업데이트
What is FLUX 3 Image Klein Base?
FLUX 3 Image Klein Base (BFL writes it "FLUX 3 [image klein] base") is the undistilled checkpoint of Black Forest Labs' FLUX 3 Klein image model, released on October 15, 2026 alongside the distilled FLUX 3 Image Klein. Both share one 5.6B-parameter image transformer and a Qwen3-VL-4B-Instruct text encoder, and both are open weights under the FLUX Kommunity License. What Base keeps is the full sampling loop: 50 steps by default, classifier-free guidance you can set from 1 to 20, and a negative prompt. BFL publishes it for fine-tuning, LoRA training and research, and it is not on BFL's own API. Here it runs as a hosted endpoint, so you get the same controllable sibling without a GPU. The distilled model is roughly three times faster and cheaper at the same size; default to it for production volume and use Base when you need the knobs.
Key features of FLUX 3 Image Klein Base
Undistilled 5.6B checkpoint, open weights
The same transformer and text encoder as the distilled Klein, before distillation. This is the checkpoint BFL publishes for fine-tuning and LoRA training, so a LoRA you train locally matches what you call here.
50-step sampling, 1 to 50 on the dial
Base samples for 50 steps by default and lets you set steps anywhere from 1 to 50. More steps buy fine detail at the cost of latency; fewer steps make cheaper drafts.
CFG guidance and a negative prompt
guidance runs 1 to 20 (default 3.5) and negative_prompt lists what to keep out. The distilled model has neither. Both are ignored when guidance is 1 or lower.
Text-to-image, editing and up to 4 references
Pass up to four image_urls to edit an image or to anchor a product, character or palette across a batch. Reference runs take longer and cost more than a text-only call.
Seven aspect ratios, four sizes up to 4 MP
21:9 through 9:21 on one model, at 0.6mp, 1mp, 2mp or 4mp. Native 4K training means the 4mp preset is a real output size, not an upscale.
Reproducible with a seed
An integer seed plus the same prompt, references and settings returns the same image. Change one setting at a time and compare.
What FLUX 3 Image Klein Base is best for
Guidance and negative-prompt experiments
Base is the only FLUX 3 Klein that exposes CFG. Use it to find the guidance value a prompt family wants before you lock a pipeline.
Matching a LoRA you trained on Base
If you fine-tuned on the Base weights locally, this endpoint gives you the same sampler behaviour without hosting the model.
Final renders at 2 MP and 4 MP
Fifty steps and native 4K training suit hero crops and print-size output where the extra seconds are fine.
Reference-anchored product and character sets
Up to four references keep geometry, identity and palette steady across a batch.
Research and side-by-side evaluation
Same architecture as the distilled model, so you can measure what distillation changed on your own prompts.
Hard prompts that drift
When the distilled model ignores part of a long prompt, raising guidance on Base is the cheapest fix to try.
FLUX 3 Image Klein Base use cases
Build a prompt tuning tool that runs one prompt on Base across a guidance sweep (2, 3.5, 5, 7) with a fixed seed, shows the four results in a row, and writes the winning value into your production config for the distilled model. Render a product catalogue with up to four reference shots per SKU so every angle keeps the same materials and colour. Produce 4 MP hero images for print and landing pages from a fixed seed, then regenerate at 1 MP for thumbnails without re-prompting. Run an evaluation set on both Klein models and log where the undistilled checkpoint earns its extra latency.
What FLUX 3 Image Klein Base can do
Editorial portrait with film-style grain
Fifty steps at the default 3.5 guidance hold skin texture, freckles and silver streaks in the hair with a shallow depth of field. This is the kind of fine detail that extra sampling steps buy.
Clean studio product render
One hard key light, one long soft shadow, a plain dark grey backdrop. Base keeps the matte ceramic and walnut grain readable without over-sharpening, which matters for catalogue and ad work.
When to use FLUX 3 Image Klein Base
Start with the distilled model. It produces the same style of image in 16 steps, and on our route it is roughly three times faster and cheaper at every size. Switch to Base in three cases. First, when you want to steer: raise guidance for a prompt the distilled model keeps half-following, or add a negative prompt to clear text and watermarks. Second, when you trained a LoRA on the Base weights locally and want the hosted output to match your test runs. Third, when a final 2 MP or 4 MP render is worth 50 steps. Billing is per image by resolution, reference runs cost more, and Base lands at about three times the distilled rate for the same size, so keep it for the shots that need it. If you want more capacity rather than more control, FLUX 2 Klein 9B and FLUX 2 Flex are the larger options in the family.
FLUX 3 Image Klein Base API examples
Call FLUX 3 Image Klein Base from any language by POSTing to /v1/tasks. Full parameter docs live at docs.unifically.com/models/image/flux/flux-3-image-klein-base.
curl -X POST https://api.unifically.com/v1/tasks \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "black-forest-labs/flux-3-image-klein-base",
"input": {
"prompt": "Studio product shot of a matte black ceramic pour-over coffee dripper on a walnut slab, single hard key light, dark grey studio background",
"negative_prompt": "text, watermark, blurry",
"aspect_ratio": "16:9",
"resolution": "1mp",
"guidance": 3.5
}
}'
Successful submission returns a task_id. Poll GET /v1/tasks/<task_id> or set a callback_url on the request to receive the finished image URL.
FAQ
FLUX 3 Image Klein Base에 대해 자주 묻는 질문
Both share the same 5.6B-parameter image transformer and the same Qwen3-VL-4B-Instruct text encoder. Base is the undistilled checkpoint. It samples for 50 steps by default and keeps classifier-free guidance and a negative prompt. The distilled FLUX 3 Image Klein runs 16 steps with no guidance control and is roughly three times faster and cheaper at the same size. Base is the one Black Forest Labs publishes for fine-tuning, LoRA training and research.
Leave it at 3.5 for most prompts. Raise it toward 5-7 when the model drifts from a detailed prompt and you want it to follow the text more closely, at the cost of a harsher, more saturated look. Drop it toward 1.5-2.5 for softer, more natural images when the prompt is short. Values run from 1 to 20.
It lists what you want kept out of the image, such as text, watermark, blurry. It only works through classifier-free guidance, so it is ignored when guidance is 1 or lower. Keep it short. Long negative lists tend to flatten the image more than they clean it up.
More steps mean more sampling passes and a longer, more expensive run. The default of 50 is where Base looks best. Dropping to 20-30 cuts latency for drafts but softens fine detail. Values run from 1 to 50. If you want few-step output as a habit, use the distilled FLUX 3 Image Klein instead; it was trained for 16 steps.
Yes. Set image_urls to up to four image URLs. The model uses them for editing and multi-reference generation, so you can keep a product, a character or a palette steady across a run. A reference run takes longer and costs more than a plain text-to-image call.
Seven aspect ratios (21:9, 16:9, 4:3, 1:1, 3:4, 9:16, 9:21, default 16:9) and four resolution presets (0.6mp, 1mp, 2mp, 4mp, default 1mp). Base was trained at native 4K, so the 4mp preset is a real output size, not an upscale.
On our route, about 8-10 seconds for a 1 MP text-to-image call at 50 steps, about 15 seconds with a reference image, and about 24 seconds at 4 MP. Fewer steps bring those numbers down.
Yes. Pass an integer seed. The same prompt, negative prompt, references, aspect ratio, resolution, steps, guidance and seed return the same image, which makes A/B tests on a single setting clean.
