Wuli-art

/

Qwen-Image-2512-Turbo-LoRA

like 214

Follow

Wuli-Art 109

Hi @Morac , this problem also roots from the original behavior of the Qwen Image 2512, which generates more realistic images compared to 2508, especially when using very short prompts.

Using prompt enhance could solve the problem:

This isn’t a problem with the acceleration LoRA. This model naturally has an extreme bias toward realism. Generating things like simple anime, flat art, 2D styles, and related derivatives is extremely difficult for this model. I’ve personally concluded that the older model is better in terms of prompt adherence, generalization, and overall quality.

@System36 We have discussed this with Qwen Image team earlier, using prompt enhance via LLM (e.g., qwen max) is highly recommended.

@aaroncisa I think the main problem is that it imposes a standard style on generation, which prevents the use of other styles and ends up killing creative freedom and generalization.

For example, when using tags like “anime,” “cartoon,” or “simple lines,” even with the LLM’s enhancer enabled, the style never comes out right. The model always “improves” the image in a way that pushes it away from the intended style and pulls it back into a generic, standardized look.

Edit: Apparently, he has excellent adherence to Lora training, which may help mitigate this problem.