Qwen-image:
The image quality is much higher than Wan2.1.
Wan2.1:
Drag the original image into ComfyUI to view generation parameters and workflow.
etc.
hunyuan:
Slightly over-trained; probably best to lower the weight a little.
Training style models is usually easier than training characters. That's mainly because it's harder to spot inaccuracies in a style compared to getting a specific character's face or features right. Right now, among the open-source models, I think HunyuanVideo does the best job at learning this particular oil painting style. Honestly, even Flux is just okay or passable at it. As for the other models, I don't think they're nearly good enough yet.
Currently, HunyuanVideo's image quality is way behind next-gen closed models like Google's, and even trails Flux. But, it is a video model, which might be why. Interestingly, I've found its anatomical understanding, especially for limbs, can be significantly improved with training – much more so than Flux or SD models. If the next version can just upgrade the base image quality, it would really fill that gap.
William-Adolphe Bouguereau：
