Bifrost Beta

[Mathaeus]

So from user perspective, that's more like Houdini, where's possible to get immediate response, but not all the time. User perspective in Houdini : as long as something is exposed as parameter on 'top'' of network, response is immediate, while tweaks inside the network (thing like mixing values or so) are taking some small time to respond (perhaps compiling into background).

[Bullit]

3Ds Max users asked for an ICE - it was the first request - and they tried to give them in form of MCG -Max Creation Graph, in the end it failed precisely because of compilation.

Maybe merging Mash with Bifrost?... but since it is supposed to be standalone.

[Mathaeus]

I'd say ''interactive'' is better word. MCG is not interactive, that's it. Compilation could be set enough fast for efficient interactivity, Houdini is great example of that (while I believe it's really complex system in Houdini, of lazy evaluation and in-advance analyzing the network, behind that).
Other than that, I think there are many factors, important for survival of such thing, like interest of community, structure of network, state of already available traditional options. Softimage ICE had almost ideal conditions, a lot of people were ready to try, they had very precise demands and wishes based on already strong classic 3d app (a lot of Houdini people simply don't know what to, what's their goal with this app), ICE is still probably most artist friendly system of this kind (or, least artist unfriendly). Finally, some of built in features like XSI hair were ''ideal enemy'' - with ICE, it was easy to show the impossible things for SI hair, while (imho false) reputation of Shave helped these achievements to look even much greater. Later, ICE wasn't successful against old but state-of-the-art Softimage kinematics....
In any case, ''innovation'' is a plain appearance of such system in widely accepted 3d app. My impression about Mayans (developers, users, support), it is much better understanding of whatever I asked, than in Houdini world. On other not-so-bright-side, a lot of old-fashion 90s style thinking, where user was basically a servant of 3d app instead of reverse, arrogant ignoring of any no-big-facility workflow and 3d app.
We''ll see, I hope.

[luceric]

I have not used MCG, although I spoke a few times with the author. It's my understanding that it generates a maxscript plugin, and then installs or reloads it then you make some types of changes, and some part of the workflow aren't truly interactive because of that implementation.

Bifrost doesn't work that way, it doesn't generate a maya plugin. There is a maya (or arnold, or max, ...) plugin that uses Bifrost to evaluate a graph, and bifrost with just-in-time compile that graph before executation, if necessary, just like Java, Python, or the graphic drivers on compile shaders do if you change the code.

That difference is subtle, but it makes the difference between being able to edit a graph during playback or not

[rray]

I don't even understand the need to compile the graph, just-in-time or not, I thought a node should be a precompiled unit like for example a function.

That it would be one of the points of having nodes. Still having to check out LLVM, it sounds really interesting, maybe it already optimizes things that way.

[Bullit]

I have not used MCG, although I spoke a few times with the author. It's my understanding that it generates a maxscript plugin, and then installs or reloads it then you make some types of changes, and some part of the workflow aren't truly interactive because of that implementation.

Bifrost doesn't work that way, it doesn't generate a maya plugin. There is a maya (or arnold, or max, ...) plugin that uses Bifrost to evaluate a graph, and bifrost with just-in-time compile that graph before executation, if necessary, just like Java, Python, or the graphic drivers on compile shaders do if you change the code.

That difference is subtle, but it makes the difference between being able to edit a graph during playback or not

After first criticism they tried to make it interactive but it seems it was only possible to make some parts of it. I think they didn't touched it in last 2 versions.

[luceric]

I don't even understand the need to compile the graph, just-in-time or not, I thought a node should be a precompiled unit like for example a function.
That it would be one of the points of having nodes. Still having to check out LLVM, it sounds really interesting, maybe it already optimizes things that way.

It doesn't do anything for graph made of a few "big nodes", but if you make a compound all with basic nodes, they are actually provided as LLVM bytecode, and then whole graph runs through the compiler optimizer, which will eliminate redundancy, expand loops, static branches are optimized, etc. It's like if code was inlined in C++, if that make sense. Then in the future perhaps they can use the LLVM backends for the GPU and run it on that.