qwen lora tests
Description
Comments (6)
Does Qwen actually handle text this well? That's crazy.
their showcase says they have strength on it. But in my test, there is limited number of texts until it shows otherwise. running with full bf16 model resulting a better texts, than using fp8 model. i am not sure which version my lora was trained on (fp8 or bf16).
one thing to note is the photo size is 1000x1600. That helps. Whenever I use text It will sometimes add extra text. If I needed it, I'd probably make sure to have a good I2I workflow to either iterate or straight up correct wrong
@makiaeveli do you mean 1000x1600 get better results for texts?
@mrdion I'm just saying it's fairly large. Some "normal" photo sizes would be around 1200x1000 or even smaller. If you use that smaller default smaller image size it won't be big enough to get the font right
@mrdion thanks for the info save me some time I was using GGUF trying to mess with the generation settings to fix the text but nothing really working I guess I was wasting my time need to get a better GPU and/or wait for the next version of qwen with even better text neat lora though first one I've seen with an actual good business use case outside of art lol.