>>106369979
one thing I'm going to test is translating the style prompt to Chinese. maybe the image metadata wasn't in English.

>>106370041
chromacousins...

>>106370085
TREAD+EQ-VAE should make a Qwen finetune (or any model finetune) orders of magnitude cheaper than even existing SDXL finetunes were. we'd just need a baker who is smart enough to implement both for Qwen before attempting.

>>106370093
>they did post-training ("aesthetics fine-tune") almost entirely on gpt4o synesthetic slop
any proof of this? I don't see it in here, they actually deny using synthetic images:
https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Image/Qwen_Image.pdf
>inally, the Synthetic Data category accounts for approximately 5% of the dataset. It is important to clarify that the synthetic data discussed here does not include images generated by other AI models, but rather data synthesized through controlled text rendering techniques (described in §§ 3.4). This excludes images synthesized by other AI models, which often introduce significant risks such as visual artifacts, text distortions, biases, and hallucinations. We adopt a conservative stance toward such data, as training on low-fidelity or misleading images may weaken the model’s generalization capabilities and undermine its reliability.

there are clearly more ways of slopping a model than using synthetic imagesets.

>>106370119
https://historia-arte.com/obras/bordando-el-manto-terrestre