ESB Grindhouse is on: myspleen, piratebay, a.b.starwars so you can take your pick. I think ROTJ is on myspleen & a.b.starwars, I just did a search and it wasn’t on tbp.
Are you able to send invites to myspleen? I have no idea how I can get one.
The same way everyone else does. Send TV’s Frink a PM.
-G
If you read the text, the drizzling is only on the final images, I did that as an example of what drizzling was (or to use the proper term “Variable Pixel Linear Reconstruction”), as I thought many had probably not heard of it, and this was a discussion about image processing.
All the previous images are without the upscale/drizzle applied. The whole thing was an example of a technique, as I said, we don’t have enough full prints to do it to this level, you would need 20 or more prints to get this level, and not all scenes would work as well as the alignment gets tricky etc. and you can lose detail in the alignment process, so diminishing returns kick in.
It certainly isn’t what I am doing with the movies, I don’t have enough prints, and don’t have enough computational grunt to do the alignment and stacking, that single image used 63GB of RAM to process. Drizzling isn’t likely to be of much use for film captures, but I mentioned it as part of the discussion and applied it to see the results.Normally, I wouldn’t have enough prints to do this, I don’t have 30 prints, but what I do have is a reel of dailies of this scene, I also have a reel of dailies from Empire, if I get time I will scan it and show it as well, if it is of interest to people here. The dailies reels have the same scene looped a dozen of more times, which is very handy for this kind of thing. I picked a scene with a static camera as (a) I already had it scanned, and (b) a scene with movement has motion blur which wouldn’t show the effect as clearly.
I used 3 adjacent frames from each dailies repeat, and an two IB prints and an Eastman, they were all colour matched and roughly aligned. I had all of that handy from a prior experiment, so could get an explanation up quickly.Film does work well with this technique, due to the randomness of the film grain, the film weave means each print is slightly differently positioned, and the photons hit different areas of the grain etc. so if you have enough, you can get back much closer to only having neg grain involved. The film is, in my opinion, under sampled when scanned, and multiple prints and scans lets us recover a lot, and remove a lot of noise.
I’m not trying to peddle hope or anything else here, just joining a discussion of super-resolution and frame stacking with a freely admitted arbitrary example. If anyone has the misconception that we are doing this to the entire film, then it is just that, a misconception.
Where this does come in handy is helping to discover the correct colours without incorporating artefacts such as the yellowing blobs and grain structure, which gives a good colour target when grading, and working out what is dirt/micro-scratches and what is not, which is not always obvious, like those dark spots on the sand, they could easily be mistaken for dirt spots and incorrectly removed.
It is also handy for efforts like Harmy’s and Adywan’s, where they might want to recreate a background plate in far more detail than any single frame allows, like the sand crawler coming over the sand dune, the starry background and planets from the crawls, and many other shots, where they can create a much cleaner version and use that as a basis for the reconstruction.Having a clean plate we also find really useful to get a more accurate grain analysis to use for de-graining algorithms.
Anyway, it is intended only as a discussion of image processing technique, not as anything specific that I am doing re the restoration of any particular film.
Nice. Thanks for the explanation.
-G
sorry, hit quote instead of edit… again.
Poita,
I feel you aren’t being completely transparent here. I doubt it’s a coincidence that this is probably the most static shot in the entire film (stays the same for many frames). I doubt you would get those kind of SNR gains with merely averaging multiple captures, but I suspect you also averaged several frames of each capture as well before and after the 1 frame you show.
You cannot sustain this kind of NR throughout the film without some serious DVNR side effects due to tearing and motion compensation errors.
Also, the same frame from multiple captures will have a lot of similar noise from the previous generations. Average a bunch of different captures and you will still be left with the noise of the previous generations. Adjacent frames however will not have the same noise and you will get greater reduction than you would if just using multiple captures.
Don’t get me wrong, this shot looks great, but show the same gains from a frame a few seconds later during the “look sir, droids” pop-up and I’d say you really are magic.
I feel you are feeding a misunderstanding of sampling theory. The reason that Superresolution works on those ccd images is because they have no anti-alias filter before capture. Some of that high-frequency information is captured in the lower frequencies due to folding around the Nyquist. Then, as the camera (or object) moves, that detail can be recovered through upsampling, motion compensating, and averaging (or medianing). I don’t think film however is aliased in the same manner (but I’ve been wrong before).
On top of this, you are drizzling, which is just a fancy name for what reduces to a spatial-only linear filter. This is just a convolution, ie. sharpener. Your comparison should not be to the original denoised frame, but to the exact same denoised image that has been nicely upsized (lanczos or similar).
I love your work, I donated to the cause, but I think you’re giving false hope. I hope you prove me wrong.
love,
-Gary
double post
What you are describing is called noise reduction.
I’d like to be invited as well, if Dark Jedi lets you.
-G
Sorry I didn’t respond sooner. Those shots have an interesting teal tone similar to some other prints I’ve seen from the 80’s. I don’t have the Super8 print as a reference, so it’s nice to see it.
A mini update: Work continues apace with finalizing the first three reels. Most of reel 3 is finalized, and I’ve just signed off on the opening scenes:
This is 1000% the correct colors for this shot.
With ep7 coming and going from the public eye, I had the idea of reediting the numerous video reviews on it and putting them into plot order. This is mostly inspired by None and his “Thee Backslacpkping” video project, especially with recontextualizing disposable video products.
I have one question though, did someone make a music only track for ep7 yet?
I would watch that.
-G
All these are pretty good, but Neverar’s last attempt is stunning.
-G
I think it was good, but I only have my memories of it.
The GOUT subs has the crawl in all the languages so I bet it could be synched into these as well as the Vader and Emperor talk in ESB and removing the abominable “Wesa free!” at the end of ROTJ if it’s in the subs. If there are any other differences I forgot please inform us. 😃
Jesus Christ! Please tell me you’re joking and that’s not really in Jedi…
No, not in the '97 SE. The line was not added until 2004.
-G
This happens every time someone extracts the ISO. Why extract it? Why do anything to it?
In general, I agree Chewie. But, my favorite method of viewing downloads lately is to put it on a USB drive and plug into my Blu-Ray player. My Blu-Ray player is not smart enough to know what to do with an iso. Just sayin’.
-G
Congrats on re-forming team!
-G
DrDre - you sir are a master. That is it. Perfection.
Thanks, I like them a lot also!
Good idea for another source to use! but… Some look a bit blue, some too bright, some with a lack of contrast. Not loving them I gotta say.
Do the same for the JSC, and the LPP, and the IB, and this one, throw out the outliers, average the results, and then you may have something.
-G
Star Wars Sample: https://we.tl/VEyCxemjLh
Empire Sample: https://we.tl/verwNofxYa
Jedi Sample: https://we.tl/YjCgGTEZUG
Holy shit!
Except for the noise reduction (pretty bad on the SW clip for example on the wall behind Luke at 0:16). The colors and the resolution look nice though.
two things doc…
This is worthy of a Nobel Prize. Not joking, millions could benefit from this technology.
I think you would end up removing any intentional color timing from the film. Think about it. Take a photo, adjust the colors without clipping any of them too badly. Take the original photo and do it again. Do it in whatever color space you want. Now hit the button on each of the differently timed photos. They all get back to being pretty neutral, right? I’m not complaining, as this is an amazing tool in color recovery, but I don’t think you can say these were the colors on opening day.
-G
I kinda like the dirt left in. Seems more authentic than anything you could do to minimize it.
-G
I find the sharp Star Wars logo against the blurry stars, and the pristine footage next to the slightly more blurry with compression artifacts footage so jarring I’d much rather just watch the original.
Sometimes, just because you can doesn’t mean you should.
-G
Williarob is interested in using this method to color correct the high quality scan of the LPP that pitta has done. So, I certainly am interested in assisting him with that.
How do I <thumbs up> <like> <love> or all of the above a post with this new forum software?
-G
F*** me Dre, stop being so awesome! I would worry about correcting all the frames individually, as I bet you would introduce flicker, but if you averaged the resulting LUT from several frames/shots and applied the same correction to a whole shot/reel, you really would have something there.
I would even argue that if you did it right, it would be just as valid of a color reference as a Technicolor print, or the JSC, which are both slightly wrong, but your result wouldn’t be any more wrong.
-G
Last I looked at it, I thought the cartoon segment on the new Edit Droid looked better than mine. Mine was based off of the Blu-Ray bonus, which had undergone some severe DVNR. The Edit Droid version has almost as much resolution, no smearing, and probably more correct colors.
-G
ROTJ v.2.5
00:30:34-00:30:38- Sarlacc has stationary, narrow and thin black line, and sometimes flickers.
The Grindhouse and the JSC also have that line. I’d leave it alone if I was making the call.
Wazzles said:
Maybe that’s why Owen looks so off to me. Could you try it on that shot at the dinner table?Sure! With the same settings . .
.@ DrDre
While doing the above, I tried different instances of JPEG-DNR: 1 application on “high”; 4 repeat applications on “low”; a split-RGB application – RED on low, GREEN on normal, BLUE on high (for progressively worse damage). Unfortunately, all the instances looked identical when inspecting pixels at high magnification.
Thanks for the integer/double explanation. In reconsideration of color-depth, although it’s good for number precision, it would be a lost cause due to down-rez rounding or truncation that would lose most or all that pixel precision anyway.
Rather, how about the resolution approach? If resolution were doubled (not pixel-doubling, but a standard averaging [not “smart resize”] for minimal processing to the original pixels), the new pixels between the original pixels would better catch your color regrade. Then, on downsizing, that new pixel information would distributed into the neighboring, original pixels. (If that would prove workable but insufficient, various resizers could be tested for both up-rez & down-rez to a best result.)
Here’s a test using that approach – with JPEG Artifact Remover on “maxium”, and using weighted average for both up and down resizing . .
. . . . . . . . . . . . . . color regrade . . . . . . . . . . . . . . . . . . . . . . . with up-rez / JPEG-DNR max / down-rez See how this up-rez/down-rez method keeps more detail, even at the highest DNR setting, than the previous, direct application at the next lower DNR setting.
(Note that the original Blu-ray has R/G/B-crush on this shot, which results in only the red showing in the beard shadow, looking quite flat. Also, a DNR-pass [a la Spaced Ranger?] on the Blu-ray to reduce or eliminate those anomalies will prevent your regrade-pass from bringing them out even stronger.)
I don’t prefer the smoothed out look from what appears to be a spatial median filter.
-G
If Harmy pulls in both Dre’s and Neverar’s methods into his next versions, there’s going to be little left for us to gripe about in terms of color.
Nice work, both of you!
-G
