Selur's Little Message Board

Full Version: Deoldify Vapoursynth filter
You're currently viewing a stripped down version of our content. View the full version with proper formatting.
In the dev version there is strange bug.

If I start from this configuration

[Image: attachment.php?aid=3296]

Then I set Denoise to "none"

[Image: attachment.php?aid=3297]

Color tweaks is set to "none" and then is disabled (as should be)

[Image: attachment.php?aid=3298]

But if I change Denoise again and set it for example to "light", Color tweaks remain disabled. 
The only way to unlock Color tweaks is to change the Coloring model, for example to  "Zhang(ECCV16)"

Please fix it.

Thanks,
Dan
Uploaded a new test version.

Cu Selur
The new dev version seems Ok.

I attached the new RC3. In this update I added another combination method called Chroma-Retention as described below

Code:
:param CombMethod:          Method used to combine coloring models with (+):
                            Allowed values are:
                              'Simple' (default)
                              'Constrained-Chroma'
                              'Luma-Masked'
                              'Adaptive-Luma'
                              'Chroma-Retention'    <--- NEW METHOD

The problem of finding a way to stabilize the DDcolor family models is very difficult.

You can find an example here

 [Image: attachment.php?aid=3299]

As you can see DeOldify is the one with more stable colors, and compared with the others is stable like a rock.

The problem is that even DeOldify isn't perfect. In the picture above the patient's legs (near the stretcher) are gray in DeOldify, while DDColor and ECCV16 were able to colorize them, but at cost of grater color instability.  

In this link you can find the sample colored with the 4 models supported by HAVC: https://archive.org/details/young-franke...720p.-test
You will see that the colors instability of DDColor, ECCV16 and Siggraph17 is terrifying.

So the problem is: it is possible to stabilize DDColor, ECCV16 and Siggraph17 and at the same time to retain the colors ?

In this case it could be possible to use these models to improve DeOldify.

The HAVC filter attempts to solve this problem. 
The problem isn't fully solved yet, but this RC3 represents a small step forward in solving this complex issue.

I hope that you have time to add the new combination method.

Thanks,
Dan
uploaded a new test version

Cu Selur
The new dev version it's Ok.

Thanks,
Dan
Wonderful work you're doing.  I run the b&w clip thru Retinex first and then colorize it (Vdieo+Artistic, fast, veryvivid) .  In the first half, the green chalkboard appears to be quite stable, the 2nd half of the video especially the dark scenes, not that great.  Perhaps this is direction you can further explore.  

Here is a link to the video:
https://gofile.io/d/6Z2d0x
Unfortunately Retinex introduce another type of artifacts see for example at time 00:02:02 of your clip.

Try these settings

[Image: attachment.php?aid=3303]

Dan
Hello Selur,

  I found a regression bug in the last dev version.

  If I select for example the method "external ClipRef same as video", Hybrid should allow to select a file in input, but instead it possible to select only a directory

[Image: attachment.php?aid=3312]  

  Moreover I decided to add the Retinex filter in HAVC, this is a patched version with no defect (no need to load the plugin in the script), so please change the name of the "B&W mode" in HAVC and ColorAdjust from "SATA - auto brighness&constrast" in "Multi-Scale Retinex on Luma".

Thanks,
Dan
"last dev version" is Hybrid_dev_2025.10.03-41135 or do you mean the last test version?
I uploaded a new test version. Let me know whether the problem persists.
I changed "SATA - auto brighness&constrast" in "Multi-Scale Retinex on Luma".

Cu Selur