The inference speed of this model on 1k or 2k images didn't show any acceleration?

[microbenh]

[Huage001]

Hi,

Our experiments indicate that at 1k resolution, the original model is indeed faster. At 2k resolution, CLEAR yields slight acceleration. Please refer to Fig. 2 and Tab. 7 of the paper for details.

A recommended use case is to adopt original model at 1k resolution and then upscale the results with CLEAR.

[Remember2015]

Hi,

Our experiments indicate that at 1k resolution, the original model is indeed faster. At 2k resolution, CLEAR yields slight acceleration. Please refer to Fig. 2 and Tab. 7 of the paper for details.

A recommended use case is to adopt original model at 1k resolution and then upscale the results with CLEAR.

if CLEAR is recommended in SR cases, what is the difference between other solutions, like aurasr/ccsr2/seesr?

[Huage001]

if CLEAR is recommended in SR cases, what is the difference between other solutions, like aurasr/ccsr2/seesr?

Hi,

Through my own experience, super-resolution-based methods may upscale an image more faithfully, while resolution-extrapolation-based methods using diffusion models may help create more details related to the images and text prompts.

You can also refer to the discussion in the project page of Demofusion (SDXL+SR vs. DemoFusion).

[Remember2015]

if CLEAR is recommended in SR cases, what is the difference between other solutions, like aurasr/ccsr2/seesr?

Hi,

Through my own experience, super-resolution-based methods may upscale an image more faithfully, while resolution-extrapolation-based methods using diffusion models may help create more details related to the images and text prompts.

You can also refer to the discussion in the project page of Demofusion (SDXL+SR vs. DemoFusion).

Thanks for your reply, I will try. Examples above really helped me got the difference between super-resolution and resolution-extrapolation, thanks~