BFL updates FLUX 3 Image Klein’s inference recipe

Black Forest Labs has updated the recommended inference setup for FLUX 3 Klein, the open-weights image model released earlier this month. The checkpoint remains unchanged. The release adds a revised ComfyUI graph, refreshed Hugging Face Space, dedicated prompt upsampler, and guidance to generate at 2K resolution to reproduce the quality shown in launch demos.

Developers who downloaded Klein at launch can keep their existing weights. Results from short prompts or low-resolution canvases may improve after updating the workflow because the new setup more closely matches the conditions used during training and internal testing.

BFL rewires the default workflow

• A new ComfyUI workflow aligned with ComfyUI v0.41.3
• An updated Hugging Face Space with the recommended defaults enabled
• A separate upsampler model, flux-3-image-klein-pup, based on a Qwen3.5-9B fine-tune
• The upsampler’s system prompt, published in the repository for use with other language models
• A recommendation to generate at 2K resolution

The pup suffix identifies a prompt-processing model. It expands terse instructions into structured descriptions of the subject, lighting, composition, camera, and visual details before Klein begins image generation.