Scaling and anti-aliasing

So, I'm finishing a series that was shot all 4k, but my timelines are HD for now. However, we may need to re-render in UHD, so I'm doing all the (very simple) title graphics in Photoshop - at UHD frame sizes. When I drop them into Resolve, they scale correctly, and I can cut them right in - but I'm getting a too much aliasing on diagonal edges. I switched the sequence to UHD, and looked at the composites- no aliasing, looks great.

Is there a trick for down scaling without aliasing? In my Project Settings the image scaling is set to smoother - but changing that setting makes no apparent difference at all, regardless.

As a test, I scaled a couple graphics in Photoshop, and cut them in as HD. No apparent aliasing. BUT, don't really want to make two versions of all the graphics if I can avoid it.

Suggestions?

Hmmm- smoother setting should work in this case. Maybe something is broken.
Try creating two UHD versions- normal and another where you add small gaussian blur in Photoshop (0.5 for start?). This should behave better after downscaling.

Another question is how do you monitor HD signal? If you do it with BM card on TV over interlaced signal (eg 50i) this will be the reason.

Thanks for the ideas.

Monitoring is Decklink HD 23.98 psf

Tried switching to 23.98 p (true p) with no change in appearance. The aliasing is visible in the viewer, too, full screen on the computer monitor, 1:1 or any size.

Tried the soft-edge characters, also no significant improvement, and they're a touch, well, soft! This seems to be an issue in the scaler - just not finding a way to solve he problem, short of scaling the text in AE and making two versions of it, or living with the aliasing.

Also, I DON'T see the problem if I generate the text in Resolve, but Resolve (14) doesn't have enough control over aspect changes, mixing fonts, etc. (Actually, I don't know how Resolve generated text would scale, either, since I'm not making the Resolve text in 4k then changing the timeline settings. I assume the text is vector, but that's just an assumption.)

Any additional tips/suggestions appreciated!

…
Any additional tips/suggestions appreciated!

Strange question but... what color science do you use? RCM or ACES?
I had similar problem with ACES.

[Rohit Gupta]

Thanks for the ideas.

Monitoring is Decklink HD 23.98 psf

Tried switching to 23.98 p (true p) with no change in appearance. The aliasing is visible in the viewer, too, full screen on the computer monitor, 1:1 or any size.

Tried the soft-edge characters, also no significant improvement, and they're a touch, well, soft! This seems to be an issue in the scaler - just not finding a way to solve he problem, short of scaling the text in AE and making two versions of it, or living with the aliasing.

Also, I DON'T see the problem if I generate the text in Resolve, but Resolve (14) doesn't have enough control over aspect changes, mixing fonts, etc. (Actually, I don't know how Resolve generated text would scale, either, since I'm not making the Resolve text in 4k then changing the timeline settings. I assume the text is vector, but that's just an assumption.)

Any additional tips/suggestions appreciated!

Are you using 4:4:4 or 4:2:2 SDI monitoring?

Color Science - Resolve

Monitoring 4:2:2, SDI via older decklink extreme to Sony PVM OLED

Yes, seems like scaling is just doing way to much sharpening, so you get aliasing. Sounds like a bug.

.../...
As a test, I scaled a couple graphics in Photoshop, and cut them in as HD. No apparent aliasing. BUT, don't really want to make two versions of all the graphics if I can avoid it.

Suggestions?

Can you give a download link to your PSD? It's very interesting.

Yes, seems like scaling is just doing way to much sharpening, so you get aliasing. Sounds like a bug.

This is an issue that has been around since forever. Antialiasing in Resolve is kind of broken, whatever filters they're using are apparently garbage. You can even see this with the built-in text tool (look at the curved and diagonal edges):



This is especially true with alpha channels, which break rather badly in many formats:



White logo on white BG, scaled down 50%, viewed at 100% zoom -- see the terrible edge aliasing where the alpha is being filtered incorrectly, creating a nasty halo?



Same white logo on white BG, NOT scaled (and therefor not filtered), viewed at 100% zoom -- nothing to look at here, because the alpha isn't being filtered and so multiplying it creates no halo.

It's a pretty show-stopping bug, and is the *one* thing left in Resolve that really irks me and keeps me from being able to get away from Adobe products entirely. The fact that it's persisted so long, and isn't fixed even in DR15, kills me. There is no other video app I know of that has such terrible filtering issues.

Yes, it can be related to whole alpha issue. I though it's different problem, but it may not be.

Sadly, I think Jed Mitchell has the answer - alpha scaling.

These graphics are VERY simple - white text writ large, almost full screen. So, if I make UHD-sized white on black cards in PS, no alpha, import them in, and swap that in, then change the transfer mode to say, add, the diagonal edges of the auto-scaled graphic are clean - no obvious aliasing. (bonus, I get more interesting interaction between the background and the title card this way, too.)

While this sucks, since I'll have to go back over 15 episodes and make these changes, it's at least a solution for now. Certainly wouldn't solve the problem if the graphic were anything else.

Fix that alpha problem please!!

Has anyone tried forcing linearized timeline to see if problem is also visible there? It sounds like filtering non-linear premultiplied RGB values produces wrong RGB results in relation to alpha and this creeps out as halo/aliasing. It could probably be cured by unpremulting the RGB before transform operation and premulting again after.

EDIT: Dark halo in the white-on-white image suggests double premultiplication. Gamma-corrected RGB values interpolated as is shouldn't produce darker result... If a gamma 2.2 timeline pixel with 1.0 value and 1.0 alpha is interpolated to 0.5 RGB and 0.5 alpha it still produces correct unpremulted value. BG*(1-alpha) also produces 0.5 and combined over will be 1.0 as expected. So something additional is done that skews RGB values. It could be explored by feeding a linear ramp and probing the resulting values.

Run into the same thing a while ago. Can't remember which one it was (maybe TIFF), but using a different file format for text with alpha, fixed the problem with alpha bleeding.

We have run into so much trouble with applying graphics inside resolve that it has eventually become policy now that nothing is ever to be finished out of resolve. It's sad and it sucks, but the fact remains that we have not been successful in finding any foolproof way to apply supers to a graded spot using the multilayer timeline inside Resolve.

My most recent experience with this involved set type from the client that originally came in as PNG's with embedded alpha. I was grading in HD, but delivering both HD and UHD. My sources were some type of XAVC Slog3 as well as a few drone shots in H.264, all of which was UHD 3840x2160. The PNG's were different sizes, all of which large enough for UHD output, all of which required scaling and positioning.

Adding those PNG's with embedded alpha looked absolutely horrendous.

Converting those PNG's to TIFF's might have made a slight improvement. I cant remember exactly, but it still looked horrendous.

Converting these graphics from the original PNG's with embedded alpha to ProRes4444's with embedded alphas using After Effects at first appeared to fix the shoddy edges, but upon closer inspection, they were still fairly ugly.

In the end, after wasting a couple hours trying to troubleshoot this again for the Nth time, I output graded flats and sent them to Flame where we could apply the graphics and output the ProResHQ masters without issue.

This has been a problem for us in Resolve for years.

[Peter Chamberlain]

Have you had a chance to review v15b5 as there was some work there related to this?
In conjunction with the above, see if setting the resize filter to smoother gets what you are looking for.

I think there is more to it than just a substandard job on alpha scaling. A little while ago I was comparing Resolve scaling options to Fusion scaling filter selections using a zone plate without alpha. Resolve's scaling was worse than Fusion's.

To avoid aliasing I've been scaling all animated out-of-house thin lined graphics in Fusion and then importing to Resolve.

Discussed before - we can't use resolve for finishing for this very reason. Doesn't pass QC
viewtopic.php?f=21&t=62426

We go to either premier or nuke studio for finals. A lot of studios go to flame. It's such a simple thing too. If iMovie can do it...

I've not tried the beta yet, but I really hope this is fixed. Also simple things like terrible aliasing on non-blurred bezier shapes in the colour tab.

It's a shame.

[Peter Chamberlain]

Once again, have you reviewed version 15b5?

Just one remark: colorists normally blur shapes in that tab!
It's not really meant for anything like rotoscoping, rather use Fusion for such tasks.

So I tried the 15v5 beta today and I'm happy to say there's definitely improvement. I took a simple white text png. with straight alpha into a clean timeline and rendered out of resolve to compare with nuke. Scaling set to 'smoother' and anti-aliasing set to 'on'.

All examples are left side resolve, right side nuke. In a lot of cases I actually now prefer the resolve scale - but at 5% zoom it breaks down. Important to note that I'm comparing renders (prores HQ) not the viewer.

100%


60%


20%


5%

If Resolve used meaningful names (what does Smoother mean?) for resizing filters, it would be possible to choose the same in Nuke. May come handy in some cases to use identical one.

Which filter did you use in Nuke by the way?

If Resolve used meaningful names (what does Smoother mean?) for resizing filters, it would be possible to choose the same in Nuke. May come handy in some cases to use identical one.

Which filter did you use in Nuke by the way?

To keep comparison 'fair' I used nuke's default 'cubic' filter... but normally I prefer something a little smoother like mitchel.

I don't know if this will work in resolve, but I've been ding it in Photoshop for ages.

When using a zoom level, in the viewer, always use halves of 100% (so, 50%, 25%, 12,5%, and so on).

It fixes aliasing problems in photoshop, and since so, I use it also on DVR.
s