[Bf-committers] It's time to get rid of cloth

Tue Oct 20 15:24:27 CEST 2020

Hi all,

I haven't emailed on this email list before, but I'm happy to see
discussion about the cloth simulator in Blender. It is the main reason why
I started using Blender, and I know a fair amount of folks in my community
(AI and robotics researchers) who would love to explore cloth capabilities.

I have used (Py)Bullet recently, and have collaborated with the folks
working on cloth simulation as part of my Google internship. While they are
working on it (see https://github.com/bulletphysics/bullet3/pull/2872) from
my experience, for more physically realistic cloth, I honestly go with
Blender. The main reason why I have used (Py)Bullet has less to do with the
simulator quality and more to do with issues I ran into with vertex pinning
/ unpinning in Blender. For reference, I am the person who wrote this:
https://devtalk.blender.org/t/pinning-and-moving-vertices-of-cloth-at-arbitrary-frame-sequences-with-updated-physics/13542

Daniel

On Tue, Oct 20, 2020 at 8:24 AM darkdefende--- via Bf-committers <
bf-committers at blender.org> wrote:

> I see that my name isn't with my email address.
> So just so clarify, I'm Sebastian Parborg.
>
> On Tue, Oct 20, 2020 at 02:20:46PM +0200, darkdefende at gmail.com wrote:
> > To fix the cloth tunneling though the cube, activate "Single sided" in
> > the cube's collision settings and it should work much better.
> >
> > Anyways, during my recent work with Bullet, I tested some things that
> > are currently not integrated into blender.
> >
> > Mainly Bullet softbodies and cloth.
> >
> > Both have their use cases, but as Brecht pointed out, the Bullet cloth
> > physics is way more unstable and doesn't have no where near as many
> > features as our current cloth sim. So outright replacing our current
> > system with that would just be a downgrade on so many levels (except
> > speed).
> >
> > I would say that all of our physics simulations currently needs a lot of
> > work to be faster, more stable, and nicer to use. The majority of them
> > feels half baked at the moment I my opinion. (And it shows as it seems a
> > lot of people does not use them in production because they are simply not
> > good enough yet)
> >
> > My current plan was to first make Bullet integration in Blender better.
> > As I deem this to be the easier task. However I have a lot of other work
> > I need to do, so progress is slow.
> > This is also because I want the Bullet refactor to align nicely with our
> > new node system. (So we don't have to do an other refactor for that)
> >
> > As Sergej pointed out, Mano Wii and I has been in talks on making the
> > collision system better for cloth and also make it have move options so
> > they user can make trade-offs if they want to.
> >
> > If you have any specific fixes or changes in mind. We would love to hear
> > them.
> >
> > Welcome back.
> >
> > On Tue, Oct 20, 2020 at 01:39:01PM +0200, Sergej Reich via Bf-committers
> wrote:
> > > Bullet's cloth system is no replacement for what Blender currently has
> and
> > > I'm not aware of any available library that we could use without
> having to
> > > do a lot of modifications to make it work.
> > >
> > > Am Di., 20. Okt. 2020 um 12:26 Uhr schrieb Brecht Van Lommel via
> > > Bf-committers <bf-committers at blender.org>:
> > >
> > > > The most important thing for cloth is that you can get it to give
> good
> > > > results in real world cases, and the cloth simulation changes were
> based on
> > > > work done for Agent 327.
> > > > https://www.youtube.com/watch?v=jU6-cI-z8HA
> > > >
> > > > In practice you use smaller time steps and the fast-moving cube will
> get
> > > > collided with anyway.
> > > >
> > > > Continuous collision detection is great for handling fast-moving
> objects,
> > > > but arguably handling cloth that is nearly at rest and continuously
> > > > colliding with the character is a much harder problem. Most cloth
> > > > simulation demo videos stop while everything is still in motion, but
> what
> > > > happens after is just as important.
> > > >
> > > > Anyway, there's definitely much room for improvement here, but it
> should be
> > > > tested on real-world cases. For example from what I've seen, Bullet
> cloth
> > > > is more reliable at high frame rates, but does not produce good
> results or
> > > > wrinkling details when turning up the quality. I've not yet seen an
> > > > open-source cloth library that we could just plug in and get much
> better
> > > > results, but it would be great if there was one.
> > > >
> > > > Some work on cloth collision is being done here:
> > > > https://developer.blender.org/D8577
> > > >
> > > > On Tue, Oct 20, 2020 at 11:52 AM Ton Roosendaal via Bf-committers <
> > > > bf-committers at blender.org> wrote:
> > > >
> > > > > Hi Joe,
> > > > >
> > > > > This is a bit harsh, there have been good cloth contributions to
> Blender
> > > > > by Luca Rood. The current module owner for it is Sebastian Parborg
> > > > though.
> > > > >
> > > > > I'd suggest to connect with them to see what steps forward could
> be.
> > > > >
> > > > > We are also in the process of making a good node physics system to
> allow
> > > > > solvers like this in a more generic setting. Design for this is
> ongoing.
> > > > >
> > > > > It's super cool to see you want to come back, just give it a bit
> of time
> > > > > to figure out who's doing what now.
> > > > >
> > > > > Als note that we (including Blender Foundation / Institute) are in
> the
> > > > > process of organizing ourselves much better, including onboarding,
> best
> > > > > practices, engineering standards amd strict requirements for a good
> > > > > design process - including the docs.
> > > > >
> > > > > We lack in a lot of areas, and the consensus is to especially put
> > > > > efforts in bringing the current state of the code in control
> (functional
> > > > > and technical specs), before recoding modules entirely.
> > > > >
> > > > > -Ton-
> > > > >
> ----------------------------------------------------------------------
> > > > > Ton Roosendaal - ton at blender.org - www.blender.org
> > > > > Chairman Blender Foundation, Director Blender Institute
> > > > > Buikslotermeerplein 161, 1025 ET Amsterdam, the Netherlands
> > > > >
> > > > >
> > > > > On 20/10/2020 00:30, Joe Eagar via Bf-committers wrote:
> > > > > > So, ten years ago I wrote several continuous collision detection
> > > > systems
> > > > > > for cloth/hair, one of which I tried to push into trunk:
> > > > > >
> > > > > > https://www.youtube.com/watch?v=QBM3TBKO9w8
> > > > > >
> > > > > > And, here's how the cloth sim performs today:
> > > > > >
> > > > > > https://www.youtube.com/watch?v=5462m9yEkU8&feature=youtu.be
> > > > > >
> > > > > > If the cloth development process is this dysfunctional (this
> patch
> > > > > actually
> > > > > > passed review?
> > > > > >
> > > >
> https://developer.blender.org/rB0666ece2e2f96571200d693d9d7bee1ca72d026f
> > > > > )
> > > > > > then I propose we eliminate the cloth code entirely and go with
> a third
> > > > > > party lib (Bullet?).
> > > > > >
> > > > > > Joe
> > > > > > _______________________________________________
> > > > > > Bf-committers mailing list
> > > > > > Bf-committers at blender.org
> > > > > > https://lists.blender.org/mailman/listinfo/bf-committers
> > > > > _______________________________________________
> > > > > Bf-committers mailing list
> > > > > Bf-committers at blender.org
> > > > > https://lists.blender.org/mailman/listinfo/bf-committers
> > > > >
> > > > _______________________________________________
> > > > Bf-committers mailing list
> > > > Bf-committers at blender.org
> > > > https://lists.blender.org/mailman/listinfo/bf-committers
> > > >
> > > _______________________________________________
> > > Bf-committers mailing list
> > > Bf-committers at blender.org
> > > https://lists.blender.org/mailman/listinfo/bf-committers
> _______________________________________________
> Bf-committers mailing list
> Bf-committers at blender.org
> https://lists.blender.org/mailman/listinfo/bf-committers
>