Nov 17, 2017

Quick tip: create fog from deep render

Hi, I'm back on track :)
I always had problems with creating good fog from zdepth passes. Our 3d guys have the habit of rendering almost everything in one layer, so when I grade a zdepth pass and merge (over) it on top of the comp, often get some kind of a halo. The fog around objects in foreground, which are not covered, get wider thanks to grading the anti-aliased pixels. If I use an unfiltered depth pass, then I have jaggy edges. So it is always hacking with erode, blur and stuff, to make it more pleasing, probably you know that. But recently I discovered a quite good method. I use only 1 frame of a basic  3d fog (volume) rendered in deep. This frame I scale and transform and the make a holdout with the current shot deep render (yes, this is the only throwback, unfortunately we need that in deep), and that I use as a fog. I don't use any contrast modifier in order not to ruin the edges, just some coloring. Because the 3d fog doesn't have any noise in it, doesn't have to be rendered for each shot and every frame. 1 frame is enough. After this setup I can use some noise on cards to break it up (brightening/darkening with the noise as masks), in order not to look so uniform.

Old graded Zdepth method (left) vs. deepfog method (right) with a character in front. Notice the hair
Of course I can use this for creating depth-based masks for color grading of the areas. Here you can see some images of this method, and if you are interested I can attach my deep fog sample, and my basic setup. Let me know in comments.

Node setup

The control node has offset and multiply values, with that you can control the coverage of fog. Here you can see two different settings:

Depth mask with different settings, offset and multiply.
Adding some slight variation with noise:

Detail with noise

Oct 4, 2016

Nuke studio edit export to mrviewer

Hi, new post again, whoooa! Sorry for the delay! But we had a big-big project in the first half of the year: we had the honor to work on the new feature movie made for the Final Fantasy XV game, called "Kingsglaive". It was an extreme effort for our studio, to do the more than 13 minutes of photorealistic cg animation, from previz to compositing, but at the end we managed to complete 433 shots! I took part as lead compositor, and I can say we learned a lot on this show. Thanks to Square Enix for getting us on board, despite of that at the end we had our most rough and challenging weeks of all time, in terms of workload and the lack of sleep :)
So for this tip: we are using nuke studio for handling our editing work. That means usually from previz, every editorial job is done in this software. Without this, and its export system, would have been much much harder to deliver all the shots in all the requested formats for Square. The problem was that we have only a few licenses, and a lot of artist should be looking at the cut. Exporting the cut to quicktimes is not a perfect solution, because it's color, quality, can't zoom in, and it's not showing always the latest state of the work, just in the moment of exporting. So I searched and found a solution: mrviewer! Its brilliand piece of open source player, with a lot of options. (Sometimes it crashes, but which software doesn't? :)
Here is a code snippet, you can export the edit from nuke studio, to mrviewer. Its not handling dissolves or anything fancy, just putting the clips one after the other, with the provided in and out points. But usually that's what we need.

from hiero.core import *
import os.path
import re
myProject = projects()[-1]
clipsBin = myProject.clipsBin()
selItems = hiero.selectedItems
  len_selitems = len(selItems)
  len_selitems = 1
print len_selitems
clipsPaths = []
if len_selitems > 1:
  for item in hiero.selectedItems:
    if type(item) == TrackItem:
      print "\n-----------\n%s"%
      filepath = item.source().mediaSource().firstpath()
      ssin = item.source().sourceIn()
      print 'Clip source (version) in: %s ' % str(ssin)
      ssout = item.source().sourceOut()
      print 'Clip source (version) out: %s ' % str(ssout)
      usedSInRelative = item.sourceIn()
      print 'Clip (shot) in: %s ' % str(usedSInRelative)
      usedSOutRelative = item.sourceOut()
      print 'Clip (shot) out: %s ' % str(usedSOutRelative)
      usedSourceInResult = ssin + usedSInRelative
      usedSourceOutResult = ssin + usedSOutRelative
      print 'Clip source startTime: %s' % item.source().mediaSource().startTime()

      clipsPaths.append([filepath, usedSourceInResult, usedSourceOutResult, ssin, ssout ])
path = 'c:/temp/XY_editorial_COMP.reel'
fileh = open(path, 'w')
fileh.write('Version 2.0\n')
for line in clipsPaths:
  fileh.write('"%s" %s %s %s %s\n' % (line[0], int(line[1]), int(line[2]), line[3], line[4]))

You will need this pythonscript installed for studio.
So the workflow is: in studio select the needed clips, prefereably cut one after the other in one track. Use the getpythonselection script, that store the clips in python objects. 
Edit the output path to the desired path in the script at line which now says:path = 'c:/temp/XY_editorial_COMP.reel'
Keep the extension ".reel"
Run this script. It writes out the reel file to the provided path. In mrviewer you can now open this as an image. It will load the clip list, but in order to see the entire cut, you need to press the edl button in the "reels" dialog.
Hope this helps for you. If you have problems, or find a bug, or just used this script without any problems, let me know.

May 7, 2015


Hi, new post at last! Yippee!
Actually the credits of this nuke tip entirely goes to my colleague Zsolt Sebők, who had the idea, and did the research and experimentation with it, and kindly permitted me to present on this blog. Thanks Sebi!
So the as compositors in an animation house, we often find ourselves in a situation that we need to import geometry from 3d scene. For mattepaints, or for render error fixing etc. Sometimes they (3d guys) export it, but sometimes we have to open and export it ourselves, if want to get it in time. It's a painful and slow process, especially when the scene is big. Further problem when geometry is present in the lighting scene only as a standin, so we have to go further back in pipeline. I assume you know what I'm talking about.
The idea is, that we can generate fairly accurate geometry from positionpass and normalpass of the renders. As a primarily animation house we do mostly full cg animation, so we always have passes like world position and normal from arnold, and with that and a clever node positionToPoints, we can generate an accurate pointcloud.  This is step one.
Nuke has this poisson mesh node, that can generate geometry from pointclouds. Probably it was designed to work with pointclouds from cameratracked scenes. But in our case we have the pointcloud from the passes, and is much more correct, than a pointcloud generated from matchmove can be. This is step two.
So the workflow is this: we make a pointcloud, that I'm sure everybody does with positiontopoints node, to clearly see the 3d scene, to help with inserting cards, tweaks etc. Then attach a poissonmesh node, and that's it. Be careful, there are several PositionToPoints node types. The one that has class (press i on the node) attribute PositionToPoints2, and loaded by PositionToPoints2.dll. We had an older version of that (PositionToPoints.dll), and that couldn't load the normal pass properly, and poissonmesh couldn't generate anything. Be careful to use unfiltered positionpass, otherwise you will get incorrect points at the edges.
We found the poissonmesh default settings are quite good, in terms of speed and quality. Maybe the size parameter would be better at 1. If you need more geo detail, increase depth a bit, but carefully, this makes much slower the process. The geo generated is much detailed where the pointcloud is present, the empty areas are approximated and rough. But usually good enough.
If not all parts of the scene needed, a good thing to stencil out the unneeded part from beauty, so nuke can handle much better (less geometry), and detail is not spent on stenciled parts.
After the geometry is generated, it can be used for mattepaints, projecting some image on that.  
We have several ideas to improve this workflow, but that will be covered later. Hope you will find this tip useful!
Here are some example screenshots:

Needed passes: beauty, unfiltered worldposition, normal

Generated pointcloud

Generated geometry on top of the pointcloud

Filtered (wrong) positionpass, points with incorrect position all over

Cloud from side

With geometry

Quite good fit of the geometry

Center part of igloo is masked, faster operation

Dec 4, 2014

Radio buttons in nuke

Hi, I always wanted to have the option of creating radiobuttons in the nuke gui, for gizmos. Its easier to use than dropdown menus, because they  need only on click to select the value. Today I checked the documentation, and I found an option. The radiobuttons are a different widget version of the enumeration knob. You can't add this knob like others via the "manage user knobs" panel, but you can with the help of a little python. For example like this:

n = nuke.toNode('Group1')
r = nuke.Radio_Knob('options', 'Option channels', ('red', 'green', 'blue', 'alpha'))

First argument is the name of the knob, second is the label, and the third is a tuple, with the radiobuttons' name in it.
Hope this was useful, cheers.

Apr 16, 2014

Mattepaint camera setup for multiple shots

Hi, so it lasted for a little while I managed to really post a new one! :) Sorry for that!
So in our current project we have a few sequences where we are using background mattepaints. Nothing new at this point. But I really like the way I create the camera and the framing for the mattepaint, from the existing shot cameras. The concept is that we should give the painter the least possible work :) (Not because he is slow) The less number of paintings and less area to paint, is the better. So I exported the cameras from maya to nuke, all animated of course. I make an environment sphere, using an uv grid as a texture, just to see where are the cameras heading. I really like the camera projection in nuke, because I can project from every shotcamera a constant color onto this sphere, using mergemat node, and then scrubbing in the timeline I can see all cameras animating the projection in realtime. Then I use a checkerboard texture projection (see the second and third image below) for a newly created mattecamera. The purpose it should cover all the area covered by the other cameras (those are projecting different constant color). I usually duplicate one of the shotcameras (usually the one with the biggest field of view) that will be the new mattecamera. Then I lower the focal length, to make the fov bigger, and rotate as needed. If there are cameras facing completely different directions, then more than one mattepaint is needed, and we should group the cameras based on facing direction. These cameras I pass back to maya, and render a basic lit environment, that the painter uses as base for his mattepaint. When the paint is ready, I save layers as images from that, and project back onto proxy geometries of the original scene geometry. Or just onto cards, if further away. Using of course the same mattepaint camera as projector camera.

Node setup, the main thing is the projection using mergemat nodes to the sphere.

Projection in 3d view

Nov 12, 2013

New post!

Hi, at last we finished (or finishing just now) our big big project of the year: we created more than 10 minutes of full cg animation for Ubisoft's upcoming big car game called Crew. Unfortunately the videos are not public yet of course, until the game is released next year, but I can tell I'm proud of these videos, turned out quite fine (or most of them :)! Hope you will enjoy by the time the game released.
So this project was very heavy work, and I couldn't really concentrate on this blog. But now I will publish some useful experience of this work. Come back soon!

Apr 15, 2013

Adding source timecode to quicktime

So I've been searching for a solution for a current problem in our pipeline: I would like to embed source timecode in a quicktime movie, that is generated from image sequence, with either ffmpeg or ffmbc or similar. The timecode should be of course the sequence frame number. The problem is the sequences are not starting from frame 1, as we (at comp) and also the editorial uses the 3d dept. frame numbers (we are mostly doing full cg animation) As far as I did research ffmpeg can't write timecode got from the filesequence, on the other side ffmbc is capable of doing this, but can't start sequences from other than 1-4 frames (which I absolutely can't understand why it can't use ffmpeg -start_frame flag).
I would be glad if it could be solved with an open source software, but currently it looks that I have to use nuke for this also. Tried RV, but didn't work. Found out, that nuke has an addTimeCode node, that writes timecode information into metadata and fortunately writenode is able to embed it in the generated quicktime.

So what are the knobs in AddTimeCode node:
Start code: is a global starting position for cases where needs to be for example 01:00:00:00
Fps and get fps from metadata: if there is no fps metadata in upstream, uncheck the get fps from metadata checkbox, and set it manually, otherwise the output will be 24 frame per second.
Start frame: acts as a timecode offset.
For creating quicktime with proper timecode from image sequences (that are not starting from frame 1, I had to set the following settings:

start code: 00:00:00:01
fps: 30, get fps from metadata: unchecked
start frame: 1, use start frame: checked
This way for example if the source filesequence starts at frame 36, the timecode of the first frame will be 00:00:01:06

Very important that in write node there is a write time code checkbox, that must be checked. So this way when opening the mov in quicktime player, click at timecode numbers, and from the pop-up list choose Timecode:Non-Drop-Frame