How Much To Run AI
Open weights

FLUX 3 Image Klein: API pricing and local setup

FLUX 3 Image Klein is Black Forest Labs’ open-weight image generator and editor, released October 15, 2026. Use its hosted API for a quick start at $0.009 per 1K image, or choose a local ComfyUI or Diffusers workflow. The 5.6B figure counts the image transformer; the Qwen3-VL-4B text encoder and VAE are additional components.

Content reviewed:

Try the official Klein demo ↗
FLUX 3 Image Klein: a 1K BFL API image costs $0.009; local distilled inference uses 16 steps, with encoder and VAE in addition to the transformer.
Route and cost illustration. Local memory depends on the complete workflow.

Use this model

API docs and console

Official model repository

Local tools and guides

Use cases

  • Create a product composition from a detailed scene description.
  • Combine up to ten reference images, assigning each a subject or style role.
  • Edit or place an element with bounding-box guidance, then inspect the surrounding pixels.

Start with a 1K API image

Create a BFL account and API key, then use /v1/flux-3-image-klein for the distilled model. Try the official Hugging Face demo for a browser preview, or use the BFL dashboard for the hosted service. Begin with one composition at 1K; add reference images once the basic request works. The response supplies a polling_url; retrieve the asynchronous result through that URL and save the completed image. Its cost field is in credits. Keep the request record with the downloaded image.

Klein supports generation, reference editing and layout control. Describe the subject, setting, light and framing. For several references, explain each image’s role. Bounding boxes guide placement and size rather than imposing a hard pixel mask, so inspect the finished edit outside the requested region too.

Example BFL request; requires your own paid API access
curl --fail https://api.bfl.ai/v1/flux-3-image-klein \
  -H "x-key: $BFL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"prompt":"A studio photograph of a blue ceramic cup on a pale wood desk, soft window light from the left, centered composition","resolution":"1k"}'

BFL API pricing per image

BFL charges the same resolution rate for text-to-image and image editing. These USD rates were checked October 16, 2026. The API and Playground use the same credit pricing; one credit equals $0.01.

At 1K, 100 generation attempts cost 100 × $0.009 = $0.90. If 25 images are accepted, generation spend is $0.036 per accepted image. At 4K, those same 100 attempts cost $3.60. Include rerolls in the attempt count rather than budgeting only the final files.

Output tierUSD / image100 attempts
768 × 768$0.008$0.80
1K$0.009$0.90
1.5K$0.011$1.10
2K$0.014$1.40
4K$0.036$3.60

Choose distilled Klein or Klein Base

The ordinary Klein checkpoint uses 16 steps with no classifier-free guidance. It is the generation and editing checkpoint behind the hosted route. Klein Base uses 50 steps with guidance 3.5 and is intended for fine-tuning, LoRA training and research.

For ComfyUI, use the official Comfy package and its matching workflow. Install the transformer, Qwen3-VL text encoder and image autoencoder in their respective folders. The optimized INT8 workflow is the route BFL documents for 8GB GPUs; an 8GB card is not a claim that every precision, reference count and output size fits.

Prepare the ComfyUI files and choose a template

Update ComfyUI before loading the Klein templates. The official guide requires 0.41.0 or newer and identifies 0.41.1 as the backport release. Accept the model repository’s license agreement on Hugging Face, then download the matching component files listed below. Paths are relative to your ComfyUI installation.

Choose “FLUX 3 Image Klein Distilled: Text to Image” in Template Library for the first generation. Its full-checkpoint counterpart is “FLUX 3 Image Klein: Text to Image”; each also has an Image Edit template. Keep the loader and sampler matched to the selected route using the settings below.

FileComfyUI folderUse
flux3_image_klein_sd_int8_convrot.safetensorsmodels/diffusion_models/Distilled generation or editing
flux3_image_klein_base_int8_convrot.safetensorsmodels/diffusion_models/Full checkpoint instead of distilled
qwen3vl_4b_int8_convrot.safetensorsmodels/text_encoders/Both routes; CLIPLoader type flux3
qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensorsmodels/text_encoders/Local caption expansion
flux2-vae.safetensorsmodels/vae/Image encoding and decoding

Keep the sampling settings matched to the template

Use the controls on the loaded workflow. The text-to-image subgraph supplies its sampling values to the inner sampler; the values stored on an internal node are not necessarily the settings used for the run.

Official templateStepsCFG
Full: Text to Image503.5
Distilled: Text to Image16Fixed at 1; no exposed CFG control
Full: Image Edit501
Distilled: Image Edit161

Generate and save the first image

Open the distilled text-to-image template or drag its downloaded JSON onto the canvas. Open the subgraph and confirm its diffusion, encoder and VAE selections. Enter a single-image prompt, choose the square frame in ResolutionSelector and keep the exposed 16-step default for this first run. There is no CFG field to adjust in this text-to-image subgraph.

The distilled text template enables prompt upsampling by default. Leave it on for a short prompt; the extra encoder expands the caption before image generation. BFL’s 8GB guidance applies to its optimized INT8 ComfyUI route, not the BF16 reference run or every canvas and reference count. If local caption expansion does not fit, follow BFL’s separate-caption recipe and supply the expanded caption before disabling the local upsampler.

Click Queue once. SaveImage writes the PNG under ComfyUI/output/. Keep the workflow JSON and seed with the output so you can repeat the configuration. For an edit, open the matching distilled Image Edit template, load one image in LoadImage and describe the finished result. Connect it through Set Reference Multi Latent as <ref_image_0>, keep CFG 1, then compare the saved image with the original.

Example first-generation prompt
A studio photograph of a blue ceramic cup on a pale wood desk, soft window light from the left, centered composition.

Resolve missing nodes or model selections

If the four Klein templates are absent or core nodes are red, update ComfyUI and restart it, then check startup errors for failed imports. A browser refresh alone does not load updated server nodes.

If a loader has no matching file, check the exact filename and folder above, then refresh the model list or restart. A gated download needs the Hugging Face license agreement first. A full checkpoint, distilled checkpoint and FLUX.2 Klein file are different choices; use the workflow for the file you actually downloaded. If generation runs out of memory, check the selected precision, canvas, references and upsampler before changing the sampler.

Read memory and speed with the complete workflow

BFL’s distilled reference-code run on a B200 at 1024 × 1024, with the prompt upsampler disabled, peaks at 31GB in BF16, 26GB in FP8 and 25GB in NVFP4. Base under the same conditions peaks at 31GB, 29GB and 26GB respectively. The separate 9B prompt upsampler adds its own memory and processing work; the reference application enables it by default. These allocations include more than the image-transformer weights.

Hugging Face’s Diffusers authors report 1.93 seconds for text-to-image and 3.87 seconds for one-reference editing at 1024 × 1024. That is a separate H100 80GB test: distilled 16 steps, no offloading, three timed runs after warmup. It describes that implementation, rather than the optimized ComfyUI 8GB route.

An early ComfyUI run on a 16GB GPU

In an October 16 walkthrough, independent author aiaicreate used the official INT8 ComfyUI workflow on an RTX 4060 Ti with 16GB VRAM and 64GB system RAM. Base generation took about 20 seconds, or 34 with the prompt upsampler; distilled generation took about 7 or 17 seconds respectively. The illustrated edit took 40 seconds with Base and 14 with distilled; a two-reference example took 69 and 27 seconds.

The author found the anime result less convincing and showed fingers merging in a cup-holding edit. These are individual examples to inspect, not a measured failure rate. The article does not specify resolution, warmup, repeated trials or peak VRAM, so its timings do not predict a particular 1K or 4K workload.

Prepare captions before calling Diffusers

Use the ModularPipeline recipe linked below for the exact checkpoint. Keep the caption-preparation stage required by the selected implementation. BFL recommends its prompt upsampler because the model was trained on its captions. Prepare that caption first, then pass it and any references to the image pipeline.

Commercial use under the FLUX Kommunity License

Klein uses FLUX Kommunity License v1.0. Its commercial-output exception covers qualifying users below US$5 million in gross annualized revenue, including affiliates, subject to the license’s definitions and restrictions. Larger organizations and other commercial uses need the applicable BFL agreement. This is a model-specific license; the encoder’s Apache-2.0 terms do not replace it.

Choose the image workflow for your project

Use the API when you want to try reference edits without managing components. Use the local distilled checkpoint when its supported workflow and license fit your hardware and data requirements. Choose Base when adapting the model is the goal. For other image pipelines, compare the Qwen-Image guide; the broader FLUX page covers the separate full Image, Video and Action products.

Before you use it

  • Reference images and bounding boxes influence composition; inspect text, identity and pixels around an edit before delivery.

How to get started

Start with one 1K image, save its settings and cost, then compare your accepted-result rate before scaling to larger output or a local deployment.

Sources

Sources (9)

Calculators