>>106254879
qwen-image is like dealing with a talented person who is stubborn and has autism. It's "stubborn" because unlike SDXL models, it doesn't change much based on seed; you only get minor changes if anything by changing the seed. It has "autism" because it needs detailed descriptions in the prompt about what goes where, otherwise it takes a "most likely" approach, i.e. you asked for shoes on the floor, you didn't say 1girl wasn't wearing any, so she gets shoes too - like that.
That said it does a good job with anatomy and stays coherent with multiple characters in a scene. I think it's actually a good starting point for i2v work, because it tends to gen characters consistently, even if you change the scene they are in.
You need at least 28GB of VRAM to run it though.
