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

[matt8000]

Ok, you just read the title.

To my understanding, the key point was that the expert people you mentioned stated that you can obtain an equivalent/superior result with AviSynth filters. I agree with them.

My post was for people who were looking to get the best results for their needs and I think I demonstrated by the provision of evidence that the product should be certainly be considered. Am I saying it is the only solution? No. Am I saying the product performed great for me? Yes. Have I proved my case? I believe so. Let the public decide for themselves.

Matt.

[lollo]

Did I say your result was not good? No. Did I say you did not provide evidence? No. Did I say that Topaz is not worth a try? I do not believe so.

I just added that AviSynth processing is superior to Topaz VEAI in my experiments (and in experts opinion). Let the public decide for themselves. (end of this useless discussion)

[matt8000]

Did I say your result was not good? No. Did I say you did not provide evidence? No. Did I say that Topaz is not worth a try? I do not believe so.

I just added that AviSynth processing is superior to Topaz VEAI in my experiments (and in experts opinion). Let the public decide for themselves. (end of this useless discussion)

I provided the post. You provided the rest.

[JoelHruska]

I welcome input from anyone who can demonstrate how to achieve superior results with AviSynth alone as opposed to AviSynth + TVEAI. I find it very strange that people are wedded to the idea of using these two utilities in an either/or fashion. I don't use AviSynth *or* TVEAI. I use AviSynth *and* TVEAI. The only exception to this is when I work with the Dione upscaler model, which is explicitly intended for deinterlacing and doesn't work very well yet.

As for scripting: Choices here depend somewhat on how much sharpening you prefer and whether you are looking to treat the show frame-by-frame, scene-by-scene, or episode-by-episode. I designed my initial project *explicitly* as a one-shot solution for people who wanted a way to treat the entire show. There are TVEAI models that create detail in Deep Space Nine -- the Theia models -- but I don't use them.

The script I'm currently using:

TFM(pp=5, mode=2, micmatching=3).TDecimate(hybrid=1)
QTGMC(ShutterBlur=3,ShutterAngleSrc=180,ShutterAng leOut=180,SBlurLimit=8,Preset="Very Slow", TR2=3, InputType=1,SourceMatch=3, MatchEnhance=1.0, MatchPreset="Very Slow", TR0=2, TR1=2, MatchPreset2="Very Slow", sharpness=0.6, SMode=2, Rep0=22, Rep1=22, Rep2=22, RepChroma=True, Sbb=3, SubPel=4, NoiseProcess=1, ChromaNoise=True, GrainRestore=0.25, NoiseRestore=0.5,DenoiseMC=True, NoiseTR=2)
pSharpen(strength=65, threshold=90, ss_x=4.0, ss_y=4.0)
MAA2(mask=1, chroma=true,ss=4, aa=128, aac=128, threads=8, show=0)
LSFmod(defaults="slow", strength=60, Smode=5, Smethod=3, kernel=11, preblur="ON", secure=true, Szrp=16, Spwr=4, SdmpLo=4, SdmpHi=48, Lmode=4, overshoot=1, undershoot=1, Overshoot2=1, Undershoot2=1, soft=-2, soothe=true, keep=40, edgemode=2, edgemaskHQ=true, ss_x=4.0, ss_y=4.0, dest_x=704, dest_y=480, show=false, screenW=1280, screenH=1024)

EDIT: I do not know why the forum is inserting a space in the script above, in the "ShutterAngleOut=180" line. There is no space in my forum post in the editor. You'll have to remove it if you want to test. I've tried to remove it repeatedly, but even re-copying and pasting the script from the AviSynth file is not correcting the problem.

This is sharper than what I've published as part of the DS9 Upscale Project. I've been experimenting with sharpening filters and strengths and I'm still experimenting with the balance between what's handled in AviSynth and what's handled in TVEAI. Anyone who wants to play with noising output to improve results in either Artemis-HQ or Gaia-CG 1.5.3 should be aware that old GCG responds more strongly to this method. If you want to push "NoiseRestore=0.5" higher than that, you need to remove "ChromaNoise=True," or your output will turn greenish. Noising Gaia-CG in TVEAI 1.5.3 can produce superior results to Gaia-CG without noise, but Gaia-CG's sharpening is also stronger than Artemis-HQ's, and if you over-sharpen in AviSynth, Gaia-CG will both magnify and sharpen the existing noise.

Artemis-HQ will absorb more sharpening than Gaia-CG 1.5.3, but it sometimes picks up background errors that GCG doesn't. Modern Gaia-CG does not produce great results in DS9 and Gaia-HQ is downright bad for it. Artemis-MQ and Artemis-LQ both produce heavy blurring. Dione output restores more stars and avoids removing some particle effects in battle scenes, but it also contains heavy deinterlacing artifacts and can only be hand-blended back into output if you want the benefit from using it. Proteus is an interesting choice, but the fact that it offers hand-tuning also means it's impossible to hand-tune the output for the entire show. I checked its output by averaging the estimated values for the six different variables after taking samples at 15 different points throughout an episode. Overall: Not bad. But not cleanly better than others, IMO.

Deliberately over sharpening in AviSynth before feeding a model through Artemis' Strong DeHalo model can produce better results in specific areas as compared to not doing so and using a standard upscale model. A comparison is below: Standard upscale on the left, deliberately oversharpened before repair is on the right. Kira's eyes are better and sharper in the latter output, and the color is better. There's a slight blurring to the background but I'll take the foreground detail, personally:

https://imgsli.com/NjI3MTY

Output on the right was deliberately broken before upscaling. Output on the left wasn't. This method can only be deployed in specific places; it will destroy output in others. Sometimes, however, Artemis Strong DeHalo footage can be combined with Artemis Antialiasing and Moire to either remove haloing artifacts or avoid heavy aliasing. In the video clip posted back on Page 8, I'm quite pleased with how I was able to blend various AI model output to improve the motion quality of the grating mounted on the underside of the Miranda-class model. While my version is not perfect, it is substantially less jaggy than the original DVD footage. An AI model that eliminates aliasing at the cost of heavy blurring can be combined with a model that doesn't to yield a "Best of Both Worlds" (if you'll pardon the pun) approach. Hit the "Magnify" button above and explore with the slider -- you'll see multiple places where deliberately oversharpening output and then using the repair model *fixed* errors compared to sharpening in AviSynth normally followed by upscaling.

Sometimes making an error worse in base footage seems to goose TVEAI into repairing it more effectively. This is very situational.

I am very interested in any tweaks or improvements to AviSynth scripting that anyone would care to suggest. I am always interested in improving output. The script above is what I have arrived at after 17 months of trial and error, but it has been meaningfully improved by outside contributions in the past and I am always curious to test ideas.

All comments above regarding the performance of various Topaz Video Enhance AI models and how they interact with AviSynth should be understood to apply solely to Star Trek DS9 and, to a lesser extent, Star Trek Voyager. My research and work has concentrated entirely on these two shows and my tuning is intended for DS9. The DS9 Upscale Project is also best-suited to DS9 Seasons 3 and later. Bad DVD transfers in Season 1 and Season 2 substantially limit the amount of recoverable detail and nothing I've discovered has solved that problem. For a quick comparison, see here:

https://imgsli.com/NjI3MTg

This is an unmodified copyright notification shot from the credits. It's one of the only frames in common between S1 and S7 of the show, since the S4+ credits are different from S1-S3. It shows how the Season 2 DVDs more closely resemble VHS, while the S6 DVDs look more like DVDs. This limits the amount of recoverable detail between the early and the later episodes of the show.

[lollo]

I find it very strange that people are wedded to the idea of using these two utilities in an either/or fashion

In my flow I simply replace the "upscale section" of an Avisynth script (nnedi3) with the "upscale feature" from Topaz. All the rest (deinterlace, denoise, ...) is still in place.

The first is better in my case, but YMMV.

https://forum.videohelp.com/threads/401910-Topaz-products

In any case, thanks for sharing your good results and your script

edit: thanks added

[JoelHruska]

There's an AI upscaler filter for AviSynth too, but I have not had great luck with it. https://github.com/Alexkral/AviSynthAiUpscale

Another interesting option is madVR + QTGMC'd output. I have not attempted to fine-tune this configuration the way I've worked via AviSynth, but I've done enough with it to know that running QTGMC on output prior to madVR produces much better results than just using madVR. QTGMC, as far as I'm concerned, is darn near a miracle filter. Right now, I'm trying to maximize sharpening while minimizing halo-ing. That's why I'm working with multiple sharpening filters at varying levels of strength -- I've been running output to experiment with the balance between edge sharpening, non-edge sharpening, and dehaloing. TVEAI's "Strong DeHalo" is (ironically) better than their "DeHalo" filter because it applies less intrinsic blurring.

I continue to experiment with various sharpening approaches because I would much rather sharpen without needing to repair than repair after sharpening. And if anyone has any advice on how to further prevent deinterlacing issues or alternative methods, I'm all ears. The TFM/TDecimate calls above appear to work perfectly ~99.9% of the time but there are occasional errors. There's a single frame missing from the introduction to the credits (In the original M2V file, Frame 6904 gets skipped by the IVTC. Frames 6902 - 6903(repeat)-6904 (skipped) - 6905 are rendered as frames 5521 and 5522 -- but frame 6904's output is missing.

This occurs in two places in the episode. In both cases, the issue can be resolved by extracting and upscaling the missing frame followed by manual reinsertion into the file stream, but I'd prefer not to have to manually fix files at all. There's no problem with audio sync, because both occur immediately before or after black frames and both occur during instrumental pieces, allowing a single frame removal in both cases to keep things sync'd. I can tolerate 2-3 single frame errors per episode when unspooling an entire TV series, but I'd like to eliminate them if it can be done.

Thank you for your thank-you. I am glad you like the output or find it useful. Constructive criticism is welcome.

[lordsmurf]

I find it very strange that people are wedded to the idea of using these two utilities in an either/or fashion.
All comments above regarding the performance of various Topaz Video Enhance AI models and how they interact with AviSynth should be understood to apply solely to ... the DS9 Upscale Project

Much of it has to do with newbies that don't know what they don't know, and are gullible suckers to marketing.

You seem to have knowledge, curiosity, open mind, possibly try to use things in new unintended ways to get results. That I can respect. If you're blending Topaz, Avisynth, and others, I find that interesting, and worth a look. You're unusual. And you're like me.

Most people are too lazy to figure out how to install Avisynth (or even use Hybrid), and instead parrot "duh, Topaz best, yay!" -- complete with super-crappy examples of their botched work, by cranking sharpening to 11, and twiddling all the knobs and buttons in the dummy-friendly Topaz GUI. Also: "I paid $200, so it's the best ever," or "I got a discount on $200 software, so it's the best ever", or "I downloaded a cracked copy of $200 software, so it's the best ever!"

Perspective. You got it, others don't.

Topaz alone is worthless software. But even a broken clock is accurate twice per day. So I'm sure Topaz has some narrow niche uses, some sources, combined with higher quality methods. It's can't be 100% bad, but I can easily say it's not useful at least half the time or more (probably MUCH more).

[JoelHruska]

"Topaz alone is worthless software. But even a broken clock is accurate twice per day. So I'm sure Topaz has some narrow niche uses, some sources, combined with higher quality methods. It's can't be 100% bad, but I can easily say it's not useful at least half the time or more (probably MUCH more)."

It's definitely true that some content is more easily upscaled than others and the interplay between sharpening and noise in Topaz is non-deterministic and cannot be easily predicted. Processing footage in AviSynth with little to no sharpening can result in DS9 footage that shows very little improvement in Topaz. Injecting noise from some filters (QTGMC is the best) will improve sharpness in older versions of TVEAI, but noise (or grain) from other filters is less effective. Sometimes, artificially injecting noise into a scene helps the upscaler a lot. Sometimes, it helps the upscaler lock on to the wrong detail, resulting in a background that's almost as sharp as the foreground. Trial and error is the only way to determine the proper balance.

The improvement to DS9 Season 1 & Season 2 from upscaling is currently so marginal, I don't intend to do those episodes until I find a better way to handle them or upscaling improves. I do intend to process them through AviSynth. There appears to be very little content it cannot improve in one fashion or another.

[DruidCtba]

And QTGMC gives way better results than TVEAI.

And do you get a result equal to or better than the one presented by friend JoelHruska?

Show it there for everyone to see and prove what you're saying friend .

Att.

Druid®.

[DruidCtba]

Ok, you just read the title.

To my understanding, the key point was that the expert people you mentioned stated that you can obtain an equivalent/superior result with AviSynth filters. I agree with them.

And what I've seen so far was NOTHING, no original processed video showing that it improved or equaled some results obtained by VEAI, I just saw blah blah blah blah blah blah, but nothing at all.

What I see is that it is visually for common people (non-professionals with lynx eyes) and VEAI greatly improves the images in the tuning, generating details on top of what was lost neither avisynth or ffmpeg do because as it was said "it was lost" it cannot be recovered, but just INVENTED, and if this INVENTION improves the final result of the product, then it produces something viable to watch over an original version scaled to 4K by a TV and its processors.

In short: it's very blah-blah-blah-blah-blah-blah, but to present something that is even better than the VEAI I haven't seen here; I confess that yes, I saw some results with avisynth close to some VEAI results, but not better yet.

Yes I know, some VEAI models still have problems like ghosting on the DIONE TV/DV model, and blurred images at some points in a 480i video on the DIONE INTERLACED ROBUST V4 model, because I saw it myself, and I hope they discover the why this happens and get it right as soon as possible.

Att.

Druid®.

[DruidCtba]

The script I'm currently using:

TFM(pp=5, mode=2, micmatching=3).TDecimate(hybrid=1)
QTGMC(ShutterBlur=3,ShutterAngleSrc=180,ShutterAng leOut=180,SBlurLimit=8,Preset="Very Slow", TR2=3, InputType=1,SourceMatch=3, MatchEnhance=1.0, MatchPreset="Very Slow", TR0=2, TR1=2, MatchPreset2="Very Slow", sharpness=0.6, SMode=2, Rep0=22, Rep1=22, Rep2=22, RepChroma=True, Sbb=3, SubPel=4, NoiseProcess=1, ChromaNoise=True, GrainRestore=0.25, NoiseRestore=0.5,DenoiseMC=True, NoiseTR=2)
pSharpen(strength=65, threshold=90, ss_x=4.0, ss_y=4.0)
MAA2(mask=1, chroma=true,ss=4, aa=128, aac=128, threads=8, show=0)
LSFmod(defaults="slow", strength=60, Smode=5, Smethod=3, kernel=11, preblur="ON", secure=true, Szrp=16, Spwr=4, SdmpLo=4, SdmpHi=48, Lmode=4, overshoot=1, undershoot=1, Overshoot2=1, Undershoot2=1, soft=-2, soothe=true, keep=40, edgemode=2, edgemaskHQ=true, ss_x=4.0, ss_y=4.0, dest_x=704, dest_y=480, show=false, screenW=1280, screenH=1024)

EDIT: I do not know why the forum is inserting a space in the script above, in the "ShutterAngleOut=180" line. There is no space in my forum post in the editor. You'll have to remove it if you want to test. I've tried to remove it repeatedly, but even re-copying and pasting the script from the AviSynth file is not correcting the problem.

Friend JoelHruska what program are you using to insert this avs script from avisynth? Is the version you are using 2.60?

From what I saw you are using a very old version of VEAI which is now in version 2.3.0, am I right or wrong? If so why use a very old version?

Another thing that wasn't clear, or my inexperience condemns me to the error, you used avisynth to improve the video in the same original proportion, right, ie 480i/576i and only then did you Upscaling in VEAI?

Att.

Druid®

[lollo]

And what I've seen so far was NOTHING, no original processed video showing that it improved or equaled some results obtained by VEAI, I just saw blah blah blah blah blah blah, but nothing at all.

You did not pay attention, friend https://forum.videohelp.com/threads/401910-Topaz-products
Check the workflows, please!

Some example of videos before/after restoration done long time ago:

transition_mc_ufo_sII2a_spot_amtv_2_cut.mp4

compare_t2a_sigla_1a_dtbc_nr_on_best_on_copie_off_d3r_on_org_v16ad-muxed.mp4

compare_thor_fc05_thor1_amtv_3_cut_org_v17.mp4

On top of the restoration, I checked Avisynth upscale versus Topaz upscale. First was the winner (small comparison in the first link)

Also understand that my videos are s-vhs captures, so my "world" can be different, and YMMV!

[lollo]

I saw some results with avisynth close to some VEAI results, but not better yet.

IMHO you cannot compare the 2: AviSynth filters is a complete suite of filters doing almost everything, Topaz (in first instance) is "only" a resizer+sharpener, anc cannot compete with the first, especially on noisy sources requiring restoration.

[JoelHruska]

"what program are you using to insert this avs script from avisynth? Is the version you are using 2.60?"

I've used StaxRip 2.1.3.0 and 2.6.0. Both work fine. There's no reason a different front end like MeGUI shouldn't work.

"From what I saw you are using a very old version of VEAI which is now in version 2.3.0, am I right or wrong? If so why use a very old version?"

I use both. I have most of the TVEAI versions Topaz has released. I use Gaia-CG 1.5.3, sometimes with larger amounts of noise than I would use for a file I intended to watch. Large amounts of noise can help Gaia-CG 1.5.3 find detail and enhance it within the scene. Because Gaia-CG includes a denoise filter, the majority -- but not all -- of the added noise is removed by the application. This has the side effect of helping to fix the over-smoothed look that Gaia-CG can produce when used against live-action footage.

Noised Gaia-CG 1.5.3 and modern Artemis-HQ 2.3.0 are very similar when optimized, but Artemis-HQ is more trustworthy as a fire-and-forget solution. Artemis-HQ produces more consistent but less sharp output. Gaia-CG 1.5.3 produces sharper, smoother, but occasionally over-sharp / overnoised output. I recommended Artemis-HQ for the entire project because my goal was to give people a solution that minimized the chance of looking bad in any given scene rather than maximizing the chance of looking good and risking damage in the process. Different scenes can absorb different amounts of sharpening or look best with various amounts of noise. Gaia-CG 1.5.3 is ill-tempered and particular about the footage you feed it. Artemis-HQ is more agreeable.

I also use 2.3.0 Strong DeHalo situationally.

Another thing that wasn't clear, or my inexperience condemns me to the error, you used avisynth to improve the video in the same original proportion, right, ie 480i/576i and only then did you Upscaling in VEAI?

Correct. The footage is first processed in AviSynth. I do not change the resolution at this time. QTGMC performance drops a great deal as resolution approaches 1080p. It can still run against 720p footage but 1440x1080 or 1280x960 are much slower. I'd like to know more about upscaling in AviSynth, but when I used the AIUpscaler filter I linked above, I always ran it after QTGMC to avoid the performance hit from operating on higher resolution footage.

[JoelHruska]

"Some example of videos before/after restoration done long time ago:"

I'd like to know how you processed "compare_t2a_sigla_1a_dtbc_nr_on_best_on_copie_off _d3r_on_org_v16ad-muxed.mp4" and the Thor footage. Those are both excellent. (Not saying the first link isn't, but those two really jumped out at me).

The more damaged your footage is, the more important AviSynth often is before applying TVEAI. I have seen one instance in which a direct Topaz ingestion yielded superior output to anything I attempted in AviSynth (this was not DS9 or any TV show). I don't think TVEAI is particularly optimized for VHS-derived video right now. It's much easier to get good results from 640x480 than 320x240.I haven't experimented with VCR footage specifically, but given that I've seen the application struggle with 320x240 it makes sense that it would also have a tougher time with VCR output.

"Topaz (in first instance) is "only" a resizer+sharpener, anc cannot compete with the first, especially on noisy sources requiring restoration."

Topaz Strong DeHalo is pretty good, but you have to blend the output in most cases rather than using it on its own. TVEAI also includes at least a light noise processor as part of certain models, but only a few of them (Proteus, Theia) expose the setting as user-configurable. Gaia-HQ is the only model I'm aware of that seems to do no noise processing at all. Gaia-HQ amplifies noise in DS9 so I don't use it.

[lordsmurf]

VEAI greatly improves the images

No.
Topaz almost entirely just oversharpens.
You can do that with SD already, for decades now, HD upscale not needed.

[DruidCtba]

"what program are you using to insert this avs script from avisynth? Is the version you are using 2.60?"

I've used StaxRip 2.1.3.0 and 2.6.0. Both work fine. There's no reason a different front end like MeGUI shouldn't work.

"From what I saw you are using a very old version of VEAI which is now in version 2.3.0, am I right or wrong? If so why use a very old version?"

I use both. I have most of the TVEAI versions Topaz has released. I use Gaia-CG 1.5.3, sometimes with larger amounts of noise than I would use for a file I intended to watch. Large amounts of noise can help Gaia-CG 1.5.3 find detail and enhance it within the scene. Because Gaia-CG includes a denoise filter, the majority -- but not all -- of the added noise is removed by the application. This has the side effect of helping to fix the over-smoothed look that Gaia-CG can produce when used against live-action footage.

Noised Gaia-CG 1.5.3 and modern Artemis-HQ 2.3.0 are very similar when optimized, but Artemis-HQ is more trustworthy as a fire-and-forget solution. Artemis-HQ produces more consistent but less sharp output. Gaia-CG 1.5.3 produces sharper, smoother, but occasionally over-sharp / overnoised output. I recommended Artemis-HQ for the entire project because my goal was to give people a solution that minimized the chance of looking bad in any given scene rather than maximizing the chance of looking good and risking damage in the process. Different scenes can absorb different amounts of sharpening or look best with various amounts of noise. Gaia-CG 1.5.3 is ill-tempered and particular about the footage you feed it. Artemis-HQ is more agreeable.

I also use 2.3.0 Strong DeHalo situationally.

Another thing that wasn't clear, or my inexperience condemns me to the error, you used avisynth to improve the video in the same original proportion, right, ie 480i/576i and only then did you Upscaling in VEAI?

Correct. The footage is first processed in AviSynth. I do not change the resolution at this time. QTGMC performance drops a great deal as resolution approaches 1080p. It can still run against 720p footage but 1440x1080 or 1280x960 are much slower. I'd like to know more about upscaling in AviSynth, but when I used the AIUpscaler filter I linked above, I always ran it after QTGMC to avoid the performance hit from operating on higher resolution footage.

Friend JoelHruska thank you very much for your reply, you are a very consistent person and having more experience using both products as complements has been very important for this topic, because you bring here facts about how far one product can go and how far the other can improve the result of this, ie: before deinterlacing with QTGMC, after image treatment 480p/576p QTGM with AviSynth and finally sharpening the image, removing noise and scaling it to 720p/1080p/2160p.

But once I thank you for your posts here and your prompt answers, as they, through their experience, will serve as a basis for many here to also take their tests and improve their work, whether at an amateur and personal level as well as at a professional level for those who are starting to work in this area.

Att.

Druid®.

[lollo]

I'd like to know how you processed

Attached the avisynth scripts. They do not precisely refer to the posted comparisons (old experiments), because they changed a little bit since then.

They are very simple, and based on two filters developed by Didèe: TemporalDegrain and LimitedSharpenFaster.

The script for the Ufo spot deals with progressive (fields from the same moment in time) and interlaced sequences, so I lossless deinterlace the second before filtering and interlace back at the end.

The script for the Thunderbirds video adds some stabilization with LordSmurf StabMod and de-haloing with Didèe Dehalo_alpha, because the original broadcasted master has some shot with evident halo. Used with caution, because it is a destructive filter.

The script for Thor (commercial vhs) adds RemoveDirtSMC, as used by videofred.

All the filters are constantly under evolution, and more effective versions are discussed in doom9 forums (i.e. SpotLess replacing RemoveDirtSMC).

On the other side, I am not a big fan of deinterlacing and upscaling, so I am not sure you'll find any utility in the scripts.


For sure, I will follow your astonishing work on improving DS9, my PAL DVDs are also not good in quality, and the show deserves much much better

ufo_sII2a_spot_amtv_2_cut_v19.avs

t2a_sigla_1a_dtbc_nr_on_best_on_copie_off_d3r_on_v21x.avs

thor_fc05_thor1_amtv_3_cut_v22y.avs

[pm-s]

You can maybe get equivalent results with AVISYNTH, but you can get even better results with vapoursynth VSGAN (and it's 100% free and way more customisable)

Thank you Selur for making me aware of it's existence

[pm-s]

TVEAI bothers me because, after deinterlacing, adds no more sharpness, and a lot of artefacts to perfectly fine analogue source


Why do people even bother with upscaling crap.

[JoelHruska]

"Why do people even bother with upscaling crap."

Because nobody has shown me a way to produce this kind of output without it. Original Voyager credits on the left. Upscaled on the right.

EDIT:

My apologies on this. The original link that was here included an image that had been additionally post-processed, without the use of TVEAI. Since this thread is about TVEAI, I've removed that image and uploaded the appropriate one. This is original VOY DVD against blended output from Topaz -- Dione and Gaia-CG.

https://imgsli.com/NjI4Nzc

Anybody who knows how to turn the left-hand image into the right-hand image *without* the use of Topaz or a similar product is welcome to contact me or drop a note in this thread. Trust me, I want to hear from you.

[Selur]

Share the original frame and I can throw some filters at it and post the output.

[Cornucopia]

That slider makes it easier to compare, I'll give you that. However,

1. The right is NOT any more detailed. No "magic".
2. The right only has higher local contrast (sharpening) and contrast expansion in the dark areas. Easy to do in ANY decent processing app.

Scott

[JoelHruska]

"1. The right is NOT any more detailed. No "magic"."

AI is not magic. AI can have a remarkable impact on certain footage, but it is not magic. I avoid AI models that inject new detail into scenes. Such models have a tendency to guess wrong more than they guess right.

2. "The right only has higher local contrast (sharpening) and contrast expansion in the dark areas. Easy to do in ANY decent processing app."

So, the initial image was wrong. I apologize for that. It had been post-processed in an additional application. I have updated the comparison in the original post, but here's the same link so you don't have to go scrolling for it:

https://imgsli.com/NjI4Nzc

That level of clarity may not seem like a particular leap forward, but it's rather difficult to get. I did not have much luck sharpening this scene at all until I deliberately started dumping noise into the AviSynth output video to give Gaia-CG more data to find. I have not spent 1/10 as much time with Voyager as I have with DS9, so any idea for improving output is welcome.

@Selur: Original VOY credits demuxed here

https://1drv.ms/v/s!AphTLFRW13WMkC8tc95eDg74ju7Q?e=19wcEf

Additional Voyager images:

https://www.extremetech.com/wp-content/uploads/2021/07/000295.jpg
https://www.extremetech.com/wp-content/uploads/2021/07/000545.jpg
https://www.extremetech.com/wp-content/uploads/2021/07/000970.jpg
https://www.extremetech.com/wp-content/uploads/2021/07/003107.jpg
https://www.extremetech.com/wp-content/uploads/2021/07/003557.jpg

Before anyone snarks at the use of JPG: The site backend limits us to 1MB image files at the moment. Major headache. Trying to get it fixed. Best I could do for the story (and these are files I've shown publicly).

[JoelHruska]

I find it very strange that people are wedded to the idea of using these two utilities in an either/or fashion

In my flow I simply replace the "upscale section" of an Avisynth script (nnedi3) with the "upscale feature" from Topaz. All the rest (deinterlace, denoise, ...) is still in place.

The first is better in my case, but YMMV.

https://forum.videohelp.com/threads/401910-Topaz-products

In any case, thanks for sharing your good results and your script

edit: thanks added

Iollo, I'm testing your method of resizing against my typical run in AviSynth. I'll post some comparisons between OG footage, my base QTGMC output, output with your scaling method, and TVEAI.

[JoelHruska]

@Iollo,

Comparisons. So, to be clear: This is your resizing script ("nnedi3_rpow2(rfactor=2, nns=4, qual=2, cshift="Spline36Resize", fwidth=1440, fheight=1080)" appended to my own script above. I also color graded the output so the comparisons would be apples-to-apples in that regard.

https://imgsli.com/NjI5MDI

AviSynth-alone on the left, Topaz on the right. I think this output is quite good. TVEAI is cleaning up noise, but it doesn't offer an improvement in terms of detail.

https://imgsli.com/NjI5MDM

This frame shows what some genuine improvement from Topaz, but it isn't huge. Her com badge is somewhat sharper from TVEAI, though I could compensate for this by increasing sharpening in my own filter. Hit Zoom and take a look at her eyes with the slider, though -- I like how TVEAI handles them.

https://imgsli.com/NjI5MDQ

This is one shot where I think TVEAI is cleanly better.

Here's an OG, unmodified frame (except for color) against your script recommendation:

https://imgsli.com/NjI5MDU

There's definitely improvement. Her badge and eyes are both sharper.

I tested the idea of throwing your 1440x1080 output through Topaz just to see what happened. Nothing really happened. After resize, the footage isn't any better, but it's not really worse, either.

[lollo]

Excellent job JoelHruska!
It is quite late now in Italy, so I will study your results tomorrow, because they need to be analyzed with maximum care.

Thanks again for your contrbution to the "restoration world", where nothing is really established by default, ad everybody is learning...

[DruidCtba]

That slider makes it easier to compare, I'll give you that. However,

1. The right is NOT any more detailed. No "magic".
2. The right only has higher local contrast (sharpening) and contrast expansion in the dark areas. Easy to do in ANY decent processing app.

Scott

But is this exactly what enchants me in the results obtained by VEAI, which so far nobody has practically presented how to achieve the same effect with AVISYNTH plugins?

This is what the QUESTION: SHARPNESS!

I look for exactly this, BETTER SHARPNESS, in the Upscaling process, and if it weren't for the problems that some modules that VEAI presents, it would definitely be for me the perfect software for Upscaling my DVD of shows that didn't come out on Blu-ray and that I like so much .

But our friend JoelHruska brought us, through his experience, new data on how to use pre-processing with avisynth and then upscaling with VEAI, showing very satisfactory results from my point of view.

In short: he is killing the snake and showing his stick .

Att.

Druid®.

[lordsmurf]

and showing his stick .

Uhhh ... what?

[lollo]

https://imgsli.com/NjI5MDI

AviSynth-alone on the left, Topaz on the right. I think this output is quite good. TVEAI is cleaning up noise, but it doesn't offer an improvement in terms of detail.

I agree. However, I do not know if you used some noise reduction for the AviSynth part. If your TVEAI model uses some noise reduction, you should also have some on the right part for a better comparison.

https://imgsli.com/NjI5MDM

This frame shows what some genuine improvement from Topaz, but it isn't huge.

I assume here the flow is the same as before. Same considerations from my side.

https://imgsli.com/NjI5MDQ

This is one shot where I think TVEAI is cleanly better.

Agree.

Here's an OG, unmodified frame (except for color) against your script recommendation:

https://imgsli.com/NjI5MDU

There's definitely improvement. Her badge and eyes are both sharper.

Yes, it is. But the sharpening I use on my videos is very soft, because I do not want artifacts. On your material you can maybe push a lot the sharpening and get better results also with AviSynth.

My impression, is that on your material, my AviSynth filtering is not completely adequate, which is normal: every video, sometimes every scene, requires dedicated procedures...

A small contribution also on my side: I compared again the full "Avisynth flow (filtering + upscale)" with a "Avisynth (filtering) + TVEAI (upscale) flow" on one of my capture. Here the result:

https://imgsli.com/NjI5Njc

I did not play a lot with TVEAI options, so the comparison is not accurate (and to be honest I compare a HuffYUV compression versus a x264 compression to save time), but I do not see a major difference, so I still prefere the free full Avisynth flow for my videos. For your work I think you should continue with your existing flow, it seems more appropriate and easier to implement.


[Attachment 60083 - Click to enlarge]

Thanks a lot for the pleasant and interesting discussion. Until next time!