flux-3-image-klein-pup : nouveautés, matériel requis, comment l'utiliser gratuitement

FLUX 3 Image Klein Prompt Upsampler turns a short request into a full-length prompt for the FLUX 3 image model. It even accepts up to ten reference images and optional bounding boxes.

FLUX 3 Image Klein Prompt Upsampler is a 9.4 billion-parameter vision-language model. It takes a brief description, a set of reference pictures, and optionally a JSON list of bounding boxes, then returns a single caption that contains scene, lighting, camera angle, subject details and exact box references. The caption is ready to feed straight into the FLUX 3 Image Klein generator, whether you are doing text-to-image, image-to-image editing, or layout-guided generation.

According to Black-forest, the upsampler is a full fine-tune of Qwen3.5-9B with the vision tower frozen. Training happened in two stages: first on pairs of short requests and their expanded prompts, then on preference data where the model’s own prompts were ranked after rendering with FLUX 3 Image Klein. The result is a model that knows exactly which details the image generator needs, without extra system prompting.

The biggest strength is that the output prompt already contains medium, style, lighting, camera angle and per-subject attributes, so the image model can produce high-quality results without generic defaults. Bounding-box coordinates are kept verbatim, letting you control placement precisely. Limits include English-only prompts, and reference images are resized to at most one megapixel, which can cause loss of tiny details. The model does not refuse requests, so external safety checks are recommended.

If you already run FLUX 3 Image Klein locally, this upsampler is a natural companion. It fits on a RTX 4060 in 4-bit mode, using about 8 GB of memory, and runs at roughly 45 tokens per second on a RTX 3060 with 12 GB VRAM. Faster cards reach up to 190 tokens per second.

Que faut-il pour le faire tourner chez vous ? En 4 bits, il prend environ 8 gigaoctets de mémoire. La plus petite machine personnelle qui le contient entièrement est RTX 4060. Sur RTX 3060 12 GB avec 32 gigaoctets de RAM, il écrit environ 45 tokens par seconde. Sur RTX 4090 avec 64 gigaoctets de RAM, il écrit environ 116 tokens par seconde.

Il n’existe pas encore d’API gratuite pour cela. Les poids sont ouverts : vous pouvez les télécharger depuis Hugging Face sous licence Apache 2.0. Arena ne l’a pas encore classé : nous ajouterons le score dès que les votes arriveront.

Compare flux 3 image klein pup with every free model, and check what your own GPU can run, on FreeLLMArena.