FLUX 3 Image Klein workflows fixed: upsample prompts, render at 2K

Black Forest Labs has updated its ComfyUI workflows and Hugging Face Space for FLUX 3 Image Klein, the 5.6B-parameter open-weights image model it released days earlier, after feedback from users (x.com). Users are told to update to ComfyUI v0.41.3 and switch to the new workflow, and the company says the best results come from running its prompt upsampling and generating at 2K resolution. The thread does not say which specific issues were fixed.

The upsampling advice is the core of the change. According to the ComfyUI model card, Klein is meant to receive full scene captions that spell out framing, subject, background and lighting; a short prompt sent as typed produces generic output, and even a one-line edit instruction such as asking for something to be made blue can reframe the whole shot. The card tells users not to run the model without an upsampler, and the ComfyUI workflows run one by default, with a switch to send your own caption instead (huggingface.co).

BFL's own upsampler is a 9B vision-language model fine-tuned from Qwen's Qwen3.5-9B. It takes the short request, up to 10 reference images and optional bounding boxes, and writes one detailed caption for the image model. BFL says it trained it first on pairs of short requests and expanded prompts, then with preference tuning in which its candidate prompts were rendered by Klein at the same seed and ranked on the resulting images (huggingface.co). For users who cannot run its upsampler, BFL also published the system prompt in its GitHub repository so any LLM can do the expansion (github.com). It instructs the model to output a single self-contained caption, keep every explicit detail in the user's request, and decide everything left unspecified, such as setting, lighting and palette.