SAP × FLUX.2-klein-4B

Stage-Aware Prompting decomposes your prompt into proxy stages aligned with the diffusion model's coarse-to-fine denoising. This lets you generate images from contextually contradictory prompts — a polar bear in a desert, a garden hose spraying fire, blue Shrek — that standard generation struggles with.

A local LLM (Qwen2.5-3B) analyses your prompt and produces the stage-aware decomposition. FLUX.2-klein-4B then generates the image in seconds on a ZeroGPU instance. No API keys needed.