If you want to generate AI images on your own PC, two open-weight model families come up again and again. Stability AI's Stable Diffusion XL (SDXL) is the long-standing workhorse, with a huge library of community LoRAs and checkpoints. Black Forest Labs' FLUX.2 [klein] is a newer, compact model family designed for fast generation and editing on consumer graphics cards.
This comparison looks at what each model is, what hardware it needs, how it's licensed and how easy it is to customise. Forge PC hasn't benchmarked or compared the image quality of these models; the details come from the official model cards and Black Forest Labs' documentation, and we haven't included speed figures.
The short answer
| SDXL base 1.0 | FLUX.2 [klein] 4B | FLUX.2 [klein] 9B |
|---|
Developer | Stability AI | Black Forest Labs | Black Forest Labs |
Size | About 3B parameters (Hugging Face listing) | 4B parameters | 9B parameters |
Architecture | Latent diffusion with two text encoders, plus optional refiner | Rectified flow transformer | Rectified flow transformer |
Generation and editing | Generates and modifies images from text prompts | Text-to-image, image editing and multi-reference in one model | Text-to-image, image editing and multi-reference |
VRAM to run (maker's guidance) | No official figure; Fooocus, an SDXL app, needs 4GB minimum on NVIDIA | About 13GB | About 24GB |
Licence | CreativeML Open RAIL++-M | Apache 2.0 | FLUX Non-Commercial License |
Best for | Low-VRAM PCs, huge LoRA library | Modern quality and editing on 16GB cards | Higher quality on 24GB cards, non-commercial use |
What SDXL is
According to Stability AI's model card on Hugging Face, SDXL is a latent diffusion model that uses two fixed, pretrained text encoders, OpenCLIP-ViT/G and CLIP-ViT/L. It's designed as an "ensemble of experts": the base model generates noisy latents, and a separate refiner model can handle the final denoising steps. Stability AI notes the base model can also be used on its own. Hugging Face lists the base model at about 3B parameters.
SDXL's biggest strength is its ecosystem. Civitai's documentation, for example, says a LoRA trained with its SDXL recipe is usable with any SDXL checkpoint, so styles and characters made by the community carry across checkpoints. Many local AI apps are built around it; Fooocus, for instance, says it supports SDXL models from Civitai and lists a minimum of 4GB of NVIDIA graphics memory.
SDXL is released under the CreativeML Open RAIL++-M licence. Read the licence terms before using the model or its images commercially.
What FLUX.2 [klein] is
Black Forest Labs describes the FLUX.2 [klein] family as its "fastest image models to date". On its model card, FLUX.2 [klein] 4B is a 4 billion parameter rectified flow transformer that generates images from text and supports multi-reference editing. Black Forest Labs says it unifies generation and editing in a single compact architecture, and runs on consumer GPUs such as the RTX 3090 and RTX 4070 with as little as 13GB of VRAM. Its example code generates an image in four steps.
There are two sizes:
FLUX.2 [klein] 4B is released under Apache 2.0. Black Forest Labs' help centre says that means you can use it commercially with no fees or approvals.
FLUX.2 [klein] 9B offers higher quality according to Black Forest Labs, but needs around 24GB of VRAM and is released under the FLUX Non-Commercial License. Commercial use of the local weights requires a licence, or you can use it through Black Forest Labs' API.
Both are much smaller than FLUX.2 [dev], which Black Forest Labs describes as a 32 billion parameter model released under a non-commercial licence.
Hardware: what can your PC run?
8GB graphics cards and below: SDXL is the realistic choice. Fooocus's 4GB NVIDIA minimum shows how far down SDXL apps can scale.
12GB to 16GB cards: FLUX.2 [klein] 4B's roughly 13GB requirement puts it within reach of 16GB cards, and Black Forest Labs names the RTX 3090 and RTX 4070 as examples of suitable consumer GPUs. SDXL has plenty of headroom here too.
24GB cards: FLUX.2 [klein] 9B becomes an option, with Black Forest Labs quoting around 24GB of VRAM.
Memory-saving options in apps such as ComfyUI can stretch what fits. See how to run ComfyUI on a low-VRAM GPU, and for buying advice, our guide to the best PC for Stable Diffusion and FLUX.
Customising with LoRAs and fine-tuning
SDXL's head start shows most here. Years of community LoRAs and checkpoints mean you can find a style for almost anything, and apps such as ComfyUI make them simple to load; see how to install ComfyUI on Windows to get started.
FLUX.2 [klein] can be fine-tuned too. Black Forest Labs' help centre says the undistilled "Base" versions of klein 4B and 9B are the right starting point. It says klein 4B trains on an NVIDIA GPU with 12GB of VRAM, such as an RTX 3060 12GB, and 32GB of system RAM, while klein 9B needs around 22GB of VRAM and 64GB of system RAM. It adds that a LoRA trains in roughly one to three hours on a consumer GPU, using open-source tools such as ai-toolkit or Hugging Face Diffusers.
The library of ready-made LoRAs for a newer model is naturally smaller, so check what's available for the style you want.
Which should you choose?
Choose SDXL if you have a graphics card with less than 12GB of VRAM, want the widest choice of community LoRAs and checkpoints, or use an app built around SDXL.
Choose FLUX.2 [klein] 4B if you have around 16GB of VRAM, want generation and editing in one model, and want a permissive Apache 2.0 licence.
Choose FLUX.2 [klein] 9B if you have a 24GB card and want higher quality for personal, non-commercial use.
There's nothing stopping you from installing both. Our best local AI image generator guide covers the apps that run them.
Our verdict
SDXL remains the most accessible local image model, running in apps that need as little as 4GB of NVIDIA VRAM, with a vast library of community LoRAs, under the CreativeML Open RAIL++-M licence. FLUX.2 [klein] 4B is the modern alternative for 16GB-class cards: it combines text-to-image, editing and multi-reference in one model, needs about 13GB of VRAM and is Apache 2.0. The 9B version needs about 24GB and is non-commercial. Pick SDXL for low-VRAM PCs and LoRA choice, and klein 4B if your card can run it. Forge PC hasn't benchmarked or compared these models' output.