MrDeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
HaDenGUp-scaling the frames and removing blur using Topaz A.I. Gigapixel
#1
Hello all!

Recently, I have been experimenting with TopazLabs A.I image enhancement tools to get better fakes.

Mostly, I create 1280*720 fakes as I think it's the optimum setting considering time, hardware restrictions and quality.

This is a frame from my latest Daisy fake.

Original - 1280*720
[Image: jHeKhzxh.png]

And this the output of Gigapixel
Upscaled - 2560*1440
[Image: PcVlNfQh.jpg]

Both of them: You are not allowed to view links. Register or Login to view.

The original frame is not only got upscaled but also looks clearer and detailed, especially eyes look better.

Check the videos to compare the quality.
2560*1440: You are not allowed to view links. Register or Login to view.
1280*720: You are not allowed to view links. Register or Login to view.

The only downside, this process takes too much time.
Up-scaling 7732 frames took ~19 hours.

Mind that if your original frame doesn't look good to begin with -check bj scene- up-scaling will not help that.
There are also other tools provided by TopazLabs. Especially Adjust and Sharpen are useful to get more vibrant colors, removing noise, etc. But unlike Gigapixel, they don't support batch-process so It has to be done manually one-by-one.

Until more powerful GPUs are released, this is the way to create fakes in higher resolutions.
#2
This is awesome, I think this will be really helpful for lower resolution videos where the face detection is not as reliable.

Can you share your workflow? I assume you upscale the entire data_dst frames, before face extraction right? What about your data_src facesets,do you also upscale those?
~ Fake it till you make it ~
You are not allowed to view links. Register or Login to view.
You are not allowed to view links. Register or Login to view.
#3
Yeah, it can be useful for lower resolution videos too but I didn't use it for that purpose.
Workflow is simple, convert as usual but before creating the video, upscale merged folder and use those frames to create the video.
#4
So you just upscale the final video, I guess thats the best way to use it, upscaling some lower resolution faces in src dataset might be a good idea too.

I tried to use Waifu upscaler but it's is designed for drawn pictures so not much improvement there but this looks way better, might run all my videos through this, especially ones done with 720p videos.

Wonder how it would look if you were to upscale both src and dst before training from say 720 to 1080p and then upscale once more after merging from 1080 to 4k and downscale to 1080/1440p.
TMB-DF on the main website - You are not allowed to view links. Register or Login to view.
Want to request a paid deepfake or have any questions reagarding the forums or deepfake creation using DeepFaceLab? Write me a message.
Donations (BTC): 1H2vT4saNyULYtm42v2Wm83VuuWHTV4PHW
#5
Now... if only the super sampling in DFL could use the godly AI Model from Gigapixel instead of DCSCN...
#6
Omg that's brilliant man!
I am a long time lurker/faker, I have just signed up only to make this comment. Will give this a try for sure, thanks again!
To moderation, please make this thread pinned so other fakers may see and use this technique in their fakes.
#7
(08-27-2019, 01:06 AM)LCC Wrote: You are not allowed to view links. Register or Login to view.Now... if only the super sampling in DFL could use the godly AI Model from Gigapixel instead of DCSCN...

Yeah.. The only problem is it takes ~8 seconds to render a single frame on 1080 GTX
So for 1 min video in 30 frames per second: 30 * 60 * 8 = 14.400 sec = 4 hours. But if it can be ported to work on the cloud that would be something else.
#8
(08-27-2019, 02:39 AM)HaDenG Wrote: You are not allowed to view links. Register or Login to view.
(08-27-2019, 01:06 AM)LCC Wrote: You are not allowed to view links. Register or Login to view.Now... if only the super sampling in DFL could use the godly AI Model from Gigapixel instead of DCSCN...

Yeah.. The only problem is it takes ~8 seconds to render a single frame on 1080 GTX
So for 1 min video in 30 frames per second: 30 * 60 * 8 = 14.400 sec = 4 hours. But if it can be ported to work on the cloud that would be something else.

Naaah, I'm pretty sure the super resolution in DFL is only upscaling the 128-256 resolution output. Which is like 16,000-60,000 pixels vs 2,000,000 of 1080p. Should go really quickly Smile
#9
(08-24-2019, 06:44 PM)HaDenG Wrote: You are not allowed to view links. Register or Login to view.Yeah, it can be useful for lower resolution videos too but I didn't use it for that purpose.
Workflow is simple, convert as usual but before creating the video, upscale merged folder and use those frames to create the video.

what is meaned with "upscale merged folder" ?
#10
(08-29-2019, 12:52 AM)Malando Wrote: You are not allowed to view links. Register or Login to view.
(08-24-2019, 06:44 PM)HaDenG Wrote: You are not allowed to view links. Register or Login to view.Yeah, it can be useful for lower resolution videos too but I didn't use it for that purpose.
Workflow is simple, convert as usual but before creating the video, upscale merged folder and use those frames to create the video.

what is meaned with "upscale merged folder" ?

He basically makes a new folder with all the images he upscaled, and uses that to compile the video.
~ Fake it till you make it ~
You are not allowed to view links. Register or Login to view.
You are not allowed to view links. Register or Login to view.

Forum Jump:

Users browsing this thread: 1 Guest(s)