FLUX Klein’s prompt upsampler is a model in its own right

Black Forest Labs’ FLUX 3 Image Klein release included a substantial extra: a prompt-upscaling model that expands short requests into detailed image instructions. OrcaRouter says the 18.8GB model is available under the Apache 2.0 licence, while Klein’s sibling model repositories carry a different licence. That makes the overlooked tool more than a footnote for people building local image-generation workflows.

Black Forest Labs Watch analysis

What happened

OrcaRouter says the upsampler is a fine-tune of Qwen3.5-9B, and that it is integrated by default into BFL’s reference inference code. It can be switched off with a –no-pup flag. The report says it handles text-to-image prompts, image edits and requests involving reference images or bounding boxes, with different instruction templates for each.

The upsampler’s weights are publicly available without an account under Apache 2.0, according to OrcaRouter. The report says that differs from the FLUX Klein model repositories, which use the FLUX Kommunity License. It also notes that the upsampler requires its own memory: BFL’s claim that an INT8 Klein image model can run with 8GB of VRAM does not establish that both models can run together in that footprint.

OrcaRouter’s technical analysis describes how the tool expands prompts and how BFL’s code connects it to image generation.

Why it matters

Prompt upsampling can shape what an image model produces before image generation even begins. Here, the tool reportedly turns brief requests into captions typically 100 to 350 words long, and its instructions specify details such as people’s appearance, backgrounds and rendered text. That gives users more to inspect than a mysterious improvement hidden behind a toggle: the report says the system prompts are included and readable.

There is a practical trade-off. An extra model may help turn an underspecified request into a more detailed image plan, but it also takes storage and GPU memory. The permissive licence may make the upsampler useful to developers even when the licensing terms of other parts of the image-generation stack are less permissive.

Our read

This is a worthwhile detail from the Klein release, not a separate image-model launch. The useful distinction is between the image model’s hardware claim and the resources needed when its prompt upsampler is also running. If you are experimenting locally, check the model and licence for each component, and try the –no-pup option if you want to compare the generated result with your own prompt untouched. More parts of an AI pipeline deserve daylight; this one has helpfully left its instruction manual lying about.

What to watch

• Whether BFL clarifies the upsampler’s intended licensing and role in its inference pipeline.
• Whether users publish reliable VRAM figures for running Klein and the upsampler together.
• How outputs compare with upsampling enabled and disabled across a range of prompts.

Discussion spark: Would you keep an upsampler that rewrites your prompt on by default, or would you rather the image model use exactly what you typed?

not affiliated with or endorsed by Black Forest Labs