Toolbag 3 vs 4
polycounter lvl 5
I remember when four version just come out, people were complaining a lot about it, and i was reading resons why this version was ipsetting and desided to stay with the third. But is it changed now? I use MT for baking textures
Replies
I still mainly work with version 3 unless a client explicitly requires 4. The noisy viewport in 4 gives me the headaches when raytracing is turned on and if that's disabled then the materials just look too different for the preview to be useful for what I do. Personally all I need out of Toolbag is a game-engine like preview, not some kind of Keyshot on the GPU (but without an offline renderer's accuracy).
Baking appears to yield same results in both versions for the limited things I do but I had less crashes in 3 so there too I default to the previous release.
Lastly I recently had an internet outage and 4 wouldn't start up during that. Not sure if it needs the net all the time or if I got unlucky with the timing but I was displeased to put it mildly.
Hey Thomas, always good to hear from you. Have you tried using the real-time denoiser when ray-tracing is on in TB4 (assuming you have an RTX card)? This can help considerably with the noise issue.
As to your comment "but without an offline renderer's accuracy" could you tell me what led to this conclusion? TB4's RT mode uses a path tracing engine that should produce comparable accuracy to most offline renderers.
The main change to baking in TB4 was ray-traced acceleration, so if you have an Nvidia RTX or AMD RX card, complex bakes like AO are much faster. The GPU-accelerated ray-tracing backends do require more VRAM though, so you may notice a regression in stability compared to TB3 when baking assets with tens or hundreds of millions of triangles - depending on your GPU. It's possible to switch to the generic (non-hardware accelerated) ray-tracing backend (Edit -> Preferences -> GPU), and this may improve stability with very dense high poly meshes or high-resolution textures.
As for the licensing problem, this was probably bad luck. Toolbag does not require an always-on internet connection, however, it does periodically need to connect to the license server and check to make sure the license is still valid. What probably happened here was the date it needed to check corresponded with the date you lost internet access. We've discussed this sort of situation and it may be possible to make this check smarter in the future.
We offer a free 30-day trial, so if you're curious if you can give that a go: https://marmoset.co/toolbag/
Many new features and improvements were added for the 4.0 release, as well as subsequent patches. 4.05 is currently in beta and you can test the latest features here: https://marmoset.co/toolbag/beta-history/
Here's a general summary of the differences between Toolbag 3 and 4
You can see the full change history here (minus the 4.05 changes): https://marmoset.co/toolbag/history/
@EarthQuake Hey Joe, yes I do have the denoiser active. Turns out I have no screenrecorder installed at the moment to do a video but with Denoise set to GPU I get a really unstable image in motion that looks like one of these painterly-filtered flickering early 2000's music videos. Also can't seem to set Denoise to quality 'High' for the GPU (in 4.04).
Keep in mind I'm trying to use Toolbag as an engine preview, that means manipulating the view all the time and/or spinning a turntable. The solution that is in there might be fine for having a fixed view and mostly just doing render setup but that's not what I'm after and also not what Toolbag up to and including version 3 offered.
As for my comment re: accuracy - I can see plenty of little artifacts in the shadow casting on my stuff. Hard edges on low poly geo sticking out in the shading, offset shadow casts on a surface, blotchy shadows, etc. It's simply my impression that TB is still pulling plenty of realtime trickery there. I certainly don't see stuff like that e.g. in Cycles or Keyshot.
Ok, did a little videograb of what I'm seeing. MKV file in the zip.
@thomasp
Yes, the painterly effect is an issue with denoisers for sure (in all apps that have them). It works better in some cases than others (scenes that require fewer samples to converge tend to be more stable). You can set the samples per pixel setting higher if your GPU can handle it, which will give cleaner early-frame denoising. If the denoise strength is set to less than 1, try turning it to 1. Otherwise, you're blending between the noisy result and the denoiser, which can make the interaction more unsettling. I tend to use a value less than 1 for final renders to retain a bit more detail, but leave it at 1 for the viewport. Generally, we're aware that the ray-tracer isn't as interactive as a game engine style renderer, and we're investigating ways to improve this.
"I can see plenty of little artifacts in the shadow casting on my stuff. Hard edges on low poly geo sticking out in the shading, offset shadow casts on a surface, blotchy shadows, etc. It's simply my impression that TB is still pulling plenty of realtime trickery there. I certainly don't see stuff like that e.g. in Cycles or Keyshot."
Toolbag's RT engine is a pure path tracer, not a hybrid raster system that rasterizes the base shading and blends in ray-traced passes for reflections, GI, etc like game engines. Shadows are ray-traced and generally very high quality. Coming from game engines, you may notice unexpected shadows on rough lowpoly geometry in all ray-tracers, this isn't something that is unique to Toolbag.
Here are renders from Modo (right), Toolbag (center), and Blender Cycles (middle). I don't have Keyshot installed at the moment but it would be surprising if their renderer is any different as this is a fundamental ray-tracing issue. The problem in all cases is that the shadows are too accurate, they are based on the mesh geometry and it's not possible to do the same tricks that raster renderers do to smooth them for low poly meshes. Raster shadows have all manner of other artifacts, though (shadow acne, resolution limitations, poor quality area light shadows, etc).
In Blender, there is a Geometry Offset set in Shadow Terminator that helps to minimize these artifacts. By default, it's set to a non-zero value, which makes the shadows less accurate (the shadows will shift).
In Toolbag there is an equivalent setting in the Render object -> Global -> Shadow Offset. This sets a scene-wide offset rather than a per-mesh offset, which is handy if you have a scene with many low poly meshes.
Here we have three renders from Toolbag, with the default offset of 0 (most accurate shadows), offset adjusted by taste, and adjusted too far. Note that adjusting the offset will shift the position of the shadow on the ground plane, and too much shadow offset introduces a different sort of artifact in the lower left corner of the geosphere.
You can see this sort of artifact with Blender's default Shadow Terminator settings too. So the main difference here is how much Blender is offsetting shadows by default (probably too much), and how the offset is configured (per-scene or per-mesh).
If you're noticing blotchy shadows, that may be the denoiser more than the ray-tracing system, and increasing the sample count is generally the solution. You'll likely see similar artifacts in other renderers with denoising options. In any case, if you have a problematic scene I'll be happy to look into it and see if there might be a bug or if I can provide a better explanation.
Just chiming in here to agree with Thomas P's comment. I just got the recent update to Toolbag 4, and almost immediately had to revert to the previous version. The raster view is just way too noisy when you're moving the camera around, even with the denoiser cranked up to max. I'm primarily using toolbag as a baker and a quick way to preview a WIP model in a proper real-time renderer without having to jump through all the hoops to get something in engine. Time spent using it as a presentation renderer is much, much lower.
Any chance you can revert some of the viewport changes, or give us a mode where it works more like it did in version 4.04?
Could you provide some more information on this, perhaps a test scene?
As far as I know, there should not be any changes to the raster viewport behavior between 4.04 and 4.05.
Hey EQ, just letting you know I haven't forgotten this. I'll take another look as soon as I have time. Totally possible it's something I screwed up on my end.
Just to clarify: I was talking about my preference for Toolbag version 3(.08) over any of the 4.x versions for my kind of work.
E.g. I have assets like full body fur on a creature done as (densely packed) hair cards that make the whole 4.x viewport strobe like mad if the denoiser is active. Impossible to see anything in there and even the still image the viewport settles into if you don't manipulate it looks soft and fuzzy, as if the Photoshop dust & scratches filter had been run over it. You basically have to render out stills to see anything at all.
Hey Thomas, thanks for the additional information. This makes sense, there are certain shading models that take longer/take more samples to converge than others. I assume you're using SSS, anisotropic, and dithered transparency for the fur? This sort of material setup is going to stress the renderer more than others and will require longer accumulation times.
One thing I would mention is that if you're using many light sources, you might try and simplify your lighting setup. Often when using ray-tracing you can get by with fewer lights while still achieving a good result, as global illumination does a better job of creating convincing shadows, bounced light, etc.
In Toolbag when you add additional lights, the samples are spread out over the light sources collectively. This means performance doesn't degrade significantly while adding lights, but it also means you will need more samples in total and the initial frame will get noisier as more lights are added. This will make denoising less effective as well. So if your scene has a bunch of lights, try seeing if you can reduce the count. You can also try turning on Advanced Light Sampling, this tends to reduce noise when using many lights but comes with a bit of a performance cost.
We're generally aware of issues with noise, denoising, and interactivity, and are actively working on improvements. If you're able to share this scene please send me a PM, as this may be a good case for our engineers to use while testing improvements.
PS: With 4.05 we made some improvements to the denoising strength slider, this now does a better job of filtering out fireflies or hot pixels, while retaining detail. So you should see better results when using a strength less than 1, which I find tends to help combat the painterly look of the denoiser at full strength.
@EarthQuake Thanks for the offer to look into my setups Joe but for one I can not share my assets and also I'm pretty confident that in this case it's beyond hopeless. I'll just stick with TB3 for realtime viewport display as long as feasible.
@thomasp hi, i was going to suggest showing your still, but read you could not so if you are trying to get rotations, i found that rendering stills (for how ever the amount, "uncompressed") and then porting them to a "stitching" application, into a video has worked as a work around if that helps at all. Just throwing a suggestion that might be beneficial, do as you wish, no need for response. Only suggested since you stated ?rendering out a still to see anything".