FLUX 3 Klein runs in ComfyUI from version 0.41.0, released on 2026-10-15 with official templates. The ComfyUI weight files are BF16 at 11.1 GB, int8 at 5.8 GB, and NVFP4 at 5.3 GB. The model card says “Do not use the model without a prompt upsampler.”

This page is the file-and-folder setup for that ComfyUI support. Every figure below is from the verified table for this site. No local timing and no interface screenshot were supplied with it, so none are shown.

Requirements

ComfyUI added native FLUX 3 Klein support in v0.41.0 on 2026-10-15. The same notes say the Klein template needs ComfyUI 0.41.0 or newer, and that 0.41.1 backported the template. Update before you open the official workflow. An older app will not have it.

The ComfyUI weights are in the Hugging Face repository black-forest-labs/flux-3-image-klein-comfy. Weight repositories for this model are gated with an “auto” gate: log in, accept the terms, and access is granted automatically. The prompt-upsampler repository is not gated.

The published name of the model is FLUX 3 Image Klein. ComfyUI also writes it as FLUX 3 Klein.

Which files to download

Use the ComfyUI package sizes when you are downloading files for ComfyUI. The Hugging Face ComfyUI repository lists:

Build Size Note from the verified table
BF16 11.1 GB ComfyUI file
int8 5.8 GB ComfyUI file. The same repo says the int8 files run on 8 GB GPUs such as the RTX 3070
NVFP4 5.3 GB ComfyUI file. NVFP4 needs a Blackwell GPU

The Black Forest Labs blog gives a different pair of generator sizes: BF16 at 11.2 GB and FP8 at 5.6 GB. Those blog figures are the generator weights, not the ComfyUI file listing. Keep them separate. If you are placing files in ComfyUI, the listing to follow is 11.1 GB, 5.8 GB, and 5.3 GB.

The text encoder on the model card is Qwen3-VL-4B-Instruct, licensed Apache-2.0. No file size for that encoder is in the verified table, so this page does not give one.

The official prompt upsampler is flux-3-image-klein-pup. It is a 9B vision-language model fine-tuned from Qwen3.5-9B, under Apache-2.0. The BF16 file is 18.8 GB. ComfyUI’s file set also includes int8 and w4a8 upsampler builds. Byte sizes from the Hugging Face API put the w4a8 upsampler at 6.72 GB. An int8 upsampler file size was not part of that reading.

The workflow also uses a VAE. The ComfyUI notes for this setup name that file flux2-vae. No size for it is in the verified table.

Skip NVFP4 unless the GPU is Blackwell. The 5.3 GB file is the wrong download for an RTX 30-series or 40-series card.

Folder layout

Put the diffusion weights in models/diffusion_models. Put the text encoder in models/text_encoders. Put the VAE in models/vae. Those are the three folders named for this workflow. A file in the wrong folder will not show up in the matching loader.

Download command

Accept the gate on the ComfyUI repository first. Then the Hugging Face CLI can download that repository by its id:

hf download black-forest-labs/flux-3-image-klein-comfy

That line is only the repository id from the verified table. It does not add a local directory, an include filter, or any other flag, because those flags were not in the table. Log in with the hf CLI before you run it. A gated file will not download on an anonymous request.

The upsampler can be fetched from black-forest-labs/flux-3-image-klein-pup. That repository is the one the verified table marks as not gated.

Built-in templates

The 0.41.0 support includes official templates for text to image and for image edit.

Klein, the distilled checkpoint, is the one behind the Black Forest Labs API. The official template default is 16 steps. The model card describes that checkpoint as 16 steps with no guidance. The ComfyUI tutorial describes the distilled template as 16 steps at CFG 1, adjustable from 4 to 16.

Klein Base is the undistilled checkpoint: 50 steps, for fine-tuning and LoRA. The official template default for Klein Base is CFG 3.5, which matches the model card and the Black Forest Labs docs. The GitHub README lists 4.

Pick the distilled template for ordinary text-to-image and edit workflows. Pick Klein Base when the work is a fine-tune or a LoRA, at the default CFG of 3.5.

Prompt upsampler switch

The model card says: “Do not use the model without a prompt upsampler.”

The official upsampler is the 9B model named above. Any LLM paired with the official system prompts is the documented alternative. In the ComfyUI templates this control is the enable_prompt_upsampler switch. Leave it on unless the prompt was already expanded by an external LLM and you are intentionally skipping the built-in upsampler.

Memory for the upsampler is a separate published figure. On an 80 GB card the official BF16 upsampler adds about 28 GB. Our own run used the w4a8 upsampler (6.7 GB file) and added about 5.8 GB of peak VRAM, not 28 GB. The B200 peaks of 31 GB in BF16, 26 GB in FP8, and 25 GB in NVFP4 were measured at 1024×1024 with the upsampler turned off. Those peaks are the official hardware numbers, not our ComfyUI run.

Multi-reference and bounding boxes

The model card, the GitHub README, and the Black Forest Labs blog agree on the capability list that matters inside these templates. The model does text-to-image, single-image editing, and multi-reference editing with up to 10 images. It also takes bounding-box layout in the prompt, renders text inside the image, and goes up to 4K.

The text-to-image template is the path for a prompt with no input image. The image-edit template is the path for one image, or for reference images up to that limit of 10. This page does not add a coordinate format for the boxes. The verified table says the prompt can carry bounding boxes. It does not give the numeric grid, so no grid is printed here.

8 GB GPU settings

On an 8 GB NVIDIA card, the ComfyUI file to use is the int8 build, 5.8 GB. The ComfyUI repository says: “The int8 files run on 8 GB GPUs such as the RTX 3070.” The Black Forest Labs blog says the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM, through ComfyUI or Diffusers.

That is the official 8 GB claim. It is not a test from this site.

By file size, using byte sizes from the Hugging Face API, the int8 generator is 5.83 GB, the int8 text encoder is 4.86 GB, the w4a8 upsampler is 6.72 GB, and the VAE is 0.34 GB. Together that is about 17.8 GB, so the full stack with the upsampler cannot be resident on an 8 GB card. The official 8 GB figure must refer to running without the upsampler resident, or to offloading or sequential loading. Whether an 8 GB card actually runs it is not something we tested.

BF16 at 11.1 GB is larger than 8 GB before activations and the upsampler, so it is the wrong first file for that card. NVFP4 is the wrong file unless the GPU is Blackwell. The int8 ComfyUI file is the one the repository names for an 8 GB GPU such as the RTX 3070.

Speed, as published

NVIDIA’s measurement, for the distilled model in INT8, in ComfyUI, at 1 megapixel, is about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070. The Black Forest Labs blog and the ComfyUI blog are the sources. Label it as NVIDIA’s measurement. It is not our run.

Screenshots

No screenshots are on this page. None were provided. This guide does not fill the gap with stock images, with another site’s ComfyUI window, or with a reconstructed interface. When a real capture from the author’s own ComfyUI is available, it can be added with the app version in the caption.