Reposted from author: Lao Bai

A genuine full fine-tune of the Qwen large model, not a merged model.

Although the number of training images was limited, it has completely eliminated Qwen's inherent "greasy" AI look, achieving photo-realistic quality. Due to dataset limitations, some finer details may still have minor flaws.

Because the training dataset was small, the drift away from the base model is minimal, so it should be compatible with almost any other LoRA.

Quantization: I only performed FP8 quantization on [Lao Bai]'s original model so that users with low VRAM and lower-spec hardware configurations can experience it.

Original Post Link