|
|
All the news now and in your digital future
Subtitle 10 Bit Uprez
Posted by hanguolaohu
|
Re: Subtitle 10 Bit Uprez August 23, 2008 08:45PM |
> Dang, I thought I could post pics within the message body.
Yeah, you can.
> if I nest the DV sequence including titles and then place it into a 10 bit sequence, does it
> automatically uprez the titles to 10 bit?
No, you'll need to do the titles in the 10-bit sequence. But if these are FCP text objects, then all you have to do is copy-paste.
> Also, should I render the 10 bit sequence before exporting the Quicktime?
Yes. The export process renders the sequence, but it's always better to do the render manually first.
www.derekmok.com
Yeah, you can.
> if I nest the DV sequence including titles and then place it into a 10 bit sequence, does it
> automatically uprez the titles to 10 bit?
No, you'll need to do the titles in the 10-bit sequence. But if these are FCP text objects, then all you have to do is copy-paste.
> Also, should I render the 10 bit sequence before exporting the Quicktime?
Yes. The export process renders the sequence, but it's always better to do the render manually first.
www.derekmok.com
|
Re: Subtitle 10 Bit Uprez August 24, 2008 08:37AM |
What you're seeing in the titles isn't the result of an 8-bit versus 10-bit thing. It's plain old 4:1:1 color subsampling in DV. You can improve (but not totally eliminate) that by going to 4:2:2. Heck, you can go to 4:4:4 if you really want, by using the Animation compressor, but there's little point, since it'll get smudged back down to 4:2:2 when you go out to any tape format but HDCAM SR.
The 8-bit versus 10-bit thing can be really confusing. It took me a long time to wrap my head around it when I was just starting out. Here's what made it click for me: Imagine you've got a ruler. It's graduated in inches: one, two, three. You measure something with it ? it's probably best not to use the thing boys most stereotypically measure for this example, but whatever works for you. Because it's only got inch marks, you have to round off. The thing you're measuring is three inches.
Then you go measure a bunch of other things, and you plot their lengths on a graph. Because your ruler is only marked off in whole inches, your graph looks like a staircase: one, two, three, four inches or whatever.
But then you go get a different ruler, one graduated in half-inches. When you re-measure all those things, your graph looks smoother, less like a staircase. Some of the things you measured that fell into the 3- and 4-inch columns are now in the 2?, 3, 3?, 4 and 4? columns. Your ruler is more precise, so your measurements can be more precise.
That's how 10-bit works versus 8-bit. You have more digits of precision ? like a more finely graduated ruler ? so you can measure finer distinctions between similar color values. Your "graph" looks less "staircasey," meaning your image has less banding in areas of subtle gradation.
But here's the thing: When you shoot in an 8-bit format ? and a ton of high-quality, professional formats are only 8-bit ? you're done. You've gone out and measured everything, and you've plotted a graph that's somewhat staircasey. If you convert that footage to 10-bit, you're not re-measuring any of the original things; you're not resampling the image in any way. You're just re-measuring the graph you drew the first time using a more finely graduated ruler. So in theory you have more precision, but in fact you're just rounding off all those extra digits to zero.
Now, this matters when you do something like apply a color correction, or even do a fade-to-black. More precision in your math means you get better results. But if you're not doing actual mathematics on your image data ? like you're just superimposing titles over the top with a simple 1-bit or 8-bit alpha matte ? then having that extra precision won't help. It won't hurt, but it won't help either.
There are some situations in which this explanation breaks down. For example, if you shoot on film, you capture a hell of a lot more color and brightness information than an 8-bit scan of the film can hold. So you can go back to the original negative and scan it with a different level of precision ? 10-bit, 12-bit, whatever ? and get more luma and chroma precision. The same is true if you shoot on Red in 12-bit linear. You can convert the raw files you get off the camera into video in any one of several ways, coming out as 8-bit or 10-bit lin or log or 12-bit lin or whatever. In those cases, you really can go back and get more precision out of your images, because you captured more precision in the actual camera itself.
But if what you record is an 8-bit format, then 8 bits is all there will ever be, unless you apply some kind of an effect and use 10-bit math to render it.
Wow. That was really long-winded. Sorry.
The 8-bit versus 10-bit thing can be really confusing. It took me a long time to wrap my head around it when I was just starting out. Here's what made it click for me: Imagine you've got a ruler. It's graduated in inches: one, two, three. You measure something with it ? it's probably best not to use the thing boys most stereotypically measure for this example, but whatever works for you. Because it's only got inch marks, you have to round off. The thing you're measuring is three inches.
Then you go measure a bunch of other things, and you plot their lengths on a graph. Because your ruler is only marked off in whole inches, your graph looks like a staircase: one, two, three, four inches or whatever.
But then you go get a different ruler, one graduated in half-inches. When you re-measure all those things, your graph looks smoother, less like a staircase. Some of the things you measured that fell into the 3- and 4-inch columns are now in the 2?, 3, 3?, 4 and 4? columns. Your ruler is more precise, so your measurements can be more precise.
That's how 10-bit works versus 8-bit. You have more digits of precision ? like a more finely graduated ruler ? so you can measure finer distinctions between similar color values. Your "graph" looks less "staircasey," meaning your image has less banding in areas of subtle gradation.
But here's the thing: When you shoot in an 8-bit format ? and a ton of high-quality, professional formats are only 8-bit ? you're done. You've gone out and measured everything, and you've plotted a graph that's somewhat staircasey. If you convert that footage to 10-bit, you're not re-measuring any of the original things; you're not resampling the image in any way. You're just re-measuring the graph you drew the first time using a more finely graduated ruler. So in theory you have more precision, but in fact you're just rounding off all those extra digits to zero.
Now, this matters when you do something like apply a color correction, or even do a fade-to-black. More precision in your math means you get better results. But if you're not doing actual mathematics on your image data ? like you're just superimposing titles over the top with a simple 1-bit or 8-bit alpha matte ? then having that extra precision won't help. It won't hurt, but it won't help either.
There are some situations in which this explanation breaks down. For example, if you shoot on film, you capture a hell of a lot more color and brightness information than an 8-bit scan of the film can hold. So you can go back to the original negative and scan it with a different level of precision ? 10-bit, 12-bit, whatever ? and get more luma and chroma precision. The same is true if you shoot on Red in 12-bit linear. You can convert the raw files you get off the camera into video in any one of several ways, coming out as 8-bit or 10-bit lin or log or 12-bit lin or whatever. In those cases, you really can go back and get more precision out of your images, because you captured more precision in the actual camera itself.
But if what you record is an 8-bit format, then 8 bits is all there will ever be, unless you apply some kind of an effect and use 10-bit math to render it.
Wow. That was really long-winded. Sorry.
|
Re: Subtitle 10 Bit Uprez August 25, 2008 02:00AM |
OK, I've done some tests. My theory is that I made 2 mistakes:
A. I made a nest of my DV sequence including titles and dropped that on a 10 bit sequence.
B. I exported a Quicktime movie without rendering that DV sequence nest.
So I believe my workflow solution is:
1. Nest only the picture of my DV sequence.
2. Drop that nest onto a 10 bit sequence.
3. Copy and paste FCP text titles from the DV sequence to track V2 of the 10 bit sequence.
4. Render
5. Export to Quicktime movie.
Someone on Creative Cow recommended "Render all YUV material in High Precision". I tried that and could not tell the difference with my workflow solution above. But despite these improvements, I still notice a bit of white in my titles:
Is this the best quality I can squeeze out of 10 bit titles?
Thanks for info.
A. I made a nest of my DV sequence including titles and dropped that on a 10 bit sequence.
B. I exported a Quicktime movie without rendering that DV sequence nest.
So I believe my workflow solution is:
1. Nest only the picture of my DV sequence.
2. Drop that nest onto a 10 bit sequence.
3. Copy and paste FCP text titles from the DV sequence to track V2 of the 10 bit sequence.
4. Render
5. Export to Quicktime movie.
Someone on Creative Cow recommended "Render all YUV material in High Precision". I tried that and could not tell the difference with my workflow solution above. But despite these improvements, I still notice a bit of white in my titles:
Is this the best quality I can squeeze out of 10 bit titles?
Thanks for info.
|
Re: Subtitle 10 Bit Uprez August 25, 2008 12:42PM |
Hey Alex,
I agree with Derek's suggestion...you need to recreate the titles in a 10-Bit sequence...not uprez the titles along with the footage. Uprezzed titles will never look as clean as they would when created in a high rez timeline. I wouldn't nest the footage in a higher rez timeline either. I would use a professional uprezzing tool like Red Giant's Instant HD 1.1 for only $99.99 (I just used this on a spot I am working on - worked great):
[www.redgiantsoftware.com]
When life gives you dilemmas...make dilemmanade.
I agree with Derek's suggestion...you need to recreate the titles in a 10-Bit sequence...not uprez the titles along with the footage. Uprezzed titles will never look as clean as they would when created in a high rez timeline. I wouldn't nest the footage in a higher rez timeline either. I would use a professional uprezzing tool like Red Giant's Instant HD 1.1 for only $99.99 (I just used this on a spot I am working on - worked great):
[www.redgiantsoftware.com]
When life gives you dilemmas...make dilemmanade.
|
Re: Subtitle 10 Bit Uprez August 25, 2008 03:28PM |
Haha... A wide screen discussion! 
Great explanation, Jeff. Lemme see if i can illustrate bit depth/rounding errors. Here's a simplified explanation. In the illustration, vertically, it measures bit depth, and horizontally, it indicates an audio sample of a sound wave, a "1" indicates audio on, a "0" indicates silence:
2 bit audio:
01100
11110
4 bit audio:
00100
01100
11110
11111
Now when you plot out the waveform based on the information given, notice how much more precise it gets when you have more bits working for you (a more defined curve on the 4 bit example as compared to the 2 bit example where the tip of the curve is actually flat). This is also otherwise known as rounding errors (where an actual analog point falls somewhere between 2 points, so it is rounded off to the nearest available point of calculation). In video, gradients look more natural with higher bitdepth, and this is also usually the result of banding- insufficient points of calculation.
In video, 8 bit depth, you have 2 (to the power of 8) which results in 256 points, and in 10 bits, you have 2 (to the power of 10) which results in 1024 points of measurement. It takes longer to render because you now have so much more numbers to calculate.
When you render, footage is decompressed and chroma subsampling is promoted to 4:4:4 YUV (and 10 bit if 10 bit rendering is selected). As the footage passes through each filter, the points are recalculated and as a result, rounding errors occur. Rounding errors are minimized when a higher bit depth is used. In floating point math (when you select high precision YUV rendering) it also calculates the decimal points (in between points of measurement), allowing a more accurate representation of the curve as you go between filters. After the footage is rendered, it is then recompressed back to the sequence codec and in your initial case, DV. DV is an 8 bit 4:1:1 codec, so the chroma is subsampled back to 4:1:1 and the 10 bit uncompressed data is then down filtered to 8 bits. This results in rounding errors, naturally, however, it is preferred, as the footage suffers less rounding errors as it went through the filter stack.
So yes, with footage that require a render, a generation is lost, as dv goes through recompression, which is why some of us render out to 10 bit codecs. It also helps prevent banding.
However, what is more obvious, especially in this case, are the graphics and text which is both goes through DV compression as well as the chroma subsampling in DV. 4:2:2 is inferior in terms of chroma subsampling when you compare it to Animation's 4:4:4, where all the chroma channels are sampled. According to 601 specs, half the chroma values are sampled in relation to the luma channel for video formats. Chroma subsampling was created to save on bandwidth because humans are less sensitive to color difference than light. In video, it barely makes a difference, but in graphics and text it is slightly more noticeable (especially when you're using an inferior chroma subsampling ratio like DV's 4:1:1). Video formats tend to hate sharp chroma edges mainly because of this, also why colored text should be slightly bigger (it completely missed the "L" in "also" in the original picture)
On another note, nesting is supposed to bypass codec compression.
However, you can try out Derek's suggestion to see if it helps.
Btw, are you uprezzing to HD? If you're uprezzing, using a more dedicated software can help eliminate problems... I'm assuming you're outputting SD for SD. Hmm... Haven't actually tried this, but what happens if you create a sample subtitle in HD and down rez it to SD. I'm wondering if oversampling could help eliminate this issue. You MAY actually eliminate it (though you also could run into downscaling issues like flickering, so you may need to add a very slight channel or gaussian blur of around 0.5 to eliminate that). Try it out on one title to check.
www.strypesinpost.com
Great explanation, Jeff. Lemme see if i can illustrate bit depth/rounding errors. Here's a simplified explanation. In the illustration, vertically, it measures bit depth, and horizontally, it indicates an audio sample of a sound wave, a "1" indicates audio on, a "0" indicates silence:
2 bit audio:
01100
11110
4 bit audio:
00100
01100
11110
11111
Now when you plot out the waveform based on the information given, notice how much more precise it gets when you have more bits working for you (a more defined curve on the 4 bit example as compared to the 2 bit example where the tip of the curve is actually flat). This is also otherwise known as rounding errors (where an actual analog point falls somewhere between 2 points, so it is rounded off to the nearest available point of calculation). In video, gradients look more natural with higher bitdepth, and this is also usually the result of banding- insufficient points of calculation.
In video, 8 bit depth, you have 2 (to the power of 8) which results in 256 points, and in 10 bits, you have 2 (to the power of 10) which results in 1024 points of measurement. It takes longer to render because you now have so much more numbers to calculate.
When you render, footage is decompressed and chroma subsampling is promoted to 4:4:4 YUV (and 10 bit if 10 bit rendering is selected). As the footage passes through each filter, the points are recalculated and as a result, rounding errors occur. Rounding errors are minimized when a higher bit depth is used. In floating point math (when you select high precision YUV rendering) it also calculates the decimal points (in between points of measurement), allowing a more accurate representation of the curve as you go between filters. After the footage is rendered, it is then recompressed back to the sequence codec and in your initial case, DV. DV is an 8 bit 4:1:1 codec, so the chroma is subsampled back to 4:1:1 and the 10 bit uncompressed data is then down filtered to 8 bits. This results in rounding errors, naturally, however, it is preferred, as the footage suffers less rounding errors as it went through the filter stack.
So yes, with footage that require a render, a generation is lost, as dv goes through recompression, which is why some of us render out to 10 bit codecs. It also helps prevent banding.
However, what is more obvious, especially in this case, are the graphics and text which is both goes through DV compression as well as the chroma subsampling in DV. 4:2:2 is inferior in terms of chroma subsampling when you compare it to Animation's 4:4:4, where all the chroma channels are sampled. According to 601 specs, half the chroma values are sampled in relation to the luma channel for video formats. Chroma subsampling was created to save on bandwidth because humans are less sensitive to color difference than light. In video, it barely makes a difference, but in graphics and text it is slightly more noticeable (especially when you're using an inferior chroma subsampling ratio like DV's 4:1:1). Video formats tend to hate sharp chroma edges mainly because of this, also why colored text should be slightly bigger (it completely missed the "L" in "also" in the original picture)
On another note, nesting is supposed to bypass codec compression.
However, you can try out Derek's suggestion to see if it helps.
Btw, are you uprezzing to HD? If you're uprezzing, using a more dedicated software can help eliminate problems... I'm assuming you're outputting SD for SD. Hmm... Haven't actually tried this, but what happens if you create a sample subtitle in HD and down rez it to SD. I'm wondering if oversampling could help eliminate this issue. You MAY actually eliminate it (though you also could run into downscaling issues like flickering, so you may need to add a very slight channel or gaussian blur of around 0.5 to eliminate that). Try it out on one title to check.
www.strypesinpost.com
|
Re: Subtitle 10 Bit Uprez August 25, 2008 04:29PM |
Alex,
I am confused.
You are at DV now (720 x 480) and you are uprezzing a "film" to DigiBeta (720 x 486)? Not much of an upres for a film IMHO. Is this for theatrical release? TV? DVD? What kind of "hassle" could it be to uprez to at least 720p?
Anyway...the Instant HD is not for your titles...it's for your footage. It is a cleaner way to uprez. As Derek & I mentioned earlier you CANNOT uprez your titles... you need to re-render them fresh in the hi rez timeline (copy them from your DV timeline & paste them into the 10-Bit timeline & re-render).
Sometimes a hassle opens your eyes to the way things should be and that hassle ends up being a blessing in disguise.
When life gives you dilemmas...make dilemmanade.
I am confused.
You are at DV now (720 x 480) and you are uprezzing a "film" to DigiBeta (720 x 486)? Not much of an upres for a film IMHO. Is this for theatrical release? TV? DVD? What kind of "hassle" could it be to uprez to at least 720p?
Anyway...the Instant HD is not for your titles...it's for your footage. It is a cleaner way to uprez. As Derek & I mentioned earlier you CANNOT uprez your titles... you need to re-render them fresh in the hi rez timeline (copy them from your DV timeline & paste them into the 10-Bit timeline & re-render).
Sometimes a hassle opens your eyes to the way things should be and that hassle ends up being a blessing in disguise.
When life gives you dilemmas...make dilemmanade.
|
Re: Subtitle 10 Bit Uprez August 25, 2008 04:37PM |
> but I don't want to redo everything in Boris 3D manually.
Ahh... then nvm the HD titles.. Boris 3D is a title generator, and doesn't care whether it's HD or SD. I thought the titles came from AE. The default text tool in FCP sucks- I only use it for convenience and offlines. Photoshop was supposed to be FCP's great text tool (according to Ken Stone).
Does copying the text onto a 10 bit sequence (bypassing the nest) help?
www.strypesinpost.com
Ahh... then nvm the HD titles.. Boris 3D is a title generator, and doesn't care whether it's HD or SD. I thought the titles came from AE. The default text tool in FCP sucks- I only use it for convenience and offlines. Photoshop was supposed to be FCP's great text tool (according to Ken Stone).
Does copying the text onto a 10 bit sequence (bypassing the nest) help?
www.strypesinpost.com
|
Re: Subtitle 10 Bit Uprez August 25, 2008 10:10PM |
i didnt do it, nobody saw me, you cant prove a thing.
glad to be able help
glad to be able help
Sorry, only registered users may post in this forum.
