Short answer: FLUX 3 Image Klein costs $0.009 per 1K image through BFL’s direct API as of October 22, 2026. BFL also lists $0.008 at 768, $0.011 at 1.5K, $0.014 at 2K, and $0.036 at 4K. These are direct API prices per output, not a subscription or GlobalGPT credit rate.
For local use: BFL’s optimized INT8 build can run on supported 8GB RTX cards through ComfyUI or Diffusers. That number is for inference, not fine-tuning; the text encoder, VAE, prompt upsampler, image size, and other loaders can need more VRAM.
For capabilities: FLUX 3 Image Klein is a 5.6B-parameter model for text-to-image generation and editing. BFL lists up to 10 reference images, native 4K output, and bounding-box control; the model card warns that boxes guide placement but do not lock pixels. Choose it for fast drafts, local workflows, or controlled reference-image work, then check the license before commercial use.
After generation, you may still need to fix a detail, rewrite a prompt, or prepare the surrounding copy: Klein’s bounding boxes guide placement but do not lock pixels, and local workflows add setup. For that next step, GlobalGPT puts image creation, editing, and writing in one dashboard. Explore its image-editing workflow for revisions after a Klein draft.
What is FLUX 3 Image Klein?
FLUX 3 Image Klein is BFL’s 5.6B-parameter image model, separate from FLUX.2 [klein] and standard FLUX 3 Image.
BFL released the model on October 15, 2026. It supports text-to-image generation, editing, up to ten reference images, native 4K output, and bounding-box control. The ficha del modelo says boxes guide placement and size but do not lock pixels; edits can change areas outside the box.
Choose a route for your task
Try it online
Open BFL’s Parque infantil and check whether Klein is available in your account.
Distilled weights
Get the distilled weights for local generation. The 8GB claim requires BFL’s optimized setup.
Base weights
Utilice el base weights for fine-tuning. The 8GB inference claim does not cover training.
How much does FLUX 3 Image Klein cost?
De BFL direct API prices vary by output size. Text-to-image and image editing have the same listed rate; reference images cost nothing extra. “1K” means about one megapixel, not 1,000 images.
| Output tier | Approx. pixels | Klein / image | 100 imágenes |
|---|---|---|---|
| 768 × 768 | 0.6 MP | $0.008 | $0.80 |
| 1K | 1 MP | $0.009 | $0.90 |
| 1,5 K | 2 MP | $0.011 | $1.10 |
| 2K | 4 MP | $0.014 | $1.40 |
| 4K | 16 MP | $0.036 | $3.60 |
At 1K, standard FLUX 3 Image costs $0.048 per image through BFL’s API; Klein costs $0.009. Price does not measure image quality.
These are BFL API prices, not GlobalGPT credits or subscription rates. Local use adds hardware and operating costs. Compare the cost per usable image, including retries.
Can it run locally on an 8GB GPU?
De BFL 8GB claim applies to optimized INT8 inference on supported NVIDIA RTX cards using ComfyUI or Diffusers. Other loaders, larger images, and supporting components may need more VRAM.
- To try Klein online, use BFL’s Playground or the Comfy Cloud template. Comfy Cloud bills separately.
- For local use, install ComfyUI 0.41.1 or newer and follow the official tutorial. Sign in to Hugging Face and accept access to the gated Comfy weights.
- Load the distilled INT8 template with its Qwen3-VL text encoder and FLUX.2 VAE. Prompt upsampling is on by default. Start at 1K and 16 steps; record peak VRAM and time.
- Test text, object count, and a one-change edit. Use the base weights for fine-tuning; the 8GB inference claim does not cover training.
En ficha del modelo reports 31GB peak VRAM with BF16, 26GB with FP8, and 25GB with NVFP4 for BFL’s reference code on a B200 at 1024 × 1024, with prompt upsampling off. With the full upsampler, memory use rose by about 28GB in BFL’s test on an 80GB GPU. Those results use a different pipeline from the optimized 8GB INT8 setup. BFL recommends upsampling short prompts to avoid generic images.
Our prompt and one-image result
Make a 3:2 studio product image of a fictional cobalt-blue glass bottle on a white shelf. Put the bottle on the left and leave clean space on the right for a headline. Add only the short word “NORTH” on the label. Soft daylight; no extra bottles or hands.
What do the benchmarks actually show?
En BFL’s own comparison, Klein scored 91 points higher in text-to-image preference and 70 points higher in editing than FLUX.2 [klein] 9B, with 38% fewer parameters in the image generator. NVIDIA reported about four seconds for a 1MP image on an RTX 5090 and under nine seconds on an RTX 4070 using distilled INT8 ComfyUI at 16 steps. These are vendor tests, not our results or end-to-end API times.
En Kingy’s Fal test, Klein distilled passed all four checks on 24 of 32 text-to-image outputs. The 32 generations and 12 edits cost $0.396 at those settings. Kingy turned off prompt expansion and used automated scoring, so 24/32 is not a general success rate.
What early users are saying
En un local-use post, Shikhar reported about 13 seconds per image with a distilled build, but gave no full GPU, resolution, or settings record. Martin Ronfort welcomed lower-precision weights for easier local use; that is an opinion, not a speed test.
Otro user reported missing limbs and fingers; a reply said the Comfy workflow had an error being fixed. This single report does not show how often the issue occurs.
What does “open weights” allow?
FLUX 3 Image Klein weights use BFL’s FLUX Kommunity License. The ficha del modelo permits qualifying users whose annual revenue, including affiliates, is under $5 million to use outputs commercially. That exception does not automatically cover commercial fine-tuning, distributing weights, or self-hosting a service. Larger companies can use BFL’s API or ask about a self-hosted license. Check the current license terms before using local weights in a product or client project.
PREGUNTAS FRECUENTES
How much does FLUX 3 Image Klein cost per image?
BFL’s direct API lists $0.009 for a 1K image as of October 22, 2026. The rate changes with output resolution.
Where can I download FLUX 3 Image Klein weights?
Use BFL’s gated Hugging Face repositories. Sign in and accept the access terms before downloading the distilled, ComfyUI, or base weights.
Can I run FLUX 3 Image Klein in ComfyUI?
Yes. Follow the official ComfyUI 0.41.1+ tutorial and use the gated Comfy weights. Comfy Cloud is billed separately.
Can it edit existing images?
Yes. BFL lists image editing and up to ten reference images. Bounding boxes guide placement but do not create a hard edit mask.
Which checkpoint should I use for fine-tuning?
Use the base checkpoint for adaptation. The step-distilled checkpoint targets faster inference.
Can I use FLUX 3 Image Klein outputs commercially?
Qualifying users under $5 million in annual revenue may use outputs commercially under BFL’s FLUX Kommunity License. Commercial self-hosting has separate terms.
Are failed BFL API generations charged?
No. BFL says rejected requests, failed generations, and outputs withheld by moderation are not charged.
Does the 8GB VRAM claim cover fine-tuning?
No. The 8GB figure covers optimized INT8 inference on supported RTX hardware, not training.
