Make your first image locally with FLUX.3 Klein

FLUX.3 Image Klein is an image model that takes a prompt² or reference images and creates a new image. Load it in ComfyUI and press Queue to save a PNG. This path uses the distilled INT8 text-to-image template in ComfyUI on an NVIDIA GPU³ to make one 1024×1024 image.

ComfyUI 0.41.0 or later supports the template. Update ComfyUI, then choose Flux.3 Image Klein Distilled: Text to Image from Workflow Templates. This first run uses ComfyUI’s standard loaders and does not require custom nodes. ComfyUI’s FLUX.3 Klein guide

The weights repository, black-forest-labs/flux-3-image-klein-comfy, has gated file access. Review and accept its terms on Hugging Face, then download the four files below from Files and versions. Put the weights, text encoder, and VAE in the folders ComfyUI scans. ComfyUI model files

The upsampler is a separate Qwen3.5 9B model that expands short requests into scene descriptions. The template can enable it by default, which uses additional memory. If a model does not appear in the selector, check that it is in the models folder used by this ComfyUI instance, then restart the app. FLUX.3 Klein model card

Keep the square 1024×1024 size, batch 1, and 15 steps, then enter a prompt. Batch 1 creates one image at a time. A short prompt is fine when the upsampler is enabled. If you turn it off to save memory, include subject, setting, composition, and lighting in a full caption, such as “Still-life photo of a small teacup on a window-side table on a rainy evening, warm indoor light, soft raindrops blurred in the background.” ComfyUI’s distilled template defaults to 15 steps and supports 4–16.

Press Queue to start. When processing finishes, SaveImage writes the PNG to ComfyUI/output/. Open the file from that folder to confirm it was saved; a preview alone does not confirm the output file. The first run can include model loading. Record loading and generation separately when comparing time. ComfyUI text-to-image workflow

If a node turns red or cannot find a model, verify the file is in the correct ComfyUI folder and restart the app. The distilled file belongs in diffusion_models, with the text encoder and VAE in their specified folders. A similarly named base variant does not match the distilled template.

If a VRAM⁴ error occurs while loading the upsampler, turn off enable_prompt_upsampler to skip the separate Qwen3.5 9B model. Write the subject, setting, composition, and lighting directly in the prompt. If the error occurs during generation, keep batch 1 and 1024×1024, close other GPU jobs, and verify that you selected the INT8 distilled files. Changing several settings at once makes the cause harder to identify.

If the composition is wrong, add the subject’s position or camera viewpoint to the prompt. The model may render small text incorrectly, so add poster copy or product labels in an image editor. To edit an existing photo, open ComfyUI’s Image Edit template, load the photo with LoadImage, and describe the intended result. ComfyUI image-edit template

BFL and NVIDIA report RTX GPU generation times at 1024×1024, distilled 16 steps, and one image per run: about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070. ComfyUI’s template defaults to 15 steps, so it is a different setting. BFL reports API⁵ response time separately, including queueing and network time. Record your own generation time with the same prompt and settings before comparing. BFL measurement conditions

The weights use the FLUX Kommunity License. Qualifying users with under US$5 million in annual revenue, including affiliates, may use outputs commercially subject to the terms; other commercial use and commercial fine-tuning require a separate license. Read the restrictions in the license before publishing or selling.