a LoRA returned the original photo, while lighting, setting and style changes worked.
All modes: It is about four times slower than the distilled 4B in our tests. There is no negative prompt, step count or guidance control, and signs or labels often come out as garbled text. One JPEG per request, up to 2048 pixels on the long side.
All modes: Every person you describe must be an adult. You are responsible for the lawfulness of what you generate and for having the rights to any source image and LoRA you use.

Use cases

What to build with the FLUX.2 Klein Base 4B LoRA API

Testing LoRAs you trained
Train on Klein Base 4B and run the file here at full quality, three at a time, without hosting a GPU.
Painted and pixel styles
Fantasy painting, colored-pencil and pixel-art LoRAs take over the whole frame on this build, even at scale 1.
Varied takes on one prompt
The undistilled model gives more variety between seeds, which helps when you are looking for a composition rather than polishing one.
Your own LoRAs on an unfiltered model
Weights are never reviewed and results are never filtered. The model holds back on its own, so load the LoRA your work needs.

Prompts

FLUX.2 Klein Base 4B LoRA prompt examples you can copy

Paste one into the playground as it is, then change one thing at a time.

Text to image
Star charts by moonlight
Moonlight falls through the open slit of an old observatory dome onto a long oak table covered in hand-inked star charts. An adult Black woman in her thirties with close-cropped hair leans over the charts in a loose ivory silk blouse slipping off one shoulder and high-waisted wool trousers, a brass divider in her hand, her face lit by a small green banker's lamp as she looks up at the camera. The great brass refractor rises behind her into the dark. Silver moonlight against warm lamp glow, 50mm, fine film grain.
Why it works. No LoRA. Two light sources, moonlight and a desk lamp, are named so the frame keeps its contrast at 2k.
Try this prompt
Image editing
The silver-haired astronomer
Place the woman from image 1 inside the observatory from image 2 at night: the dome slit is open to a full moon and silver light falls across the brass telescope. She leans against the telescope mount in a midnight-blue velvet slip dress with thin straps and a low neckline, her cream cardigan slipped down to her elbows, bare shoulders, braid over one shoulder, looking at the camera with a quiet smile. Keep her face and silver braid.
Why it works. Two sources: a studio portrait and an empty observatory. Naming each piece of the outfit made it come out exactly as written.
This mode starts from your own image: add it in the playground before you run.
Try this prompt
Text to image
A fantasy observatory stair
light_fantasy, a detailed fantasy painting of the inside of an old observatory at night: an adult woman in her thirties in a midnight-blue velvet gown with bare shoulders climbs the iron spiral stair toward the platform of a huge brass telescope and looks back over her shoulder at an adult man in his thirties in a loosened white shirt who holds a lantern at the foot of the stair. A single moonbeam falls through the open dome slit, star charts cover the curved walls, dust glitters in the light.
Why it works. A fantasy-painting LoRA trained on Base 4B with its trigger first. Base shows this style strongly at scale 1.
Try this prompt

Pricing

FLUX.2 Klein Base 4B LoRA API pricing

One image starts at $0.018.

The playground shows the exact price for your settings before you run. Failed or timed-out tasks are usually not charged. Some models charge for content refusals as stated on their model page; check those billing terms before submitting.

See every price tier

Get started

How to use the FLUX.2 Klein Base 4B LoRA API in three steps

Step 1
Bring LoRAs trained on Base 4B
Base 4B is what most Klein 4B LoRAs are trained on, so they apply here at full strength. LoRAs trained on the distilled 4B loaded too in our tests; files for FLUX.1 or the 9B Klein models change nothing.
Step 2
Start the scale at 1
Style LoRAs show strongly on this build. Begin at 1, try 0.5 for a lighter touch, and stay below 4: at 4 the image turned to noise.
Step 3
Write sentences, trigger first
Put the LoRA trigger phrase first, then describe subject, clothing, place and light. There is no negative prompt or guidance control.

Open the playground Get an API key

Parameter reference and code samples are in the API and Readme tabs at the top of this page.

Why SpicyAPI

Why run FLUX.2 Klein Base 4B LoRA on SpicyAPI

No filter from us
We add no content filter of our own. What the model renders is what you get back.
Tested, not guessed
Content labels come from real generations we ran and checked, with the test date on the page.
Pay per result, in dollars
No subscription and no credits. Failed or timed-out tasks are usually not charged. Some models charge for content refusals as stated on their model page; check those billing terms before submitting.
One key for every model
One API key and balance cover image, video, audio and chat models, so you can explore more models from the same account.

FAQ

FLUX.2 Klein Base 4B LoRA API FAQ

About FLUX.2 Klein Base 4B LoRA

FLUX.2 Klein Base 4B LoRA puts Black Forest Labs' undistilled, Apache 2.0 FLUX.2 [klein] Base 4B behind your own LoRAs: the build Klein LoRAs are trained on, with text to image in nine aspect ratios up to 2048 pixels and editing from up to three source images. LoRAs load without review, and nothing is filtered.