Skip to content
Claudexia TeamGUIDE

Stable Diffusion and Flux locally: licenses, VRAM, ComfyUI

Run Stable Diffusion and Flux on your own PC as of 17 November 2026: which models exist, licenses, VRAM per the docs, setup in ComfyUI and diffusers.

To run Stable Diffusion and Flux yourself, use ComfyUI or the diffusers library: download the weights from Hugging Face, accept the license and generate on your GPU. The latest versions as of 17 November 2026 are Stable Diffusion 3.5 and FLUX 3 Image Klein, each with its own license and its own memory floor. Below is what the documentation says, with no measurements of our own.

What exists on 17 November 2026

Stability's Stable Image page calls Stable Diffusion 3.5 the company's most powerful image model and lists three variants: Large for professional work at 1 megapixel, Turbo with four steps, and Medium for ordinary hardware. Stability's 2026 news has no entries about new image models. Stable Audio 3.0 came out on 20 May 2026.

Black Forest Labs has three generations of models: FLUX.1, FLUX.2 and FLUX 3. The first two shipped earlier. The company introduced FLUX 3 on 23 July 2026 as video, images and audio in one model, in early access through the API with closed weights. Two members of the FLUX 3 family have open weights: FLUX 3 Action with 7 billion parameters (for controlling robots) and FLUX 3 Image Klein with 5.6 billion. Klein came out on 15 October 2026 for generating and editing images, and it accepts up to 10 references. The 5.6 billion figure covers the transformer, while the text encoder is counted separately.

Licenses: what the model cards say

ModelSizeLicenseWhat the card saysTerms on Hugging Face
Stable Diffusion 3.5 Large, Mediumnot stated on the cardStability Community Licensefree for research, non-commercial use, and commercial use under $1 million annual revenue; above that, an Enterprise licensemust be accepted
FLUX.1 [schnell]12 billionApache-2.0personal, scientific and commercial purposesmust be accepted
FLUX.1 [dev]12 billionFLUX.1 [dev] Non-Commercial Licenseoutputs may be used for personal, scientific and commercial purposes under the license termsmust be accepted
FLUX.2 [klein] 4B4 billionApache-2.0open weights for commercial usenot required
FLUX.2 [klein] 9B9 billionFLUX Non-Commercial Licensenon-commercial licensemust be accepted
FLUX.2 [dev]32 billionFLUX Non-Commercial Licenseoutputs may be used for personal, scientific and commercial purposes under the license termsmust be accepted
FLUX 3 Image Klein and Base5.6 billionFLUX Kommunity License v1.0non-commercial use, plus commercial generation of outputs for organizations with revenue under $5 million, affiliates countedmust be accepted

The older Stable Diffusion XL 1.0 and 1.5 sit under CreativeML Open RAIL++-M and CreativeML OpenRAIL-M. We did not go through their license terms. The SDXL 1.0 base and SD 1.5 repositories on Hugging Face are open with no terms to accept. The last column of the table above shows the gated flag on the repositories as of 18 November 2026.

What matters in the license texts themselves:

  • The Stability Community License requires you to register with Stability for commercial use. The $1 million annual revenue threshold is counted over you and your affiliates, separately or together, and all revenue counts, including revenue unrelated to Stability's models. Above the threshold the license ends and you have to ask for a new one. If you distribute the model or a derivative, attach the license and show "Powered by Stability AI". You may not use outputs to create or improve foundation generative models, other than the licensed models themselves and their derivatives.
  • The FLUX 3 Image Klein card lists what section 4 of the license prohibits: military purposes, surveillance, biometric processing, illegal and fraudulent activity, and using the model or its outputs to improve another model with a similar function. The commercial exception, the card stresses, does not allow commercial fine-tuning or distribution.
  • The ComfyUI page on Klein adds conditions for organizations with revenue under $5 million: filter or review the outputs, and disclose that an image is AI-generated where the law requires it. Any other commercial use, commercial fine-tunes included, needs a license from Black Forest Labs.
  • Both vendors mention trade restrictions. Stability requires compliance with Trade Control Laws, and the Klein card says the license covers export controls and sanctions.

This is a retelling of the texts. We give no legal assessments and do not promise that any method is lawful or safe for your account. You read and accept the vendor's terms.

VRAM requirements: only what the vendors write

What we runWhat the documentation saysSource
FLUX 3 Image Klein, INT8 in ComfyUI or DiffusersNVIDIA GeForce RTX cards with 8 GB, example RTX 3070BFL blog of 15 October 2026, README
FLUX 3 Image Klein, reference codepeak memory on a B200 at 1024×1024 without the upsampler: 31 GB in BF16, 26 GB in FP8, 25 GB in NVFP4; the upsampler adds about 28 GB on an 80 GB cardHugging Face card
FLUX.2 [klein] 4Bfits in about 13 GB, RTX 3090 and 4070 and upHugging Face card
FLUX.2 [klein] 9Babout 29 GB per the card, RTX 4090 and upHugging Face card
Stable Diffusion 3 and 3.5 in diffusershard on cards under 24 GB even in fp16 without memory-saving tricksdiffusers documentation

A few notes. The Black Forest Labs figures of 8 GB and 31 GB were obtained under different conditions: the first is for INT8 files in ComfyUI, the second for the reference code on a B200 card. The pages we read do not say how much memory the INT8 build needs together with the upsampler. FP8 in the reference code works on Ada, Hopper and Blackwell cards, and for the RTX 30 series and older BFL recommends the INT8 files in ComfyUI. For Stable Diffusion 3.5 the Hugging Face cards give no memory figures. They do show an example of 4-bit NF4 quantization through bitsandbytes, and diffusers describes offloading parts of the model to the CPU, dropping the T5 text encoder and quantizing it to 8 bits.

The files are big too. Per the README, the fast Klein model with the upsampler takes 52 GB, and the whole collection 107 GB. The upsampler alone weighs 18.8 GB in BF16 and 9.8 GB in INT8, according to the card.

Software requirements per the documentation: ComfyUI runs on Windows (NVIDIA or AMD cards recommended, Windows 10 or newer needed), on macOS with Apple Silicon chips and on Linux, and the recommended Python is 3.13. The diffusers installation page says Python 3.10+ and PyTorch 2.6+, but the metadata in setup.py on the main branch (version 0.44.0.dev0) and in release 0.43.0 of 12 November 2026 says Python 3.11 or newer, so take 3.11 or higher. It runs on cuda, xpu, mps (Apple Silicon) and cpu, and the documentation gives no speed for the CPU. According to its README, AUTOMATIC1111 WebUI asks for Python 3.10.6 on Windows. Its last release, v1.10.1, came out on 9 February 2025, the last change in the repository dates from 2 March 2026, and the README mentions neither FLUX nor Stable Diffusion 3.5.

Running through ComfyUI

ComfyUI Desktop exists for Windows, for macOS on Apple Silicon and for Linux. The ComfyUI page on Klein gives this order: update ComfyUI to the latest version, download the weights from Hugging Face and load the FLUX 3 Image Klein template. The model goes through standard nodes: Load Diffusion Model, Load CLIP with type flux3 for the Qwen3-VL 4B text encoder, and Load VAE. No API key is needed.

The black-forest-labs/flux-3-image-klein-comfy repository on Hugging Face is built for ComfyUI and holds BF16, INT8 and NVFP4 files for both models. The card says the INT8 files are meant for cards with 8 GB such as the RTX 3070, and for cards without FP8, which means the RTX 30 series and older. First sign in and accept the terms on the repository page. Then download the files you need straight into the ComfyUI folders with the command from the card:

hf auth login

hf download black-forest-labs/flux-3-image-klein-comfy \
  diffusion_models/flux3-image-klein-int8-convrot.safetensors \
  text_encoders/qwen3vl_4b_int8_convrot.safetensors \
  vae/flux2-vae.safetensors \
  text_encoders/qwen3.5_9b_flux3_prompt_upsampler_int8_convrot.safetensors \
  --local-dir ComfyUI/models

File sizes from the card's table: INT8 transformer 5.8 GB, encoder 4.9 GB, upsampler 9.8 GB, VAE 0.3 GB. The card says to always run the upsampler, and its file in the card's download command is labeled required. The model was trained on full scene descriptions, so a short request without the upsampler gives "generic output". The workflows run the upsampler by default. The enable upsampler switch turns it off, but the card allows that only when the request is already a full scene description.

Running through diffusers

You need Python 3.11 or newer and PyTorch. Install PyTorch with the selector at pytorch.org for your GPU and system, which is where the diffusers installation page sends you. Everything else takes one command. The diffusers package pulls in neither torch, nor transformers, nor accelerate. The examples below import transformers, and enable_model_cpu_offload without accelerate fails with an ImportError. Sentencepiece and protobuf are there for the T5 tokenizer of SD 3.5, and we did not check whether your version of transformers needs them. The commands come from the documentation and the model cards, and we did not run them.

pip install transformers accelerate sentencepiece protobuf
pip install git+https://github.com/huggingface/diffusers.git

The second command installs the main branch, as the FLUX.2 [klein] 4B card does.

The SD 3.5 repository is gated: first accept the terms on the model page, then sign in with hf auth login. The FLUX.2 [klein] 4B repository is open, so there are no terms to accept. This model is under Apache-2.0 and, per the card, fits in about 13 GB. The example is taken from the card:

import torch
from diffusers import Flux2KleinPipeline

device = "cuda"
dtype = torch.bfloat16

pipe = Flux2KleinPipeline.from_pretrained("black-forest-labs/FLUX.2-klein-4B", torch_dtype=dtype)
pipe.enable_model_cpu_offload()

image = pipe(
    prompt="A cat holding a sign that says hello world",
    height=1024,
    width=1024,
    guidance_scale=1.0,
    num_inference_steps=4,
    generator=torch.Generator(device=device).manual_seed(0),
).images[0]
image.save("flux-klein.png")

For Stable Diffusion 3.5 Large the card gives a short example. The diffusers documentation recommends CPU offload for ordinary hardware and says the memory-saving tricks for SD3 also fit 3.5. Below is an example modeled on the documentation, with the model id swapped for 3.5 Large:

import torch
from diffusers import StableDiffusion3Pipeline

pipe = StableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3.5-large",
    dtype=torch.float16,
)
pipe.enable_model_cpu_offload()

image = pipe(prompt="a photo of a cat holding a sign that says hello world").images[0]
image.save("sd35.png")

We did not run either snippet. FLUX 3 Image Klein has a modular pipeline in diffusers. There the upsampler, a separate model with 9 billion parameters, loads together with the other components and runs on every request, so Klein is heavier on memory than its 5.6 billion size suggests.

If you have no GPU of your own

Black Forest Labs sells API access. The distilled Klein is called flux-3-image-klein, and the price starts at $0.009 per 1K image, per the company blog of 15 October 2026. There is also a browser sandbox. On its Klein page Comfy lists Comfy Cloud with five free runs without a card, and a Standard plan at $20 a month. Terms and prices change, so check the vendor pages. For downloaded weights that you run yourself or in Comfy Cloud, the license stays the same, as that Comfy page says: the license governs commercial use wherever you run the model. Access through the Black Forest Labs API comes under separate service terms, and the weights license does not carry over to it. Read the vendor's API terms.

Images through the Claudexia API

Stable Diffusion and Flux are not in the Claudexia catalog right now (per the live price list on 17 November 2026). On the same date the Premium channel has the image models Nano Banana, Nano Banana 2, Nano Banana Pro, GPT Image 2, GPT Image 2.5 and GPT Image 2.6, and the Cheap channel has Nano Banana Pro and three GPT Image versions. Payment in Claudexia works in rubles through SBP (Russia's Faster Payments System), with a Russian card or in crypto, and the minimum top-up is $1. See the current lineup on the models page. If you need an image from code without a GPU, the API comparison is in our article on choosing between GPT Image and Nano Banana as an image API.

We have separate reviews of the models themselves: Nano Banana Pro, GPT Image 2 and Qwen Image 2.1, which has open weights. Free routes are collected in our piece on Midjourney alternatives, and running text models on your own machine is covered in our Ollama guide (installation on Windows, macOS and Linux, and choosing a model). Neighboring reviews in the group: Runway, Higgsfield and HeyGen.

FAQ

Is Stable Diffusion free?

The weights are free for research and non-commercial use. For commercial use they are free under the Stability Community License while annual revenue stays under $1 million: you and your affiliates are counted together, and all revenue counts, even revenue unrelated to the models. Commercial use requires registering with Stability. Running it yourself costs no license fee. You pay for electricity and the GPU.

Which should I pick, Stable Diffusion or Flux?

We will not name a best model. We have no measurements of our own, and the comparisons from Black Forest Labs are internal. Choose by terms instead: do you need a free commercial license with no revenue threshold (FLUX.2 [klein] 4B and FLUX.1 [schnell] under Apache-2.0), is there enough memory, and does the license fit your income.

How much VRAM do I need?

It depends on the model and the build. For FLUX 3 Image Klein INT8 in ComfyUI the vendor says 8 GB, FLUX.2 [klein] 4B needs about 13 GB, and for SD3 diffusers warns about 24 GB without memory-saving tricks. Details and sources are in the table above.

Can I use the outputs in ads and client work?

It depends on the model's license and your revenue, so see the table and the license texts. This is a retelling of the terms, not legal advice.

Why does Hugging Face ask me to accept terms?

For gated repositories (SD 3.5, FLUX.1, FLUX.2 [klein] 9B, FLUX.2 [dev], FLUX 3 Image Klein) the page is visible, but the files open only after you accept the license and sign in to your account. For FLUX.2 [klein] 4B the gated flag was off as of 18 November 2026, so no agreement is needed there.