It's been a while. Recently, I have switched my image generation model to QWEN. It is amazing how well it adheres to the prompts! I can now draw scenes and concepts that weren't possible in Flux1-dev.

This is one of the first set of LoRAs I trained for QWEN. It is surprisingly easy to train. While it learns slower than Flux1, it seems to converge better and more predictable.

This is a fp16 LoRA so make you sure set Diffusion in Low Bits to fp16. Make sure you set CFG Scale to be greater than 1, and set negative prompts! I use 3.

Also, QWEN works better with Chinese prompts. It works for me since that is my native language. If you are having trouble getting QWEN to draw what you ask it to, try translating your prompt to Chinese. Though most of the time, English prompts work just fine.