Reconstruct image fail. about renoise-inversion

[garibida]

Hi,
What model and scheduler do you use? Do you use my examples?
If your code is short and you can share it, it will be easier for me to help :)

[yuxu915]

hi, I use:

model_type = Model_Type.SDXL
scheduler_type = Scheduler_Type.DDIM

I didn't modify anything in inversion_example_sdxl.py.

[yuxu915]

hi, @garibida , I've also test inversion_example_turbo.py and get similar fail results like:

However, I test inversion_example_sd.py , get successful results.
The problem seems to be present in the SDXL model. Would you please check it?
Thank you.

[garibida]

Hi,

Did you use the environment.yaml or requirements.txt to create your environment? I created multiple environments and I wasn't able to reproduce this issue...

[ajrheng]

I had the same problem initially using my own environment and libraries, but creating the environment using environment.yaml solved the issue. Strange that slightly different versions of pytorch/diffusers can cause the model to fail completely.

[haofanwang]

It's related to diffusers version. A newer version has some differences.

[haofanwang]

I found the reason. The core difference is in the pipeline_stable_diffusion_xl_img2img.py between 0.24.0 and latest versions. You can load the pipeline from 0.24.0 and then everything goes well.

To be more clear, there is a typo in 0.24.0, and it has been fixed in following updates.

In latest version,

timesteps, num_inference_steps = self.get_timesteps(
            num_inference_steps,
            strength,
            device,
            denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None,
        )

in 0.24.0

timesteps, num_inference_steps = self.get_timesteps(
            num_inference_steps,
            strength,
            device,
            denoising_start=self.denoising_start if denoising_value_valid else None,
        )

The denoising_start is different, which leads to different timesteps and noisy results.

[yuxu915]

@haofanwang hi, your answer perfectly solved my question, thank you!

I found the reason. The core difference is in the pipeline_stable_diffusion_xl_img2img.py between 0.24.0 and latest versions. You can load the pipeline from 0.24.0 and then everything goes well.

To be more clear, there is a typo in 0.24.0, and it has been fixed in following updates.

In latest version,

timesteps, num_inference_steps = self.get_timesteps(
            num_inference_steps,
            strength,
            device,
            denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None,
        )

in 0.24.0

timesteps, num_inference_steps = self.get_timesteps(
            num_inference_steps,
            strength,
            device,
            denoising_start=self.denoising_start if denoising_value_valid else None,
        )

The denoising_start is different, which leads to different timesteps and noisy results.