Run FLUX 3 Image Klein Locally: VRAM, ComfyUI and Licence

Yes, FLUX 3 Image Klein runs locally: Black Forest Labs says its int8 build works on NVIDIA RTX cards with as little as 8 GB of VRAM (it names the RTX 3070), through ComfyUI or Diffusers. In ComfyUI 0.41.0 or later you need four files from BFL's ComfyUI repository: the 5.83 GB int8 diffusion model, the 4.86 GB int8 Qwen3-VL-4B text encoder, the 0.34 GB FLUX.2 VAE and a prompt upsampler (6.72 GB in its smallest build) that BFL tells you not to skip. The images are free to use commercially only if you and your affiliates have less than US$5 million in annual revenue.

Every figure below comes from a primary source read on November 2, 2026: the Hugging Face model cards and file listings, BFL's reference code and its October 15 announcement, the ComfyUI release notes and docs, the Diffusers docs and the licence text. Local AI Master has not run FLUX 3 Klein on its own hardware, so the speed figures are the ones BFL published (measured by NVIDIA), and any total worked out here by adding up file sizes is marked as calculated.

A note on the name: Black Forest Labs calls it FLUX 3 Image Klein, ComfyUI's docs write FLUX.3 [klein], and the files say flux3-image-klein. They are all the same model.

---

What is FLUX 3 Image Klein?

A 5.6B open-weights image model that generates and edits in one checkpoint, announced on October 15, 2026. It is the second open release in Black Forest Labs' FLUX 3 family, after the FLUX 3 Action robotics model. The 5.6B figure covers the image transformer only (Hugging Face counts 5.55 billion parameters in the BF16 file); the text encoder is a separate Qwen3-VL-4B-Instruct model.

What one checkpoint does, per the model card and announcement:

• Text-to-image, single-image editing and multi-reference generation with up to 10 reference images.
• Bounding boxes in the prompt. You place elements with boxes on a 0 to 1000 grid, then recolor, move, replace or remove them through the prompt. BFL is explicit that boxes guide placement and scale; they are not a hard mask.
• Native 4K generation, according to the announcement.
• Text rendering, with the model card's own caveat that rendered text "may be inaccurate or subject to distortion".

The release is two image checkpoints plus a helper model:

The Base guidance value is 3.5 on the model cards, in Comfy's templates and in Diffusers; the table in BFL's reference repository lists CFG 4.

Treat the upsampler as part of the model. BFL's card says "Do not use the model without a prompt upsampler": the image model takes full scene captions, and "a short prompt sent as typed gives generic output." ComfyUI's templates, the reference code and BFL's demo all run it by default. Any other LLM can stand in if you give it BFL's system prompts (system_prompts.json in the upsampler repository) and paste the caption it writes into the image model.

Where the files live:

On November 2, Hugging Face's download counter showed 1,003,813 for the ComfyUI repackage, against 155,600 for the Diffusers-format distilled repository.

---

How much VRAM does FLUX 3 Klein need?

8 GB is the official floor, with the int8 files in ComfyUI or Diffusers. Comfy recommends the BF16 files only above 24 GB, and the NVFP4 files need a Blackwell GPU such as an RTX 50-series card.

What the official sources say, word for word where it matters:

The 31 GB figures come from BFL's own inference code on a data-centre GPU, run without its --offload option, and that code has no int8 path at all. They are not the requirement for a consumer card in ComfyUI, where the 8 GB guidance applies.

File sizes by precision

Decimal gigabytes, from the Hugging Face file listings:

The BF16, int8, NVFP4 and w4a8 files are in the ComfyUI repository. The FP8 builds sit only in the Diffusers-format repositories (variants/fp8r), for the reference code and Diffusers; the ComfyUI repository has no FP8 file. BFL's announcement gives the generator weights as 11.2 GB in BF16 and 5.6 GB in FP8, and its model card adds that a quantized file's size "is not a measure of total GPU memory use."

What a ComfyUI install downloads (calculated)

These totals are the file sizes above added together, so they tell you about disk space and download time, not peak VRAM:

Add 5.81 GB if you also want the int8 Base checkpoint. The all-BF16 sum lines up with the "about 39 GB of weights" in the Diffusers docs, and the same set without the upsampler (20.31 GB) with its "about 20 GB". BFL's reference-code route downloads more, because the Diffusers-format repository carries the transformer twice (as model.safetensors and as transformer/ shards): BFL puts the distilled model plus upsampler at 52 GB.

Which build to use on your GPU

Mapped from the guidance above, nothing added:

One honest gap: BFL's speed chart below covers the RTX 5090, 5070 and 4070 but no 8 GB card, so there is no official speed figure for the RTX 3070 configuration it says will run. For FLUX.2 cards and the VRAM maths behind them, the FLUX.2 local setup guide has the per-card table.

---

How fast is it? Black Forest Labs' published timings

About 4 seconds for a 1024x1024 image on an RTX 5090 and under 9 seconds on an RTX 4070, with the distilled checkpoint and the int8 files in ComfyUI, according to BFL. The charts in the October 15 announcement were measured by NVIDIA, with the RTX results in ComfyUI; BFL describes them as generation time on the GPU itself, while its API latency figures are the ones that add queueing, network and prompt upsampling.

Distilled checkpoint, 16 steps, seconds per image:

Base checkpoint, 50 steps, at 1024x1024: 12.5 s text-to-image and 27.4 s for an edit on the RTX 5090, 34.0 s and 83.0 s on the RTX 5070, 38.0 s and 90.7 s on the RTX 4070, 43.8 s and 97.1 s on the DGX Spark. The reference README adds one more pair: about 3 seconds per 1 MP image for the distilled checkpoint against 12 seconds for Base on a single H200.

The practical reading: iterate on the distilled checkpoint at 1024x1024, and save 2048x2048 edits for the final pass, where every card in the chart takes several times longer.

---

How to install FLUX 3 Klein in ComfyUI

Update to 0.41.0 or later, accept the licence on Hugging Face, download four files into your model folders, and open the FLUX 3 Image Klein template. No custom nodes are needed.

Step 1: Update ComfyUI to 0.41.0 or later. Native support arrived in ComfyUI v0.41.0 on October 15, 2026 (pull request #17064 added the FLUX.3 model, its Qwen3-VL-4B text-encoder setup and a FLUX.3 Layout Prompt node), and Comfy's FLUX 3 Klein docs say the templates need 0.41.0 or later. ComfyUI Desktop and Comfy Cloud follow stable releases, so they can lag a nightly install. If the templates are missing from the Template Library, the install is too old.

Step 2: Accept the licence on Hugging Face. The ComfyUI files are in the gated repository black-forest-labs/flux-3-image-klein-comfy. Sign in, open it, accept the FLUX Kommunity License, then log the command-line tool in with a read token (hf auth login).

Step 3: Download the four files straight into the model folders. This is BFL's command from the model card; run it from the folder that contains ComfyUI/, or change --local-dir to your models folder:

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

Comfy's built-in templates load the smaller w4a8 upsampler instead, so if you use them, swap the last file for text_encoders/qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors. For the Base checkpoint, add diffusion_models/flux3-image-klein-base-int8-convrot.safetensors. On a card with more than 24 GB, swap in the -bf16 versions of the model and upsampler.

Step 4: Check where the files landed.

ComfyUI/
  models/
    diffusion_models/
      flux3-image-klein-int8-convrot.safetensors
    text_encoders/
      qwen3vl_4b_int8_convrot.safetensors
      qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors   (or the int8 file from BFL's command)
    vae/
      flux2-vae.safetensors

Step 5: Load a template. In the Template Library, search for "FLUX 3 Image Klein Distilled: Text to Image" (16 steps) or "FLUX 3 Image Klein: Text to Image" (Base, 50 steps); both have an Image Edit version. BFL's own workflows (text-to-image, image-to-image and multi-image, for each checkpoint in BF16 and int8) are in the workflows folder of the same Hugging Face repository. Do not confuse any of these with ComfyUI's "FLUX 3 Image" Partner Node, which calls BFL's paid API and runs the larger FLUX 3 Image model.

Step 6: Confirm the loaders. Load Diffusion Model points at the flux3-image-klein file, Load CLIP points at qwen3vl_4b_int8_convrot.safetensors with its type set to flux3, and Load VAE points at flux2-vae.safetensors. The upsampler also loads through Load CLIP with type flux3 and runs in a Text Generate node; its enable_prompt_upsampler switch is on by default.

Step 7: Keep each checkpoint's sampling defaults. The distilled checkpoint runs at 16 steps (4 to 16 is the trained range) with cfg 1, and Comfy's docs note that raising cfg does not help. Base runs at 50 steps with cfg 3.5 for text-to-image, and the Base image-edit template uses cfg 1.

Step 8: Queue the prompt. The result is saved to ComfyUI/output/. A short prompt is fine while the upsampler is on, because it rewrites the request into a full caption first.

Editing with reference images and boxes

• Connect each reference to Set Reference Multi Latent, in order, and refer to them in the prompt as <ref_image_0>, <ref_image_1> and so on, up to 10. Comfy's docs say to add each reference to both the positive and the negative prompt, the way BFL's reference pipeline does.
• Write the result, not the change. The model expects a complete caption of the finished image with the edit folded in; BFL's card warns that a one-line instruction such as "make it blue" reframes the shot. With the upsampler on, it writes that caption from your instruction and the references.
• Reference size is a cost dial. The templates resize each reference to roughly 1 megapixel. Keeping references at native size keeps more detail, at a higher cost.
• Boxes for layout. The Create Bounding Boxes node draws regions, and the FLUX.3 Layout Prompt node appends them to the prompt as the layout table the model reads: [top, left, bottom, right] on a 0 to 1000 grid, each element tagged with an id in the caption.

Coming from FLUX.2? Only the VAE carries over. BFL's flux2-vae.safetensors has the same SHA-256 hash as the ae.safetensors file in the FLUX.2 [dev] repository, so it is the FLUX.2 autoencoder, but the transformer, the text encoder and the CLIP loader type are all new, and FLUX.2 Klein's Qwen3 text encoders do not apply.

---

Can you run FLUX 3 Klein with Diffusers?

Yes, as a modular pipeline, from a Diffusers install built from GitHub. The model card's install line is pip install git+https://github.com/huggingface/diffusers.git. This is the distilled checkpoint with the component offloading that the Diffusers docs recommend for GPUs that cannot hold everything:

import torch
from diffusers import ComponentsManager, ModularPipeline

manager = ComponentsManager()
pipe = ModularPipeline.from_pretrained("black-forest-labs/flux-3-image-klein", components_manager=manager)
pipe.load_components(dtype=torch.bfloat16)
manager.enable_auto_cpu_offload(device="cuda")

out = pipe(
    prompt="a red bicycle leaning on a white wall",
    generator=torch.Generator().manual_seed(1),
    output=["images", "prompt"],
)
out["images"][0].save("bicycle.png")
print(out["prompt"])  # the full caption the image model received

Things to know before you rely on it:

• The upsampler. The current Diffusers docs say the pipeline runs the prompt upsampler (the 9B model from flux-3-image-klein-pup) on every request, and output=["images", "prompt"] returns the caption it wrote. The example on the model card instead loads the upsampler beside the pipeline with Transformers and passes its caption as prompt. Either way, the upsampler runs.
• Memory. Everything in BF16 is about 39 GB of weights per the Diffusers docs (the transformer, the 4B text encoder and the 9B upsampler, with the VAE in FP32), and about 20 GB without the upsampler.
• Quantized builds. The FP8 (fp8r) and NVFP4 single-file checkpoints load through Flux3Transformer2DModel.from_single_file with a BFLQuantizationConfig; the Diffusers BFL quantization guide lists the kernel requirements for each.
• Defaults. The distilled pipeline runs 16 steps and takes no guidance or negative prompt; the Base pipeline runs 50 steps at guidance 3.5 and generates 1024x1024 by default.

---

Can you use FLUX 3 Klein commercially? The licence in plain words

Yes, for the images, if you and your affiliates have less than US$5 million in gross annualized revenue and you filter or review what you publish. Everything else commercial needs a licence from Black Forest Labs. The weights are under the FLUX Kommunity License v1.0, and you accept it by downloading or using them.

The conditions attached to commercial use as a Qualifying User (Section 2.e): you must either run content filters that prevent unlawful, unsafe or infringing output, or review outputs for unlawful or infringing content before you distribute them, and you must label outputs as AI-generated where the law requires it. Black Forest Labs claims no ownership of the images (Section 2.d). If you share the weights or a fine-tune, Section 3 requires a copy of the licence and BFL's attribution notice to go with them.

The Qwen3-VL-4B text encoder, the prompt upsampler and BFL's reference code are each Apache 2.0, which does not change the terms on the image model. If you need commercial use above the revenue line without a BFL licence, the Apache 2.0 options in the FLUX line include FLUX.2 [klein] 4B and FLUX.1 [schnell]. The site's image model licence comparison sets the FLUX terms against SDXL, Qwen-Image and the rest. This is a plain-language summary, not legal advice; the licence text is the authority.

---

FLUX 3 Klein vs FLUX.2 Klein vs FLUX.1: what changed?

Fewer parameters than FLUX.2 Klein 9B, a vision-language text encoder, layout boxes, 16 distilled steps instead of 4, and a new licence with a revenue threshold. Only differences the vendors document are listed:

BFL also reports, from its internal evaluations, that FLUX 3 Klein scores 91 Elo higher than FLUX.2 [klein] 9B in text-to-image and 70 higher in image editing, with 38% fewer parameters. Those are the vendor's own comparisons, not independent tests.

Three practical consequences:

• Nothing but the VAE transfers. New transformer, new text encoder, new loader type. LoRAs made for FLUX.2 or FLUX.1 belong to different networks; for FLUX 3 Klein, BFL points LoRA training at the Base checkpoint, with Diffusers and Ostris's AI Toolkit supporting it from day one.
• Do not judge speed by parameter count. FLUX 3 Klein is smaller than Klein 9B but its distilled checkpoint runs 16 steps where FLUX.2 Klein runs 4. Compare published timings, not sizes.
• When FLUX.2 still makes sense: if you need Apache 2.0 at any revenue (Klein 4B), or you already depend on a FLUX.2 workflow. The FLUX.2 setup guide covers it, the FLUX.1 local guide covers the original family, and the local image model comparison ranks the alternatives.

---

Is FLUX 3 Klein uncensored?

No. Black Forest Labs built safety measures into the weights and puts filtering duties in the licence:

• In the weights. The announcement lists dataset filters during pre-training "to limit exposure to harmful content" and targeted fine-tuning during post-training "to suppress harmful behaviors". BFL says the checkpoints were red-teamed internally and evaluated by its external safety partner, Cinder, and the model card says child sexual abuse material and non-consensual intimate imagery were specifically evaluated and mitigated before release.
• In the licence. Section 4 bans generating unlawful content, including CSAM and non-consensual intimate images, and Section 4.d bans circumventing the model's usage restrictions or re-enabling functionality it disables. Commercial use as a Qualifying User requires content filters or review of outputs (Section 2.e).
• In the reference code. BFL's inference code ships optional checks that screen prompts, reference images and outputs for copyrighted characters, brands and public figures (using a vision-language model) and for NSFW content (using a classifier), refusing anything flagged. Turn them on with --safety in the CLI or Advanced, Safety checks in the demo. BFL's code also writes EXIF tags (Software: AI generated, Make: Black Forest Labs) into the images it saves.
• The upsampler adds no filter. Its model card says it "does not refuse requests"; screening is the job of the checks above.

For how filtering works across local image models in general, and the legal lines that apply whichever model you run, see the uncensored local image generation guide.

---

Troubleshooting FLUX 3 Klein in ComfyUI

The model is less than three weeks old, so this list is short and every item is tied to the docs or to a public ComfyUI issue.

"Could not detect model type" or "ERROR UNSUPPORTED DIFFUSION MODEL"

Two documented causes. First, a ComfyUI install older than 0.41.0 does not know the FLUX.3 architecture at all, so update. Second, the file is not a ComfyUI-format build: in issue #17232, a third-party FP8 conversion made for another app failed with exactly this error, and kijai, the ComfyUI developer who wrote the FLUX 3 support, replied that "those files simply aren't in ComfyUI compatible format, int8 is the preferred format to use." Load the BF16, int8 or NVFP4 files from black-forest-labs/flux-3-image-klein-comfy: those are the only builds the ComfyUI repository and Comfy's docs list, while the fp8r files in BFL's Diffusers-format repositories are documented for the reference code and Diffusers. More on these errors: ComfyUI quantization load errors.

CLIP Text Encode fails, or the text encoder will not load

Check the two documented requirements first: ComfyUI 0.41.0 or later, and Load CLIP set to type flux3 for both qwen3vl_4b and the upsampler file. An open report, issue #17286, shows CLIP Text Encode failing with AttributeError: 'NoneType' object has no attribute 'device' on FLUX 3 Klein workflows; no fix had been confirmed in that thread by November 2.

"CUDA error: an illegal memory access was encountered" while the upsampler runs

In the still-open issue #17115, the Text Generate node running the FLUX 3 upsampler (the int8 file) crashed a few hundred tokens in on an RTX 3090 Ti. The reporter's A/B runs showed the crash only with the node's mtp input on auto, its default; every run with mtp off finished. Commenters hit the same error on other RTX 30-series cards with a different Qwen3.5-9B-based model, and one of them said launching with --disable-dynamic-vram avoided it. A tester on an RTX 5090 could not reproduce it, which is why the thread suspects RTX 30-series (Ampere) cards. The report adds that after the crash the queue stops processing new prompts until ComfyUI restarts, even though the interface still loads. These are user reports, not an official fix; the ComfyUI out-of-memory and DynamicVRAM guide explains that flag.

Out of memory on an 8 GB or 12 GB card

• Use the int8 model and int8 text encoder, the configuration BFL says fits 8 GB, rather than BF16.
• Keep reference images near 1 megapixel, as the templates do; native-size references cost more.
• Take the upsampler out of the run: write captions with another LLM using BFL's system prompts and switch enable_prompt_upsampler off. BFL's reference code documents the same split (upsample first, then generate with --no-pup) for when "the upsampler and the model do not fit on the GPU together".
• If the traceback mentions HostBuffer or DynamicVRAM, follow the DynamicVRAM fix.

NVFP4 or FP8 files will not run on your card

NVFP4 needs a Blackwell GPU, FP8 in BFL's reference code needs Ada, Hopper or Blackwell, and RTX 30-series cards have no FP8 support, which is why BFL ships int8 builds. In ComfyUI, use int8 unless you are on Blackwell.

Images look generic or ignore what you typed

That is what BFL says happens when a short prompt reaches the model directly: "a short prompt sent as typed gives generic output." Switch enable_prompt_upsampler back on, or write a full scene caption (medium and style, scene, lighting and camera, each subject's colours, materials and position in the frame). For edits, describe the finished image rather than giving a one-line instruction.

Apple Silicon

Neither Black Forest Labs nor Comfy publishes guidance for Macs; every official memory and speed figure is for NVIDIA hardware. On October 22, issue #17309 reported that on an M4 Max every FLUX.3 workflow using reference images (edit and multi-reference) failed with RuntimeError: Invalid buffer size: 58.21 GiB on ComfyUI 0.41.3 and 0.42.4. The reporter closed it the next day without a linked fix, so test an edit workflow before you commit to Klein on a Mac.

"Access to model ... is restricted" when downloading

The full message reads "Access to model black-forest-labs/flux-3-image-klein-comfy is restricted. You must have access to it and be authenticated to access it." Accept the licence on the repository page while signed in, then run hf auth login so the command-line tool sends your token. The upsampler repository is not gated; the three image repositories are.

For anything not listed here, the ComfyUI error messages index explains what each common error means.

---

Key takeaways

1. FLUX 3 Image Klein is a 5.6B open-weights model from Black Forest Labs (October 15, 2026) that generates and edits with up to 10 references and layout boxes, using a Qwen3-VL-4B text encoder and the FLUX.2 VAE.
2. 8 GB of VRAM is the official minimum, with the int8 files in ComfyUI 0.41.0 or later. BF16 is for cards above 24 GB, and NVFP4 needs Blackwell.
3. Run the prompt upsampler. BFL says short prompts sent as typed give generic output; the templates run it by default.
4. The images are free for commercial use only under US$5M revenue (you plus affiliates), with filtering or review and AI disclosure where required. Fine-tunes and anything above the threshold need a BFL licence.
5. It is not an uncensored model. BFL filtered the training data and fine-tuned the weights against harmful output, and the licence bans unlawful content, including CSAM and non-consensual intimate images.

---

Sources

• Black Forest Labs, "FLUX 3 Image Klein", October 15, 2026: parameter count, builds, 8 GB guidance, timing charts, Elo comparison, licence summary, safety measures, API pricing.
• Hugging Face model cards and file listings for flux-3-image-klein, flux-3-image-klein-base, flux-3-image-klein-comfy and flux-3-image-klein-pup, read November 2, 2026: checkpoints, sampling defaults, memory figures, every file size, download counts.
• FLUX Kommunity License v1.0: the sections cited in the licence table.
• Reference inference code README: memory table, GPU requirements per precision, H200 timings, safety checks.
• ComfyUI v0.41.0 release notes and Comfy's FLUX 3 Klein tutorial: version requirement, loaders, templates, sampling settings.
• Diffusers FLUX 3 Image Klein docs: modular pipeline, memory, quantized loading.
• Comparison models: the FLUX.2 [klein] 9B and 4B model cards, BFL's FLUX.2 [klein] announcement (January 15, 2026), BFL's FLUX.1 launch post (August 1, 2024) and the FLUX.1 [dev] model card.