Face Distortion

[zrealli]

Hi,
when the face region is relatively small, it tends to become distorted. Is this due to the high compression ratio of dc-ae? Is there a solution to this problem? Would using a 2K image better?
Thank you very much for your wonderful work!!

[lawrence-cj]

Face distortion happens when the face is relatively small in the image. This problem will be mitigated in Sana-1.5 with DC-AE 1.5 later this year.

[zrealli]

Face distortion happens when the face is relatively small in the image. This problem will be mitigated in Sana-1.5 with DC-AE 1.5 later this year.

Thanks for your response! Could it be addressed by training facial dataset at a 2K resolution?

[lawrence-cj]

Thanks for your response! Could it be addressed by training facial dataset at a 2K resolution?

It will help if you try to fine-tune on some facial datasets with smaller faces.

[Deng-Xian-Sheng]

Facial datasets usually have large faces
Should we fine-tune a dataset with only faces or a dataset with very small faces?

[zrealli]

Thanks for your response! Could it be addressed by training facial dataset at a 2K resolution?

It will help if you try to fine-tune on some facial datasets with smaller faces.

I tried fine-tuning 200k+ facial datasets at 1.5K resolution, which resulted in a slight improvement, but it's still not a perfect solution. Looking forward to SANA-1.5 and new DC-AE!

[Deng-Xian-Sheng]

wow 200k OMG

…

--------------------------------

----- 原始邮件 ----- ***@***.***> ***@***.***>等3人 主题：Re: [NVlabs/Sana] Face Distortion (Issue #52) 日期：2024年12月04日 14:21:29 Thanks for your response! Could it be addressed by training facial dataset at a 2K resolution? It will help if you try to fine-tune on some facial datasets with smaller faces. I tried fine-tuning 200k+ facial datasets at 1.5K resolution, which resulted in a slight improvement, but it's still not a perfect solution. Looking forward to SANA-1.5 and new DC-AE! A.candid.snapshot.of.travelers.boarding.a.train.at.a.busy.station.with.luggage.in.hand.and.exciteme.jpg (view on web) — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[Muinez]

Face distortion happens when the face is relatively small in the image. This problem will be mitigated in Sana-1.5 with DC-AE 1.5 later this year.

@lawrence-cj will DC-AE remain an AE or switch to a VAE?

[Deng-Xian-Sheng]

Impossible to convert to VAE ----- 原始邮件 ----- 发件人：Muinez ***@***.***> 收件人：NVlabs/Sana ***@***.***> 抄送人：靳灿奇 ***@***.***>, Comment ***@***.***> 主题：Re: [NVlabs/Sana] Face Distortion (Issue #52) 日期：2024年12月05日 13点34分 Face distortion happens when the face is relatively small in the image. This problem will be mitigated in Sana-1.5 with DC-AE 1.5 later this year. @lawrence-cj will DC-AE remain an AE or switch to a VAE? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[Muinez]

Impossible to convert to VAE

why?

[zideliu]

Impossible to convert to VAE

why?

I think it's because the feature dimensions are different

[Deng-Xian-Sheng]

Impossible to convert to VAE

why?

If you think about it, this is interesting. They went to so much trouble (and published a paper) to switch from VAE to AE and achieve 32x compression.

Now, if you ask them to go back to VAE, they will punch the screen when they hear this.

It is estimated that training such a model costs 200 US dollars, in terms of renting computing power.