How to Run FLUX 3 Klein in ComfyUI on 8 GB VRAM (The Upsampler Alone Is 9.8 GB)

Black Forest Labs says FLUX 3 Image Klein runs locally in ComfyUI on an 8 GB card like the RTX 3070. That's true for the transformer. The prompt upsampler the model card calls "required" is a 9.8 GB file on its own, so the full int8 stack you download is 20.8 GB of weights for an 8 GB GPU.

You can still make it work, as long as you know where those 20.8 GB end up and which file you can swap out.

What FLUX 3 Klein is

FLUX 3 Image Klein came out on October 15, 2026. It's a 5.6B-parameter open-weights model for text-to-image, editing, and multi-reference generation with up to 10 reference images. It's the second open-weights release in the FLUX 3 family, after FLUX 3 Action.

BFL's headline comparison: against FLUX.2 [klein] 9B it gains 91 Elo in text-to-image and 70 Elo in editing, with 38% fewer parameters. Those are BFL's own internal evals, so treat them as a claim until someone independent reproduces them.

There are two checkpoints, and they want very different settings:

If you load Base with distilled settings, or the other way around, the results will be bad and you'll blame the model. People in the HF threads already did exactly that.

The 20.8 GB math for 8 GB VRAM

Here's what the ComfyUI repo's model card tells you to download for the int8 path:

The card's exact line is "The int8 files run on 8 GB GPUs such as the RTX 3070, and on GPUs without FP8 support (RTX 30 series and older)." Nowhere does it say they all fit at once, and they obviously can't. The transformer plus the text encoder is already 10.7 GB. The upsampler is a Qwen3.5-9B LLM that's bigger than the image model it feeds.

So "runs on 8 GB" means ComfyUI loads the stages one after another and moves whatever isn't in use out to system RAM. That works, but the 20.8 GB still have to live somewhere, and if it isn't VRAM it's RAM. On a 16 GB desktop, expect swapping and slow first runs. I'd consider 32 GB of system RAM the realistic floor for this setup. That's my estimate, not something BFL publishes.

Why you can't just skip the upsampler

The card is blunt about it: "Always run a prompt upsampler." The model was trained on full scene captions that cover framing, subject, background and lighting, and "a short prompt sent as typed gives generic output." The upsampler takes your one-liner and expands it into one of those captions. For edits, it rewrites the instruction as a complete caption of the reference image with the change included.

BFL staff said the same thing in the launch-day "wrong weights?" thread. One user ran the shipped t2i template at 16 steps and CFG 1 and called it worse than FLUX.2 klein 4B. The replies said the Comfy workflow had an error, that prompt upsampling "should be used all the time", and a day later that the upsampler "has a profound impact on model quality."

That makes the 8 GB marketing feel a bit generous to me. The model that fits on a 3070 is 5.8 GB. The setup that produces the results BFL shows needs another 9.8 GB of LLM.

How to run FLUX 3 Klein locally in ComfyUI

Update ComfyUI before anything else. In thread #7 a BFL staffer pointed people to ComfyUI v0.41.2 for "much better and updated workflows" after the broken launch ones. If you're on an older build, you probably have the bad templates.

Then pull the int8 stack straight into your models folder. This is the card's command:

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 

Load one of BFL's workflows from the card. There are editing, text-to-image and layout variants for both the base and distilled checkpoints. They run the upsampler by default. Only turn off "enable upsampler" if the prompt you're pasting is already a full caption.

Check the sampler settings before you queue anything. Distilled gets 16 steps and no guidance. Base gets 50 steps at CFG 3.5. For editing you can pass up to 10 reference images.

Which files to pick depends on your card:

• On RTX 30 series and older, stick with int8. These cards have no FP8 support, and the model card points them to the int8 "convrot" files.
• On Blackwell (RTX 50), the NVFP4 transformer is 5.3 GB.
• If you're short on RAM, swap the int8 upsampler for qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors. It's 6.7 GB instead of 9.8 GB, which brings the stack down to 17.7 GB. The card offers this file but doesn't say how much quality you lose, so compare the two on your own prompts.

The cheaper trick: bring your own upsampler

The card gives you two options: "Use ours", or "any LLM you prefer with a system prompt that expands the request into one full scene caption." BFL publishes the system prompts it uses in system_prompts.json in the flux-3-image-klein-pup repo. There are four of them (t2i, i2i, t2i_bbox, i2i_bbox), and that repo's card says outright that "any other LLM can do the job" with them. If you already have a local LLM running, or an API key, you can do the caption expansion there and paste the result into ComfyUI with the upsampler switched off. Then your GPU only has to hold the image stages.

One user in discussion #41 reports that a different enhancer, MOSS Image's in "Caption" mode on a strong LLM, produces prompts that work better on Klein than MOSS's own T2I mode. That's a single report, but it's cheap to try.

How fast is it?

BFL's numbers for distilled int8 at 1 MP in ComfyUI: about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070. At 16 steps and 1024×1024, they list 8.9 s on the 4070. Base at 50 steps ranges from 1.7 s on a GB300 to 43.8 s on a DGX Spark.

BFL doesn't publish a 3070 number, which is odd, since that's the card the model card names. BFL says those are GPU generation times only, so they leave out the time a 9B LLM spends writing your caption, and they obviously don't include ComfyUI shuffling 20.8 GB between RAM and VRAM. On an 8 GB card, expect your end-to-end time to be well above the brochure figure.

The quality fight in the comments

The HF discussions are rough. You'll find "CANNOT create faces at all" (#37), along with threads about warped faces, body malformations, and a "Thank God for Qwen" post. Some of that is the broken launch workflow and missing upsampler. Some of it might be real. Thread #22 says the model "works way better with higher resolutions", and one user in #7 got "dramatically better" results with more steps and CFG 3.

My read: judge it only after you've updated to the fixed workflows with the upsampler on. A lot of the early complaints came from people running it without either.

What the license actually allows

The weights use the FLUX Kommunity License v1.0. Non-commercial use is free. Commercial use of outputs is allowed for "Qualifying Users" with under US$5M in gross annualized revenue, affiliates included. Above that, you need a commercial license or the API, where the distilled model starts at $0.009 per 1K image.

The card also lists out-of-scope uses: military, surveillance, biometric processing, and using the model or its outputs "to improve another AI model that performs a function similar to a FLUX" model. That last one matters if you were planning to generate synthetic training data. The Qwen3-VL-4B text encoder and the upsampler are Apache-2.0, so the restrictions come from the FLUX weights alone.

For a freelancer or a small studio, the license is fine. If your company's revenue is anywhere near $5M, talk to legal before you build a product on it.

Sources

• FLUX 3 Image Klein ComfyUI model card
• BFL launch post
• HF discussion #7: "Did you upload the wrong weights?"
• HF discussion #37: faces
• HF discussions list
• BFL licensing

What GPU are you generating images on, and what's your real time per image with the upsampler on, measured from clicking Queue to seeing the image? And if you're on 8 GB, did you keep BFL's upsampler or bring your own LLM?