so where's all the Topaz Video Enhance AI discussion?

[Corvius]

My God... what is wrong with you guys?

Such extreme views on some software. I've never seen anything quite like this before, except maybe when dealing with actual fake scam applications. That's really why you think VEAI is?

I use Hybrid, VEAI, Avisynth, Adobe Premiere, and Davinci Resolve. Everything has its use and place.

To each his own I guess...

[VoodooFX]

Such extreme views on some software.

Probably to scare off the computer-illiterate people, adepts know its use cases.

Bigger terrorism is when all over the places Sir lordsmurf is claiming that all logos are irremovable.

[lordsmurf]

Bigger terrorism is when all over the places Sir lordsmurf is claiming that all logos are irremovable.

While your tool is better than others, it's still a case of replacing a logo with overall semi-blurry artifacted video. It's not removal. And due to human perception, the artifacty video can be more distracting than the original logo.

[VoodooFX]

While your tool is better than others, it's still a case of replacing a logo with overall semi-blurry artifacted video. It's not removal. And due to human perception, the artifacty video can be more distracting than the original logo.

Here is a random logo removal, a last one I did:
https://gateway.ipfs.io/ipfs/QmYfA6zMUeENoAFEKtHMQK1e1yYGbs1MuNoGJJ78fMJPLo/vlcsnap-20...29m59s841.webp
https://gateway.ipfs.io/ipfs/QmYfA6zMUeENoAFEKtHMQK1e1yYGbs1MuNoGJJ78fMJPLo/vlcsnap-20...30m08s622.webp
https://gateway.ipfs.io/ipfs/QmYfA6zMUeENoAFEKtHMQK1e1yYGbs1MuNoGJJ78fMJPLo/vlcsnap-20...30m44s998.webp

Please, paint on any picture the place where logo was, shouldn't be hard for you to spot the blurry distracting artifacts.

[rrats]

For the love of God, upload images on the VideoHelp's own image server, it's not that hard and it makes em actually load

[lollo]

Here is a random logo removal, a last one I did

Impressive!

[lordsmurf]

For the love of God, upload images on the VideoHelp's own image server, it's not that hard and it makes em actually load

Or better yet, don't attach images at all. What betrays "logo removal" is the motion in video clips. I can use MS Paint and "delogo" an image. But the motion is what actually matters.

[VoodooFX]

Or better yet, don't attach images at all. What betrays "logo removal" is the motion in video clips. I can use MS Paint and "delogo" an image. But the motion is what actually matters.

Or better stop posting nonsense on things where you hardly have clue about.
FYI, there are three basic logo removal methods, two of them has nothing to do with motion, so looks like you know only about one of them.

Still waiting for your MS Paint skills painting the place where logo was, use the forensic filtering if you need.

Attached the clip with motion.

[lollo]

Attached the clip with motion.

Impressive!

I had no doubt about your results, even without watching the video instead of the pictures. LS probably wanted to stress about the fact that showing the full video is more convincing than static images.

[lordsmurf]

Or better stop posting nonsense on things where you hardly have clue about.

No clue? Amusing.

Impressive!

That clip doesn't impress me in the least. Why? Two factors.

(1) Cel animation is far easier to work with, be it encoding, or restorations. In this exact case, the logo can be consider unwanted noise, and attempted removal is a restoration.

(2) The excess grain and errant noise makes it easy to hide flaws. This is an old restoration trick, adding less-bad noises to conceal imperfect removal of others.

Now, if Voodoo here wants to impress me, I'll find him a far more typical clip of a logo that needs removal. Not a cartoon, not a faint watermark, not animation. I'll probably be wasting my time, but I at least want to provide a nice long varied clip. Then it's a case of "STFU or prove your magic method".

Logo removal is one of those idiotic irksome claims that goes back decades. After a while, years, literal decades, you get tired of it.

LS probably wanted to stress about the fact that showing the full video is more convincing than static images.

Lots of faulty and crappy methods (and people) try to hide their junk in stills, or by "cherry picking" (selective editing, to the point of dishonesty and BS). Stills for video is an easy one, when you're trying to show something that is 100% affected by motion. Nice try.

-- and that brings us full circle to the topic of this thread: Topaz, artifacts, crappy upscale/sharpening, and mostly-BS "AI" used for marketing to the lemming masses.

[VoodooFX]

LS probably wanted to stress about the fact that showing the full video is more convincing than static images.

Nah, he is sincerely ignorant on the subject.

Anyway, progress was made, he found bunch of excuses instead of claiming that all logos are irremovable.
Examples of "expert" advises:
https://www.digitalfaq.com/forum/video-editing/8974-how-remove-watermark.html#post55837
https://www.digitalfaq.com/forum/video-restore/9334-remove-video-logo.html#post58637

I'll probably be wasting my time, but I at least want to provide a nice long varied clip. Then it's a case of "STFU or prove your magic method".

See, I don't claim that all logos are removable without a trace, it's you who claim[ed] that they are all irremovable, but somehow you couldn't locate the size as a truck logo:



Maybe here you can show your MS Paint skills: Attached non-animation video with various scenes, slow, fast, horizontal & vertical movement.
Hint: Size of a Ford Focus logo.

EDIT:

something that is 100% affected by motion

Please explain how Layer() function affects motion.

[lordsmurf]

You had two options here:
(1) Discuss current logo removal methods (using your InpaintDelogo, initial released in 2019, not really mature until later releases), which may make past years/decades of advice less accurate (albeit still not really wrong).
(2) Be a troll/dick.

You chose the latter.
So I decline further interaction.

But if that Fleischer sample truly is a before/after (and you're not just trolling), then it is interesting. I'd need to see the scripts/whatever, and be able to recreate myself. And then attempt on projects where I've wanted to remove logos for years/decades (and nothing was ever satisfactory, not even with InpaintDelogo).

I see logo ghosts on the live action sample, and it's noticeable on an HDTV. Also bad encoding glitches (too much noise, not enough bitrate). You're hiding leftover removal in noise, I'm not fooled, that's a standard trick. Sometimes it works well, sometimes laughably not at all.

This side conversation is derailing the thread.

[VoodooFX]

There is nothing fundamentally new in the past ~decades, so there is nothing new in InpaintDelogo too, just various minor tweaks.
Same you could do ~decades ago, but obliviously you prefer to troll.

You're hiding leftover removal in noise, I'm not fooled, that's a standard trick.

I'm not hiding anything, samples are transparent encodes of the sources on hand, you come-up with weird excuses for your ignorance.

I see logo ghosts on the live action sample, and it's noticeable on an HDTV.

Cool, where is your sample with a painted location where logo was?

[lollo]

Impressive!

That clip doesn't impress me in the least

But if that Fleischer sample truly is a before/after (and you're not just trolling), then it is interesting.

[Cornucopia]

The Fleischer clip is fairly straightforward. Use the features of the logo against itself.
Threshold (of local contrast) sense edge detection, with possible median, to create a mask, then apply simple darken/dyn range reduction to the area convered by the mask, then downstream 2nd mask of edge+surrounding border to apply minor edge blend blur. Slight addition of gaussing noise and subsequent re-encode would basically hide any remaining incongruities.


Scott

[lordsmurf]

I'm not seeing what's humorous here. I'm not impressed, for the reason mentioned. But I do see some potentially interesting granularities in the processing, having now seen a before/after (assuming it's legit, not trolling).

For the record, poisondeathray, StainlessS, jagabo and johnmeyer impress me greatly (and some others I'm surely forgetting off-hand here), I'm sometimes in awe of what they've done. So it's not like I can't be impressed. And I think you (lollo) are quite competent at Avisynth.

Topaz doesn't impress me in the least, hatchet job of upscale+sharpen.

InpaintDelogo is more advanced that previous attempts to de-logo, but it's not flawless magic. Voodoo needs to take credit for what he's accomplished (kudos!), but not try to overstate the ability of his tool, nor be a tool himself.

The Fleischer clip is fairly straightforward. Use the features of the logo against itself.
Threshold (of local contrast) sense edge detection, with possible median, to create a mask, then apply simple darken/dyn range reduction to the area convered by the mask, then downstream 2nd mask of edge+surrounding border to apply minor edge blend blur. Slight addition of gaussing noise and subsequent re-encode would basically hide any remaining incongruities.
Scott

The granularity of the edge/mask/NR is what may be interesting here. But I'd want to see the script. Avisynth has so many possibilities, and nobody can know everything about it, even though some of us try. I know some of this has gotten better in recent past years, but I've not been able to keep up to date on it.

Generally with "logo removal" the post-removal attempts to hide are hatchet jobs. (Most often, the removal attempt itself is a hatchet job.) And indeed, the live action sample above has many tell-tale artifacts of this nature, especially viewed large. But with toons being easier, any slight improvements in methods might make it more believable at a non-small (preview) size. Or at least easier to ignore.

Noting that almost no logos are grades of transparent these days, so it has limited values. Even with the current-best methods, most logo "removals" still create distracting blurry blobs. (I never understand the willful ignorance to state the obvious here. It ain't removed.)

I've seen actual frame-by-frame repair, and I'm sure you have as well. It can actually remove harsh problems, and that would include logo-type noise (constant film issues frame-to-frame, types of cyclical damage). This is probably something I would have gotten into at work, had health not cut my career short. The software was starting to become more viable, and the need greater where I was (as we'd started to enter uglier sources, and the set-aside projects).

[lollo]

I'm not seeing what's humorous here.

Your whole aptitude. You started with "it is not possible to remove a logo in an effective way", continuing with "you are a troll" and ending with "interesting results". If you read back yourself, you'll find it's weird.

I was impressed by the showed results, you replied you're not. Fine for me, but then when you changed your "tone" I had to smile

The funny thing is that I also think that logo removal is difficult and produces somehow contradictory results, and VoodooFX also agree with us
But I also thinks that all efforts from doom9's avisynth developers, such as "master" VoodooFX, deserve more consideration and support!

[lordsmurf]

I'm not seeing what's humorous here.

Your whole aptitude. You started with "it is not possible to remove a logo in an effective way", continuing with "you are a troll" and ending with "interesting results". If you read back yourself, you'll find it's weird.

Not really. The cartoon sample wasn't a logo, but a watermark. Not the same. That watermark had at least 50% opacity, and that can often be corrected. It's very content based. It can give you a sample of a watermark that I had issues removing, and decided to just manually fix the segment I needed.

Logos tend to be located in corners, and not a simple transparency. Unless the content is very cartoon-like (large area of solid colors), those can't be removed without artifacts. And those artifacts are often more distracting than the logo. At least the logo was static, the blurry blob is an obnoxious amoeba moving on the screen.

The funny thing is that I also think that logo removal is difficult and produces somehow contradictory results, and VoodooFX also agree with us

It's why I don't understand his attitude in this thread. ("He started it!")

But I also thinks that all efforts from doom9's avisynth developers, such as "master" VoodooFX, deserve more consideration and support!

I agree. But don't feed me dog shit and call it pudding. Be honest about limitations.

[lollo]

But don't feed me dog shit and call it pudding

Nobody did that.

All started with:

Bigger terrorism is when all over the places Sir lordsmurf is claiming that all logos are irremovable.

and you had evidences with 2 videos that it was possible.

In any case, VoodooFX does not need me to defend him, I was just impressed by (some of) his results and your comments. I stop here!

[MrEos]

I spent $200 dollars just to find out I had to spend the last 5 days learning how to ivtc. Starting with Stargate SG1 S1E1.

With that being said. I'm still going to include VEAI in my workflow.

Which, during that 5 days, turned into DVDfab to VOB > DGIndex (force film) > Virtualdub2 > (5 days learning ivtc techniques) > Learned the SeperateFields() command > Realized there are Blended frames that are resulting in residual chroma ghosting at certain scene changes that no amount of ivtc tweaking is getting rid of. Thinking an override file is the only way to resolve..

Then decided to just go back to handbraking a file out, throwing it at VEAI, and being done with it. Because the thought of trying to jump through that many hoops for... 220+ episodes. Is not sustainable.

With that being said. Where do I find the repository of scripts built by this community to fix all these mistakes in common DVD series such as Stargate, so I don't have to jump through all these hoops to get a proper detelecined source to work with, free of blended frames.

That's what this place really needs. Community guides for fixing common Sci-Fi DVD series.

Where the first post is always updated with the most up-to-date method of acquiring the best possible detelcined source possible for throwing at programs like VEAI or for further refinement in Vdub/Synth, etc.

That might make the thought of spending more time with more advanced solutions reasonable to consider. Because I'm not writing out override files for 220 episodes by myself. Less time spent on fixing this stuff could mean more time spent learning to use filters to make your claims about topaz come true.


Forgot the most important part of the post. Getting knowledgable on Hybrid/Avisynth/VirtualDub2, is like playing whack-a-mole. And that's just to IVTC. Nevermind how many dead end web searches a person will go through to make those programs try to match or exceed topaz on a grand scale such as 220+ episodes.

You have no idea how hard it was to just to find this post. https://forum.videohelp.com/threads/401305-TIVTC-settings. while having come across it once before but forgetting to bookmark it. It took luck to stumble across it again.

[Alwyn]

Then decided to just go back to handbraking a file out, throwing it at VEAI, and being done with it.

How'd it turn out?

[MrEos]

For the time effort involved in a Handbrake > VEAI

vs learning the more complicated methods.

Somewhat worth it.

It came down to spending the 100+ for the blu-ray set, or spending 200 for VEAI and having something to throw at a whole collection of dvd's after that.

I can see jumping down to Vdubing the more favored episodes. But, for a collection that is likely to be re-ripped at some point in the future when technology advanced. I didn't see the point in trying to perfect it on my own. That's why I just went back to batch converting with handbrake it's like 99 percentile issues from handbraking vs trying to TFM().TDecimate() and figuring out if DGindex should honor pulldown or force film. etc.

The whack-a-mole game of learning it is giving me gray hairs.

And I hate that for as long as everyone has been at this. That one single thread I linked to about sums up the total combined effort to bring forth a solution to fix any specific telecined DVD source. Like the internet is just going to waste... Like I'm starting from scratch instead of having something to work off that has had 1,000's of people working on over the years. Not including the plugins of course.

If I can thank VEAI for anything, it was to send me on a quest to make a proper progressive source backup. But it's not a chore for the faint of heart.

https://www.reddit.com/r/Stargate/comments/kaxnme/cba_for_upscaling_your_stargate_dvds_for_personal/

is very close to what I settled on for now, until I feel like playing whack-a-mole again.

[johns0]

Can you please stick to the topic of this thread and not bash each other over whose logo removal is best?

[Alwyn]

Logo not mentioned in last 3 posts (not including yours), and not after 19Sept.

[forart.it]

For those who wanna test various super resolution "AI" (= neural networks) models, we just found this interesting repo (and relative colab):

Repository to use super resolution models and video frame interpolation models and also trying to speed them up with TensorRT. This repository contains the fastest inference code that you can find, at least I am trying to archive that. Not all codes can use TensorRT due to various reasons, but I try to add that if it works. Further model architectures are planned to be added later on.

Currently working networks:

• ESRGAN with rlaphoenix/VSGAN and HolyWu/vs-realesrgan
• RealESRGAN / RealESERGANVideo with xinntao/Real-ESRGAN and rlaphoenix/VSGAN
• RealESRGAN ncnn with styler00dollar/realsr-ncnn-vulkan-python and media2x/realsr-ncnn-vulkan-python
• Rife4 with HolyWu/vs-rife
• RIFE ncnn with styler00dollar/VapourSynth-RIFE-ncnn-Vulkan and HomeOfVapourSynthEvolution/VapourSynth-RIFE-ncnn-Vulkan
• SwinIR with HolyWu/vs-swinir
• Sepconv (enhanced) with sniklaus/revisiting-sepconv
• EGVSR with Thmen/EGVSR and HolyWu/vs-basicvsrpp
• BasicVSR++ with HolyWu/vs-basicvsrpp
• RealBasicVSR with ckkelvinchan/RealBasicVSR
• RealCUGAN with bilibili/ailab
• FILM with google-research/frame-interpolation
• PAN with zhaohengyuan1/PAN
• IFRNet with ltkong218/IFRNet
• M2M with feinanshan/M2M_VFI
• IFUNet with 98mxr/IFUNet
• eisai with ShuhongChen/eisai-anime-interpolator
• SCUNet with cszn/SCUNet
• GMFupSS with 98mxr/GMFupSS
• ST-MFNet with danielism97/ST-MFNet
• VapSR with zhoumumu/VapSR

Also used:

• TensorRT C++ inference with AmusementClub/vs-mlrt
• ddfi with Mr-Z-2697/ddfi-rife (auto dedup-duplication, not an arch)
• nix with lucasew/nix-on-colab
• custom ffmpeg with markus-perl/ffmpeg-build-script

VSGAN-tensorrt-docker

PLEASE DON'T ABUSE.

[Alwyn]

Some questions for you Topaz Video AIers out there:

How do you incorporate TVAI into your workflow? Is it capture>TVAI>NLE>Export?

What output formats does TVAI have?

I know there's a trial available, but I thought a current user who knows the program would be better to ask.

Thanks for your feedback.

[lollo]

Is it capture>TVAI>NLE>Export?

For your Analog captures that flow is not appropriate, because TVEAI does not handle interlaced nor telecined material effectively.

Use: Capture -> AviSynth/VapourSynth/VirtualDub -> TVEAI

The export can be a sequence of images, or a final h264 encode. Not much freedom on the parameter of the last, just sort of "quality" setting (similar to crf). If you plan a further NLE processing the h264 is probably not appropriate.

I use sometime TVEAI just as upscaler, at the end of an AviSynth flow, in alternative to nnedi3 or similar.

Although there is nothing TVEAI can do that cannot be matched or beaten by AviSynth/VapourSynth, you may find an interest in it, because your preference to GUIs and straighter approach.

[rgr]

But remember that VEAI likes to change colors, especially with RGB input. After the color mosaic conversions, I made a script where VEAI least (slightly) distorts the colors.

ffms2("input.mts")
convertbits(bits=10)
levels(64,1,1016,64,940,coring=false) #adjust to your needs, here convert 16-255 > 16-235

[Cyrax9]

Some questions for you Topaz Video AIers out there:

How do you incorporate TVAI into your workflow? Is it capture>TVAI>NLE>Export?

I'm going to try to answer these in order as best I can Alwyn. For your first question it really depends on the source footage and whether it's a personal project or a professional one. (I can't go into detail, but I've literally had professional projects where the source footage is the company pointing to a Youtube link and saying "just use this," because for one reason or another they don't have access to proper masters.) For professional projects, if what's provided is already deinterlaced source material, I typically export an image sequence from ffmpeg or my NLE and just run that through TVAI when speed is prioritized over quality. (E.g. "make it better, but do it fast.) If I have time to actually work on the footage, I may try and do some preprocessing before exporting the image sequence, even if it's just some basic color correction. In some cases the source material is truly atrocious, (e.g. blatantly wrong aspect ratio, "fake" HD that was achieved by literally tossing the footage into a timeline and scaling it to the size of a frame, etc.) in which case I'll try to fix those issues before I even touch TVAI. I've found for example that downscaling "fake HD" actually improves the output as TVAI has no clue what to lock onto from HD clips that are non–AI upscaled SD clips.

What output formats does TVAI have?

There may be others that I haven't looked into, but right now I know that TVAI supports H.264, ProRes (LT, 422 HQ, 4444, not sure about 422 without the HQ option,) and image sequences in JPEG, PNG, and TIFF, with both 8 and 16–bit versions of the latter two. I highly recommend working with image sequences if you have the storage space for them, even if you need a huge mechanical HDD to hold your image sequences until processing is complete. Why image sequences? Because in my experience TVAI's implementation of ffmpeg doesn't properly match the duration of the output file that other implementations of ffmpeg use, and thus doesn't match the duration of the source video properly either, even on constant frame rate material. With image sequences one frame in = one frame out, and I know that I'll get the results that I expect. Likewise, while not nearly the issue it was in early versions of TVEAI version 2, should TVAI version 3 crash on me, I won't lose hours of processing and as long as I can get my settings back, can just tell it to pick up from the frame that I left off on as if nothing had happened. (On older versions of TVEAI I also noticed that 852x480 source material frequently crashed the app if it wasn't converted to an image sequence first, although I have not noticed this issue in TVAI 3.)

I know there's a trial available, but I thought a current user who knows the program would be better to ask.

Thanks for your feedback.

Definitely give the trial a try, I was pleasantly surprised with the output on certain videos that I expected far less from. In some cases the improvement is minimal, in others it's actually worse, and in some it's a drastic leap in perceived quality. Source material matters a lot.

Also, one thing I've been curious about since a lot of people have been using TVAI to upscale classic 90s Sci–Fi, is if there's a way to preserve the variable frame rates of the source material, particularly for shows that blended 23.976fps film footage and 29.97fps material used for VFX. Basically, I'm wondering if there's a way to split the source material into a series of 23.976 and 29.97fps clips so that I could deinterlace the material that should be deinterlaced and IVTC the material that should be IVTC'd instead of applying the "wrong" method to different portions of the video. If there's a way to do this, I may experiment with it on a personal project and then see if I could apply the same process to extract constant framerate sections from variable framerate source material to cell phone footage on certain professional projects when needed. (Occasionally I do wind up having to work professionally with material from non–professional sources.)

Right now one of my other biggest headaches is trying to figure out how to get the Mac version of AviSynth+ set up with TDeint and QTGMC so that I can use it for preprocessing on there. Unfortunately the documentation for the Mac version of the software is almost nonexistent, which isn't good when I need a tutorial or two so I can get started with it.

[Alwyn]

@Cyrax9, thanks very much for that. Food for thought.