Local setup guide
How to Run FLUX 3 Klein in ComfyUI
FLUX 3 Klein ships open weights and a ComfyUI template, so you can run it on your own GPU without an API key or per-image credits. Here is what to install, which files go where, how the four shipped workflows differ, and what the licence allows.
Running FLUX 3 Klein locally, or in a browser
Local ComfyUI keeps the whole loop on your machine: prompts, reference photos and output files never leave your disk, and nothing is billed per image. It is also the setup-heavy path, because the templates expect a particular ComfyUI version, five files and the right checkpoint per job.
If all you want is to see what the model does with your wording, skip the install. This site runs FLUX 3 Klein on its own servers and accepts up to three reference images, so a prompt here comes back in a minute.
This box opens the generator with FLUX 3 Klein, the model this site runs. Nothing is installed and no file is downloaded.
What you need before you start
The templates were built for one combination of version, files and checkpoints. Get one of these wrong and the workflow will not load, or will fail while sampling.
ComfyUI 0.41.0 or newer
The FLUX 3 Klein templates need ComfyUI 0.41.0 or later and shipped in the 0.41.1 backport. If the template is missing from the list, update ComfyUI first.
Access to the gated weights repository
The weights live in black-forest-labs/flux-3-image-klein-comfy on Hugging Face. Open it and accept the licence while signed in, or the files stay locked.
Five files in three folders
Two diffusion checkpoints, one text encoder, the prompt upsampler and the FLUX.2 VAE, filed under models/diffusion_models/, models/text_encoders/ and models/vae/.
GPU memory for the checkpoint you pick
ComfyUI does not state a requirement. Black Forest Labs quotes the generator weights at 11.2 GB in BF16 and 5.6 GB in FP8, and says the INT8 build runs on RTX GPUs with as little as 8 GB of VRAM; its reference code peaks higher at 1024x1024. Full figures in the FAQ.
Model files and where they go
The templates load the int8 (convrot) single-file checkpoints; bf16 and nvfp4 builds sit in the same repository. Place the downloaded files in these folders.
| Folder | Used by | |
|---|---|---|
| flux3_image_klein_base_int8_convrot.safetensors | models/diffusion_models/ | Full FLUX 3 Klein checkpoint; the text-to-image and image edit templates load it |
| flux3_image_klein_sd_int8_convrot.safetensors | models/diffusion_models/ | Step-distilled checkpoint for 4 to 16 step runs |
| qwen3vl_4b_int8_convrot.safetensors | models/text_encoders/ | Text encoder for both checkpoints; load it with CLIPLoader and type flux3 |
| qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors | models/text_encoders/ | Prompt upsampler used by the text-to-image templates |
| flux2-vae.safetensors | models/vae/ | VAE for FLUX 3 Klein — the FLUX.2 VAE |
Paths are relative to your ComfyUI folder. The repository is gated: accept the licence at huggingface.co/black-forest-labs/flux-3-image-klein-comfy first. Source: docs.comfy.org/tutorials/flux/flux-3-klein, read on October 17, 2026.
The four workflows, step by step
The template library ships four FLUX 3 Klein workflows: text-to-image and image edit, each in a full and a step-distilled version.
Text to image, full checkpoint
Open "FLUX 3 Image Klein: Text to Image", type the prompt in the subgraph, set width and height or pick a frame from ResolutionSelector, and leave steps at 50 with cfg 3.5. The Qwen3.5 9B upsampler runs by default; switch enable_prompt_upsampler off to sample your own words.
Text to image, distilled
This one loads flux3_image_klein_sd_int8_convrot.safetensors and runs in 16 steps with cfg fixed at 1, so raising cfg does not help. You can drop as low as 4 steps for a quick pass.
Image edit, full checkpoint
Load the source photo in the LoadImage node and connect it to Set Reference Multi Latent. Write the instruction as a full description of the finished image; the shipped example is "In <ref_image_0>, change the dress color <obj_1> to red". The canvas comes from Empty Flux 2 Latent, 1024x1024 in the template, and guidance is 1 here.
Image edit, distilled
One reference, one instruction, 16 steps and no guidance. Keep steps between 4 and 16, leave cfg at 1, and compare the input with the result in ImageCompare.
Prompt and reference tips that change the output
Five details from the ComfyUI documentation. They decide more of the result than any sampler setting.
Write the result, not the change
The editing templates expect a full caption of the finished image with the edit folded in: the shipped example describes the new colour and the object it applies to, rather than saying "change this".
ComfyUI numbers references from zero
Connect extra images to Set Reference Multi Latent in order and address them as <ref_image_0>, <ref_image_1> and so on. Note the difference: ComfyUI starts at <ref_image_0>, while the generator on this site starts at <ref_image_1> and takes up to three images.
Reference size is a cost dial
The templates resize each reference to roughly 1 megapixel, and the output latent takes that size. A reference kept at native size carries more detail into the edit and costs more.
Step count follows the checkpoint
50 steps and guidance 3.5 for text-to-image on the full checkpoint, guidance 1 for its editing workflow, and 16 steps with guidance fixed at 1 on the distilled one. Do not mix the two settings.
There is a Create Bounding Boxes node
The templates ship it for planning regional edits: it draws boxes over the source image, and Preview Any lists them.
What the licence says
Only what the publisher states, quoted as published. This is not a reading of the licence: if you intend to sell what you generate, read the full text.
| Licence as published | |
|---|---|
| FLUX 3 Klein weights | FLUX Kommunity License: non-commercial use, plus commercial output generation by Qualifying Users with less than US$5M in gross annualized revenue, subject to its terms |
| Text encoder (Qwen3-VL-4B-Instruct) | Apache-2.0 |
| Repository access | Gated on Hugging Face; you accept the licence before the files unlock |
| Full licence text | LICENSE.md inside the weights repository (linked below) |
Sources: bfl.ai/blog/flux-3-image-klein, huggingface.co/black-forest-labs/flux-3-image-klein-base and the LICENSE.md in the gated weights repository, read on October 17, 2026. Nothing here is legal advice.
Where to download the weights
This site runs FLUX 3 Klein on its own servers and does not host or sell the weights; nothing here is downloadable. The files live on Hugging Face and the steps above come from the ComfyUI documentation.
Running FLUX 3 Klein locally
Which ComfyUI version do I need for FLUX 3 Klein?
0.41.0 or later. The templates shipped in the 0.41.1 backport, so if "FLUX 3 Image Klein" is not in your template list, update ComfyUI and look again.
Can I download the FLUX 3 Klein weights from this site?
No. This site runs the model in your browser and does not host the files. Get them from the gated repository black-forest-labs/flux-3-image-klein-comfy on Hugging Face, after accepting the licence.
How much VRAM does FLUX 3 Klein need?
The ComfyUI documentation states none. Black Forest Labs quotes the generator weights at 11.2 GB in BF16 and 5.6 GB in FP8, and says the INT8 build runs on RTX GPUs with as little as 8 GB of VRAM. Its reference code peaks at 31 GB BF16, 29 GB FP8 or 26 GB NVFP4 at 1024x1024, plus about 28 GB for the upsampler on an 80 GB GPU. Vendor figures only.
Why does the workflow rewrite my prompt before sampling?
Because the text-to-image templates run the Qwen3.5 9B prompt upsampler by default, which expands a short request into a full scene caption. Set enable_prompt_upsampler to off to sample your own words.
Do the ComfyUI templates and this generator use the same reference numbering?
No. ComfyUI starts at <ref_image_0>; this site's generator starts at <ref_image_1>. Rewrite the tokens if you move a prompt between them.
Try the prompt before you install
Write the prompt above, run it once, then carry the wording you like into ComfyUI.