Set against a Stable Diffusion AI image generator, FLUX.2 Klein comes in two sizes, 4B and 9B. Klein 4B is released under Apache 2.0, making it the commercially usable option, while Klein 9B carries a non-commercial licence. That makes 9B a research and evaluation model unless you obtain a separate commercial licence from Black Forest Labs. The case for Klein 4B is speed: it is a distilled model the vendor positions for interactive, production and latency-critical work. Whether that speed holds on your GPUs, at your resolution and with your prompts is something to measure before you commit.
Klein 4B, Klein 9B, SDXL and SD 3.5 Large at a glance
| FLUX.2 Klein 4B | FLUX.2 Klein 9B | SDXL | SD 3.5 Large | |
|---|---|---|---|---|
| Denoiser | 4B rectified flow transformer | 9B flow model with an 8B Qwen3 text embedder | Latent diffusion U-Net, two text encoders | Multimodal Diffusion Transformer (MMDiT) |
| Sampling | Distilled to few steps; undistilled base also published | Step-distilled to 4 inference steps | Base model plus optional refiner for the final denoising steps | n/a |
| VRAM figure or memory option | ~13GB per the model card; other sources range from ~8GB to 17.8GB (see below) | ~29GB | CPU offload available when VRAM is short | 4-bit NF4 quantisation via bitsandbytes |
| Licence | Apache 2.0 | FLUX.2-dev Non-Commercial License | Community Licence, free below $1M annual revenue | |
| Editing | Text-to-image and multi-reference editing in one model | Text-to-image and multi-reference editing in one model | Image-to-image via SDEdit-style denoising | n/a |
How fast FLUX.2 Klein generates in production
Black Forest Labs describes the Klein family as its fastest image models, with end-to-end inference “as low as under a second”. Those vendor speed figures were measured on a GB200 in bf16. ComfyUI’s guide quotes different numbers on an RTX 5090: about 1.2 seconds for the distilled 4B model and about 17 seconds for the undistilled base 4B. The vendor’s figure and ComfyUI’s figures come from different hardware and different settings, so they cannot be compared with each other, and neither one tells you how your own GPU will behave.
InferenceBench tested Klein 4B on a single NVIDIA H100: 0.19 seconds at 512x512 and 0.57 seconds for a 4-step generation at 1024x1024. The same benchmark reports multi-GPU scaling at 87% efficiency up to four GPUs and 74% on an eight-GPU node. That is one third-party run of one model on a single H100. We would use it to sanity-check the vendor’s claims, not to plan capacity.
In ComfyUI’s RTX 5090 numbers the base 4B takes roughly fourteen times as long as the distilled 4B (17 s against 1.2 s) on the same card. The base model is there for flexibility and fine-tuning, while the distilled 4-step model is built for deployments where speed comes first. Few-step distillation has published precedent: Sauer et al. (2023) showed that Adversarial Diffusion Distillation samples large image diffusion models in 1 to 4 steps and reaches SDXL-level performance in four. If a Klein deployment feels slow, the first thing to check is which checkpoint is loaded and how many steps the sampler is running.
How much VRAM FLUX.2 Klein 4B and 9B need
The published memory figures for Klein differ by source, precision and resolution:
| Source | Model | Figure | Conditions stated |
|---|---|---|---|
| Black Forest Labs model card and launch blog | 4B | ~13GB | RTX 3090/4070 and above |
| Black Forest Labs GitHub repo | 4B | ~8GB | RTX 3090/4070 and up |
| ComfyUI guide | 4B distilled / base | 8.4GB / 9.2GB | RTX 5090 |
| InferenceBench | 4B | 16GB at 512x512, 17.8GB at 1024x1024 | Single H100, bf16 Diffusers pipeline |
| Black Forest Labs model card | 9B | ~29GB | “Accessible on NVIDIA RTX 4090 and above” |
| WillItRunAI | 9B | Fits a 16GB GPU at FP8 | With text encoder offloading |
Treat no single row as settled. For 9B, the text encoder is big: the vendor card pairs the 9B flow model with an 8B Qwen3 text embedder, so whether that encoder sits on the GPU, gets offloaded to the CPU or is quantised will move peak memory a lot. WillItRunAI describes the 9B encoder as T5-XXL plus CLIP-L, which contradicts the vendor card. We trust the card on the architecture and treat WillItRunAI’s memory numbers as indicative only.
The 9B card gives ~29GB without stating a precision, which is more than the 24GB of the RTX 4090 it names as accessible.
Running a Stable Diffusion-style generator locally with FP8 or GGUF
Yes, Klein runs on consumer GPUs, and quantisation reduces VRAM for Klein and SD 3.5 alike. Black Forest Labs released FP8 and NVFP4 versions of every Klein variant, built with NVIDIA for RTX GPUs, and claims FP8 is up to 1.6x faster with up to 40% less VRAM. “Up to” marks a best case, not a typical result.
GGUF is a different tool. Unsloth publishes a GGUF build of Klein 4B that keeps important layers at higher precision. Diffusers loads prequantised GGUF checkpoints through from_single_file on model classes. During inference the weights stay in a low-memory dtype and are dequantised to the compute dtype in each module’s forward pass, so the memory saving comes with a dequantisation on every pass. The Diffusers Flux2 documentation says the same about quantisation in general: memory efficiency comes with a trade-off in inference latency. Optimised CUDA kernels claw back about 10% of GGUF inference time, according to the Diffusers documentation.
Here is how we read the trade-off. FP8 is the first option to try when you need both speed and memory on a supported card, and GGUF is mainly for fitting a model onto hardware it would not otherwise fit, accepting slower generation. Stable Diffusion has equivalent escape routes: SDXL supports CPU offloading through enable_model_cpu_offload, and SD 3.5 Large documents 4-bit NF4 quantisation with bitsandbytes.
Is Stable Diffusion free to use commercially, and what about Klein?
It depends on the version. Older Stable Diffusion models (SD 1.5, SDXL) are under the CreativeML Open RAIL-M licence, which a 2026 terms.law summary describes as allowing commercial use with no revenue limit, subject to use-based restrictions. SD 3.5 Large uses Stability AI’s Community Licence, which is free for research, non-commercial and commercial use by organisations or individuals with less than $1M in total annual revenue. Above that threshold you need an Enterprise Licence. Stability AI’s licence page adds that the threshold counts annual revenue “regardless of the source of that revenue”.
Klein splits cleanly. Klein 4B is Apache 2.0. Klein 9B uses the FLUX.2-dev Non-Commercial License, whose definition of non-commercial purpose excludes revenue-generating activity, anything that interacts directly with or affects end users, and training models for commercial use. The same definition does allow for-profit companies to test and evaluate in a non-production environment. One clause causes confusion: the licence claims no ownership of outputs and allows commercial use of them, but it separately prohibits using the model itself for any commercial or production purpose. So you can benchmark 9B internally, but you cannot serve it to customers without a commercial licence from Black Forest Labs.
That is an engineering reading of the licence texts. A commercial deployment still needs your own legal review.
Stable Diffusion image-to-image versus Klein’s multi-reference editing
Both families edit existing images, but they do it in different ways. The Stable Diffusion image-to-image pipeline in Diffusers takes a text prompt and an initial image and uses the diffusion-denoising mechanism from SDEdit, where the core tension is between staying faithful to the input and producing a realistic result.
Klein treats text-to-image and multi-reference editing as jobs for a single unified model. ComfyUI lists the supported edits as style transforms, semantic edits, object replacement and removal, multi-reference composition and iterative edits. In FLUX.1, inpainting and outpainting were handled by FLUX-Fill and in-context editing by the FLUX-Kontext models. For deployment, Diffusers provides a Klein KV pipeline that caches the reference image’s attention keys and values on the first denoising step and reuses them on later steps, which speeds up inference when reference images are involved. If your pipeline edits more than it generates, that cached path is the one to profile.
Expect to tune Klein’s image-to-image settings against your own inputs, not copy SDXL settings across.
FLUX.2 Klein against a Stable Diffusion AI image generator
A Stable Diffusion AI image generator is a latent diffusion model: rather than working on pixels, it runs the diffusion process in the compressed latent space of a pretrained autoencoder, which cuts compute while keeping visual fidelity. Cross-attention layers condition the denoiser on inputs such as a text prompt. Generation starts from noise in that latent space, which is denoised step by step and then decoded back to an image. Our explainer on how the diffusion process in AI image generators works covers the mechanism in more depth, and AI art generation with Stable Diffusion covers the creative side.
AI image generation architectures have since moved apart. SDXL kept the U-Net but made it three times larger, mostly through more attention blocks and a second text encoder, and it adds a refiner for the final denoising steps. SD3 and FLUX.1 moved to fully transformer-based denoisers, and FLUX models are rectified flow transformers trained in an autoencoder’s latent space. FLUX.2 is a new architecture pretrained from scratch, not an update of FLUX.1.
On speed, InferenceBench reports Klein 4B as 5 to 10 times faster than SDXL on equivalent hardware, which it attributes to fewer parameters, the transformer architecture and 4-step flow matching. The same benchmark gives Klein 4B an average CLIP score of 0.335 and names real weak spots: text rendering that is decent but not flawless, and complex multi-attribute prompts that sometimes assign properties to the wrong object. Black Forest Labs says Klein 9B matches or exceeds models five times its size, but that is the vendor’s claim.
Be wary of the “drop-in replacement” framing too. InferenceBench calls FLUX.2 a drop-in improvement over SDXL with the same Diffusers pipeline, but Diffusers documents dedicated Flux2 Klein and Klein KV pipelines.
When should a production pipeline choose a small, fast diffusion model over a larger one?
Pick the small, distilled model when latency is part of the product: live previews, interactive editing, or a generate-and-revise loop where a person is waiting. Cost per image is the other reason. InferenceBench reports about 2,500 Klein 4B images per dollar of GPU time on its single-H100 run. Look beyond Klein 4B when output needs legible typography or complex multi-attribute prompts, which InferenceBench names as its weak spots. Stability AI positions SD 3.5 Large on improved typography and complex prompt understanding, a vendor claim that is worth testing against Klein on your own hard cases.
For Klein 4B versus 9B, check the licence first. If the images reach customers, 4B is the open option, and 9B is the model to evaluate while you decide whether a commercial licence is worth negotiating.
Profiling the shortlist on your own GPUs
Choose model size and quantisation from a profiled workload, not from launch-day benchmarks. TechnoLynx profiles and benchmarks candidate generative image models (Klein 4B and 9B, SDXL, SD 3.5) on a team’s own workload and GPUs, so the model and quantisation choice rests on measured latency, memory and cost per image instead of vendor figures.
Run these steps in order, changing one variable at a time:
- Fix the workload. Collect a prompt set from real traffic, including your hardest cases (text in images, multi-object scenes), and record the resolutions, the split between generation and editing, and the batch sizes you actually serve.
- Pin variant and precision. Distilled or base checkpoint, bf16, FP8 or GGUF, and the step count. The ComfyUI 4B numbers show distilled against base alone moving latency by more than an order of magnitude.
- Measure peak VRAM at your largest resolution. Include the text encoder and the VAE.
- Measure end-to-end latency under load. Time text encoding, denoising and decoding together, warm, at the concurrency you expect, and record the median and the tail, not a single best run.
- Score quality on your prompts. Side-by-side review of the hard cases is what tells you whether Klein’s weak spots matter for your product.
- Compute cost per image at sustained throughput on the GPU you would actually rent or buy.
- Check the licence before anything leaves a non-production environment.
Our piece on monitoring ML models in production for image-generation stacks covers what to watch, and shipping deep learning stages to production covers the packaging side.
Who can benchmark Klein and SDXL on our own GPUs?
TechnoLynx runs that profiling sequence on your hardware: building a prompt set from your traffic, testing distilled and base variants at bf16, FP8 and GGUF, measuring peak VRAM and latency under realistic concurrency, and setting cost per image against the quality your reviewers accept.
This is not a fit for every team. A single creator running Klein on one workstation is well served by the ComfyUI guide’s ready-made workflows. Licence interpretation is a legal question, and our part is the engineering only. A team that just needs images, with no wish to own a model, may be better off with a hosted API than with any self-hosted model.
FAQ
What is the difference between FLUX.2 and FLUX.2 Klein?
FLUX.2 [dev] is a 32B flow matching transformer under a non-commercial licence. Klein is the compact family of 4B and 9B models built for sub-second generation and editing.
How is FLUX.2 different from FLUX.1?
FLUX.1’s variants share a 12B architecture. FLUX.2 is a new architecture pretrained from scratch, not a retrained FLUX.1.
Is FLUX.2 Klein 9B distilled?
The released 9B model is step-distilled to 4 inference steps. ComfyUI’s model file list also includes an undistilled 9B base checkpoint.