According to VitalikButerin, a single round of FLUX development inference takes approximately 5 minutes due to the limitations of his 4070 GPU, which only has 8 GB of VRAM. This necessitates the use of enable_sequential_cpu_offload(), resulting in slower performance. He mentioned that a 4-bit quantization might fit but has not tested it yet. Additionally, manual edits take around 10 minutes, and a second round of inpainting takes about 1 minute.