How to Use Flux AI in 2026: Online, in ComfyUI and With the BFL API

You can use Flux in three ways: in a browser studio such as Kenerate AI, locally with ComfyUI or Python diffusers on your own GPU, or through Black Forest Labs' (BFL) API. The quickest start is the browser: pick a Flux model, type a prompt that leads with the subject, choose a size and generate. Run it locally if you have a GPU with roughly 13 GB of VRAM or more and want full control. Use the API when you are building Flux into your own app.

Flux is the image model family from Black Forest Labs. "Flux" now covers two generations: FLUX.1 (dev, schnell and the older pro models) and FLUX.2 (dev, pro, flex, max and the compact klein models). This guide walks through each route step by step, explains which licence covers which model, and lists the mistakes that waste the most time. Model facts come from BFL's own model cards and docs, read on 23 October 2026. Kenerate facts come from our model config. Kenerate AI is our product.

Which Flux model should you use?

Start with the job, not the name. Fast drafts and free tests suit the small, distilled models. Finished images with fine detail or text suit FLUX.2 pro, flex or max. Local use depends on your GPU and on the licence.

Two points from the model cards are easy to miss. First, the "non-commercial" licences cover the weights. The FLUX.1 [dev] and FLUX.2 [dev] cards both say outputs can be used for personal, scientific and commercial purposes "as described in" the licence, so read the licence text before you ship. Second, FLUX.2 [klein] 4B is the only FLUX.2 model on the list under Apache 2.0.

How do the three ways to use Flux compare?

How do you use Flux on Kenerate AI, step by step?

1. Open the image studio and sign in. New accounts get 25 free credits.
2. Open the model picker and choose the FLUX family. On the free plan you can use Flux 2 Dev, Flux 2 Klein, Flux 2 Turbo and Flux 2 Flash. The other Flux models need a credit pack.
3. Write your prompt. Put the subject first, then the action, style and setting (more on this below).
4. Pick a size. Every Flux model offers square, portrait, landscape, 16:9, 9:16 and HD sizes up to 1536 × 1536. Flux 2 Dev, Pro, Flex and Max add 2K sizes (2048 × 2048, 1536 × 2048 and 2048 × 1536).
5. Optional: set a seed so you can repeat a result while you change one word at a time. Every Flux model on Kenerate supports a seed.
6. Optional: add reference images to edit or combine pictures (see the next section).
7. Generate, then upscale the keeper if you need a bigger file.

Here is what each Flux model costs on Kenerate, from our model config and credit calculator:

With the 25 sign-up credits, that is eight Flux 2 Dev images. Credit packs start at $15 for 1,300 credits on the pricing page. For a deeper look at the FLUX.2 family, see our Flux 2 model page.

How do you edit images with Flux references?

Every FLUX.2 model can edit and combine images. On Kenerate, upload up to 3 reference images (4 on Flash and Turbo), then describe only the change and what must stay. For example: "Turn the pencil sketch into a photoreal modern house at sunset, same layout, same windows."

BFL's FLUX.2 prompting guide recommends giving each reference a job, such as subject, style or background, and naming that job in the prompt. The two FLUX.1 models work differently: they take one image and remix it with a strength slider. Low strength keeps more of the original.

How do you run Flux locally in ComfyUI?

ComfyUI is the easiest local route because it ships ready-made Flux templates.

1. Install or update ComfyUI. The ComfyUI docs say missing nodes usually mean an outdated version.
2. Open Workflow Templates and pick the Flux template you want, for example FLUX.2 Dev.
3. Download the model files the template lists into the matching folders. For FLUX.2 [dev], the ComfyUI tutorial lists flux2_dev_fp8mixed.safetensors in models/diffusion_models, mistral_3_small_flux2_bf16.safetensors in models/text_encoders and flux2-vae.safetensors in models/vae.
4. Accept the model licence on Hugging Face where the repo asks for it. The FLUX.2 [dev] card says access requires accepting the FLUX Non-Commercial License and BFL's Acceptable Use Policy.
5. Type your prompt in the text node, set width and height, and click Run.

For the klein 9B model in particular, our FLUX.2 klein 9B ComfyUI workflow covers the files and settings in detail.

How do you run Flux with Python (diffusers)?

If you prefer code, Hugging Face diffusers has a pipeline for each generation. The FLUX.1 [schnell] model card uses FluxPipeline with 4 steps and guidance_scale=0.0:

pythonCopy
import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()  # saves VRAM; remove it on a big GPU
image = pipe("a red fox in fresh snow, birch forest, soft winter light",
             guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256).images[0]
image.save("fox.png")

The settings differ per model, and copying the wrong ones is a common reason for bad results:

The klein card notes that Flux2KleinPipeline needs diffusers installed from the GitHub main branch.

How much VRAM does Flux need?

The model cards give these figures:

• FLUX.2 [klein] 4B: fits in about 13 GB of VRAM, listed for RTX 3090 / 4070 and above.
• FLUX.2 [klein] 9B: fits in about 29 GB, listed for RTX 4090 and above.
• FLUX.2 [dev]: 32 billion parameters. The card's example runs a 4-bit quantized version on an RTX 4090 and points to the diffusers docs for consumer GPUs.
• FLUX.1 [schnell] / [dev]: 12 billion parameters. enable_model_cpu_offload() trades speed for lower VRAM.

If your card is smaller than that, quantized (fp8 or 4-bit) files or a browser studio are the practical options.

How do you use the Flux API from Black Forest Labs?

1. Create an account and an API key in BFL's dashboard.
2. Send a POST request to https://api.bfl.ai/v1/<model>, for example flux-2-pro, with your key in the x-key header and a JSON body with prompt, width and height.
3. The response gives you an id and a polling_url. BFL's quick-start says to always poll that URL rather than build your own.
4. Poll until the status is Ready, then read the image URL from result.sample.
5. Download the image straight away. BFL says the signed result URLs are valid for 10 minutes and are not meant to be served to users directly.

BFL's quick-start lists up to 24 active tasks per account, with a 429 error above that, and a 402 error when you run out of credits.

How should you write a Flux prompt?

BFL's FLUX.2 guide recommends the order Subject + Action + Style + Context, because earlier words carry more weight. It suggests 30 to 80 words for most images. FLUX.2 also reads structured JSON prompts and hex colour codes tied to a named object, and it renders text you put in quotation marks.

A simple example: "A ceramic coffee mug on a walnut desk, steam rising, minimalist product photo, soft window light from the left, the word 'MORNING' printed on the mug in black serif letters."

For 40 ready-made examples, see our Flux prompts guide.

What are the most common Flux mistakes?

1. Writing a negative prompt. BFL states that FLUX.2 does not support negative prompts, and none of the Flux models on Kenerate take one. Describe what you want instead: "sharp focus throughout" rather than "no blur".
2. Burying the subject. A long style preamble before the subject weakens it. Lead with the subject.
3. Wrong steps or guidance for the model. Schnell and klein are distilled for about 4 steps. Running them like dev (50 steps, guidance 3.5) wastes time and can look worse.
4. Assuming the weights licence allows anything. Dev and klein 9B weights are non-commercial. Read the licence before you build a paid product on them.
5. Unclear reference roles. With several references, say which one is the subject, the style and the background.
6. Changing everything at once. Fix the seed and change one thing per run so you can see what helped.
7. Letting API results expire. BFL's result URLs last 10 minutes, so save the file in the same script.

Bottom line

To use Flux, pick the route that fits you. A browser studio needs no setup and lets you compare models. Kenerate runs 11 Flux versions, with Flux 2 Dev, Klein, Turbo and Flash on the free plan. ComfyUI or diffusers suit you if you have a 13 GB+ GPU and want LoRAs, privacy and unlimited runs. BFL's API suits developers. Whichever you pick, lead the prompt with the subject, skip negative prompts, and match steps and guidance to the model. You can also start straight from our Flux AI image generator page.