FLUX 3 Image Klein Base
FLUX 3 Image Klein Base is the full undistilled checkpoint of Black Forest Labs' open-weights FLUX 3 image model, covering text-to-image generation and reference-based editing. It keeps the complete model capacity, which makes it the variant to fine-tune, customize, and optimize. It suits research, post-training work, and workflows that favor control over generation speed.
Guidance and negative prompts
How to use CFG scale and negative prompts with FLUX 3 Image Klein Base: how guidance changes adherence, what a negative prompt removes, and what step count to run.
Introduction
FLUX 3 Image Klein Base is the full undistilled checkpoint of FLUX 3 Image Klein. It takes the same captions, sizes, references and bounding boxes as the distilled model, and adds two controls the distilled model does not have: CFGScale, which sets how strongly the image follows the prompt, and negativePrompt, which lists what to keep out. It runs 50 steps by default.
A wide shot at standing height of a spa treatment room for a wellness resort website: a massage table made up with white linen and a folded sage towel in the center, a low teak cabinet on the right holding rolled towels and amber oil bottles, and a large window on the left with sheer linen curtains onto green palms. Limewashed walls in warm sand, a pale stone floor. Soft late-afternoon light through the curtains, gentle long shadows. Photoreal hospitality photography, 24mm lens, calm sand and sage palette.
This guide covers what CFGScale changes, what a negative prompt can and cannot remove, and how the step count trades time for detail. Writing the caption itself is covered in the prompting guide for FLUX 3 Image Klein, and it applies to this model unchanged.
How guidance changes the result
CFGScale takes values from 1 to 20 and defaults to 3.5. The three images below share a caption and a seed, so the guidance value is the only difference.
A medium shot from standing height looking down at a 30 degree angle, the blanket filling the lower two thirds of the frame: a summer picnic on a lawn for a homeware catalog, with a red and white striped cotton blanket, a wicker hamper with its lid open, two tall glasses of pink lemonade with paper straws, a ceramic bowl of strawberries and a folded linen napkin. Behind it, out of focus, a row of plane trees and a park bench. Bright afternoon sun from the right, crisp dappled shadows on the blanket. Photoreal lifestyle catalog photography, 35mm lens, fresh red and green palette.
A medium shot from standing height looking down at a 30 degree angle, the blanket filling the lower two thirds of the frame: a summer picnic on a lawn for a homeware catalog, with a red and white striped cotton blanket, a wicker hamper with its lid open, two tall glasses of pink lemonade with paper straws, a ceramic bowl of strawberries and a folded linen napkin. Behind it, out of focus, a row of plane trees and a park bench. Bright afternoon sun from the right, crisp dappled shadows on the blanket. Photoreal lifestyle catalog photography, 35mm lens, fresh red and green palette.
A medium shot from standing height looking down at a 30 degree angle, the blanket filling the lower two thirds of the frame: a summer picnic on a lawn for a homeware catalog, with a red and white striped cotton blanket, a wicker hamper with its lid open, two tall glasses of pink lemonade with paper straws, a ceramic bowl of strawberries and a folded linen napkin. Behind it, out of focus, a row of plane trees and a park bench. Bright afternoon sun from the right, crisp dappled shadows on the blanket. Photoreal lifestyle catalog photography, 35mm lens, fresh red and green palette.
At 1, guidance is effectively off and the prompt is followed loosely: the hamper turns into a tote and the strawberries look like apricots. At 3.5 every item in the caption is present and in place. At 7 the arrangement is the same as 3.5 with deeper color and harder contrast, and nothing gets more accurate.
Treat the default as the working value. Raise it a few points when a render looks flat and you want more punch, and avoid 1, which also switches the negative prompt off.
import { createClient } from '@runware/sdk'
const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()
const [result] = await client.run({
model: 'bfl:flux@3-image-klein-base',
positivePrompt: 'A medium shot from standing height looking down at a 30 degree angle, the blanket filling the lower two thirds of the frame: a summer picnic on a lawn for a homeware catalog, with a red and white striped cotton blanket, a wicker hamper with its lid open, two tall glasses of pink lemonade with paper straws, a ceramic bowl of strawberries and a folded linen napkin. Behind it, out of focus, a row of plane trees and a park bench. Bright afternoon sun from the right, crisp dappled shadows on the blanket. Photoreal lifestyle catalog photography, 35mm lens, fresh red and green palette.',
CFGScale: 7,
width: 1248,
height: 832
})import asyncio
import os
from runware import Runware
async def main():
async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
results = await client.run({
"model": "bfl:flux@3-image-klein-base",
"positivePrompt": "A medium shot from standing height looking down at a 30 degree angle, the blanket filling the lower two thirds of the frame: a summer picnic on a lawn for a homeware catalog, with a red and white striped cotton blanket, a wicker hamper with its lid open, two tall glasses of pink lemonade with paper straws, a ceramic bowl of strawberries and a folded linen napkin. Behind it, out of focus, a row of plane trees and a park bench. Bright afternoon sun from the right, crisp dappled shadows on the blanket. Photoreal lifestyle catalog photography, 35mm lens, fresh red and green palette.",
"CFGScale": 7,
"width": 1248,
"height": 832
})
asyncio.run(main())curl https://api.runware.ai/v1 \
-H "Authorization: Bearer $RUNWARE_API_KEY" \
-H "Content-Type: application/json" \
-d '[
{
"taskType": "imageInference",
"taskUUID": "0f1e2d3c-4b5a-4968-8776-5a4b3c2d1e0f",
"model": "bfl:flux@3-image-klein-base",
"positivePrompt": "A medium shot from standing height looking down at a 30 degree angle, the blanket filling the lower two thirds of the frame: a summer picnic on a lawn for a homeware catalog, with a red and white striped cotton blanket, a wicker hamper with its lid open, two tall glasses of pink lemonade with paper straws, a ceramic bowl of strawberries and a folded linen napkin. Behind it, out of focus, a row of plane trees and a park bench. Bright afternoon sun from the right, crisp dappled shadows on the blanket. Photoreal lifestyle catalog photography, 35mm lens, fresh red and green palette.",
"CFGScale": 7,
"width": 1248,
"height": 832
}
]'runware run bfl:flux@3-image-klein-base \
positivePrompt="A medium shot from standing height looking down at a 30 degree angle, the blanket filling the lower two thirds of the frame: a summer picnic on a lawn for a homeware catalog, with a red and white striped cotton blanket, a wicker hamper with its lid open, two tall glasses of pink lemonade with paper straws, a ceramic bowl of strawberries and a folded linen napkin. Behind it, out of focus, a row of plane trees and a park bench. Bright afternoon sun from the right, crisp dappled shadows on the blanket. Photoreal lifestyle catalog photography, 35mm lens, fresh red and green palette." \
CFGScale=7 \
width=1248 \
height=832{
"taskType": "imageInference",
"taskUUID": "0f1e2d3c-4b5a-4968-8776-5a4b3c2d1e0f",
"model": "bfl:flux@3-image-klein-base",
"positivePrompt": "A medium shot from standing height looking down at a 30 degree angle, the blanket filling the lower two thirds of the frame: a summer picnic on a lawn for a homeware catalog, with a red and white striped cotton blanket, a wicker hamper with its lid open, two tall glasses of pink lemonade with paper straws, a ceramic bowl of strawberries and a folded linen napkin. Behind it, out of focus, a row of plane trees and a park bench. Bright afternoon sun from the right, crisp dappled shadows on the blanket. Photoreal lifestyle catalog photography, 35mm lens, fresh red and green palette.",
"CFGScale": 7,
"width": 1248,
"height": 832
}Response
{
"data": [
{
"taskType": "imageInference",
"taskUUID": "0f1e2d3c-4b5a-4968-8776-5a4b3c2d1e0f",
"imageUUID": "a9b8c7d6-e5f4-4a3b-9c2d-1e0f9a8b7c6d",
"imageURL": "https://im.runware.ai/image/os/a14d18/ws/2/ii/a9b8c7d6-e5f4-4a3b-9c2d-1e0f9a8b7c6d.jpg"
}
]
}Removing recurring details
Some details turn up whether or not the caption asks for them. The kitchen below got potted herbs and a cutting board from a caption that lists only the fixtures. negativePrompt takes a list of the things to keep out, and it is the quickest way to strip that default staging from a scene.
A wide shot at counter height, the island centered: a newly built apartment kitchen with flat white cabinets, a white marble island with two oak bar stools, a stainless steel range and a matte black faucet. Behind it, a window over the sink onto a courtyard. Soft morning daylight from the window. Photoreal real estate listing photography, 24mm lens, bright white palette.
A wide shot at counter height, the island centered: a newly built apartment kitchen with flat white cabinets, a white marble island with two oak bar stools, a stainless steel range and a matte black faucet. Behind it, a window over the sink onto a courtyard. Soft morning daylight from the window. Photoreal real estate listing photography, 24mm lens, bright white palette.
Negative promptpotted plants, flowers, fruit bowls, vases, cutting boards, cookbooks, clutter
The caption names cabinets, an island, stools, a range and a faucet, and nothing else. Without a negative prompt, the model staged the back counter with herbs, a cutting board, canisters and bowls. With those props listed in negativePrompt, the counters are bare, the way a new-build listing is shot. Write the negative prompt as plain nouns, the things you would point at in the frame, rather than as a sentence.
What a negative prompt doesn't remove
Text is the exception. A product with a printed surface comes back with a label, and listing "text, label, lettering" in the negative prompt does not take it away. Say what the surface looks like in the caption instead.
A centered close-up at bottle height: an amber glass skincare serum bottle with a black dropper cap standing on a pale travertine shelf in a bathroom, a sprig of eucalyptus beside it. Soft morning light from a window on the left, a gentle highlight on the glass. Beauty product photography, 100mm lens, warm minimal palette.
Negative prompttext, label, lettering, logo, writing
A centered close-up at bottle height: a plain unlabeled amber glass skincare serum bottle with a black dropper cap and no printing on the glass standing on a pale travertine shelf in a bathroom, a sprig of eucalyptus beside it. Soft morning light from a window on the left, a gentle highlight on the glass. Beauty product photography, 100mm lens, warm minimal palette.
The negative prompt left the label in place, filled with letters that spell nothing. Describing the bottle as "plain unlabeled" with "no printing on the glass" gave a clean bottle with the same shelf and light. When the label should carry real copy, quote the words, as the text rendering guide shows.
Choosing a step count
steps defaults to 50 and accepts values from 1 to 50. The base model is billed by compute, so fewer steps cost less, and on the portrait below the difference is small.
A head-and-shoulders portrait, the subject centered and looking just off camera: a man in his forties with close-cropped gray hair, light stubble and laugh lines, wearing a navy crew-neck sweater, for a company team page. Behind him, a softly blurred open-plan office with warm lights. Soft key light from a window on the left and a gentle fill from the right. Photoreal corporate portrait photography, 85mm lens, shallow depth of field, natural skin texture.
A head-and-shoulders portrait, the subject centered and looking just off camera: a man in his forties with close-cropped gray hair, light stubble and laugh lines, wearing a navy crew-neck sweater, for a company team page. Behind him, a softly blurred open-plan office with warm lights. Soft key light from a window on the left and a gentle fill from the right. Photoreal corporate portrait photography, 85mm lens, shallow depth of field, natural skin texture.
A head-and-shoulders portrait, the subject centered and looking just off camera: a man in his forties with close-cropped gray hair, light stubble and laugh lines, wearing a navy crew-neck sweater, for a company team page. Behind him, a softly blurred open-plan office with warm lights. Soft key light from a window on the left and a gentle fill from the right. Photoreal corporate portrait photography, 85mm lens, shallow depth of field, natural skin texture.
Measured at 1024 × 1024:
| Steps | Time | Cost against the default |
|---|---|---|
| 50 (default) | about 8 s | 1× |
| 28 | about 6 s | about 0.6× |
| 16 | about 4 s | about 0.4× |
The three portraits are hard to tell apart, down to the stubble and the laugh lines. Drop to 28 or fewer for drafts and check your own subject before lowering the step count for final renders. If you need neither guidance nor a negative prompt, FLUX 3 Image Klein is faster and cheaper again at its default settings.
Tips for best results
-
Start at the default guidance. 3.5 follows a full caption closely. Raise it for more contrast, not for more accuracy.
-
Keep guidance above 1. At a
CFGScaleof 1 the prompt is followed loosely andnegativePrompthas no effect. -
List nouns in the negative prompt. Name the props the way you would point at them in the frame: "plants, cutting boards, cookbooks".
-
Handle text in the caption. A negative prompt won't clear a label. Describe the surface as plain, or quote the words it should carry.
-
Use fewer steps for drafts. 28 steps costs a little over half the default and looked the same as the default on the portrait above.
-
Reuse the Klein guides. The base takes the same inputs as the distilled model, so the editing, multi-reference, bounding box and text guides for FLUX 3 Image Klein apply here as well.