Same file looks blurrier and with some artifacts compared to Supersplat

[MotivaCG]

Sparks looks great! but unfortunately, it degrades the quality compared to supersplat web viewer

You can get the test file from https://wetransfer.com/downloads/493bfd6235042deb9aa6120c6bcd8c3020250707075148/848258e49f8610cab6ecb4457b15fdf120250707075308/923e87?t_exp=1752133908&t_lsid=9dbdce9f-aace-40a4-8933-efdb82931da1

Regards, Víctor

[dmarcos]

Thanks so much for filing. We’re looking into it

[asundqui]

@MotivaCG
One thing we can try: In the Spark Editor we added a "Debug" folder and in there are two buttons "Non-AA preset" and "AA preset". When I click the "Non-AA preset" button I believe it sets the gaussian splat blurring parameters to be equal to what SuperSplat does... It looks noticeably brighter and sharper to me.

Does this explain most of the remaining difference? Or does it still seem softer/blurrier with Spark?

https://sparkjs.dev/examples/editor/

[MotivaCG]

@MotivaCG One thing we can try: In the Spark Editor we added a "Debug" folder and in there are two buttons "Non-AA preset" and "AA preset". When I click the "Non-AA preset" button I believe it sets the gaussian splat blurring parameters to be equal to what SuperSplat does... It looks noticeably brighter and sharper to me.

Does this explain most of the remaining difference? Or does it still seem softer/blurrier with Spark?

It looks different but I'm worried that still worse than supersplat. Apart from sharpness, for example, you can look at the sleeve of the T-shirt where you can see several problems.

Regards.

[asundqui]

It looks different but I'm worried that still worse than supersplat. Apart from sharpness, for example, you can look at the sleeve of the T-shirt where you can see several problems.

@MotivaCG Can you explain in more detail what you mean by the sleeve having several problems? I'm trying to look carefully between Spark and SuperSplat and it's hard for me to tell... maybe we could add an option to the Spark editor/viewer where you can select the background color so we could match the gray in SuperSplat.

One more stupid question from me, have you tried turning off Radial sort?

This makes the splats sort by z-coordinate, which will match SuperSplat. The default "radial sort" sorts by r=sqrt(x^2+y^2+z^2), which is more stable under viewpoint rotation, but is different from the training target. You may have to rotate the viewpoint slightly after unclicking "radial sort" for it to update the sort order. To me the results look slightly different with it checked on/off - I'm wondering if this explains more of the differences you're seeing?

[mrxz]

The problems on the sleeve seem to be caused by the Radial Sort when I tested it. The remaining "blurriness" compared to SuperSplat seems to be due to different calculations to the focal length in Spark compared to SuperSplat. I've experimented with matching the SuperSplat behaviour in this regard and that brings the resulting visuals very close (#102).

Alternatively lowering the preBlurAmount to 0.1 seems to get closer as well, but clear differences remain and lowering it further makes the result noticeably aliased.

See the following comparison showing the result after disabling radial sorting, applying the above mentioned tweaks (#102) compared to SuperSplat:

[asundqui]

This is a very nice comparison. @mrxz how did you get the cameras to line up perfectly between SuperSplat and Spark?
Btw I just added a PR to allow background color changing in the editor, to help with comparisons in the future... #103

[mrxz]

how did you get the cameras to line up perfectly between SuperSplat and Spark?

As long as all camera parameters match (fov, near, far, position and orientation), the views line up. I currently have Spark, SuperSplat and Mkkellogg setup in a consistent way, where the loaded model is always at the origin. For each test I choose a suitable camera position and use that for all three.

[simonbethke]

While I totally see why spark should look similar to other tools, I still think that this is not how a correct splat should look:

[dmarcos]

I'm reopening this issue to confirm fix by @MotivaCG. 0.1.6 contains a fix that should address the artifacts. Can you confirm if the file looks good to you now? Thanks

[dmarcos]

@MotivaCG Try setting SparkRenderer focalAdjustment to 2.0

[MotivaCG]

Unfortunately it still seems to have less "depth" of colors and some issues with alphas, take a look at the sleeve and her hair.

[mrxz]

@MotivaCG Based on the images it looks like Spark is still using radial sorting, did you disable it?

Could you try the following setup:

const sparkRenderer = new SparkRenderer({
  renderer: renderer,
  preBlurAmount: 0.3,
  blurAmount: 0.0,
  focalAdjustment: 2.0,
  view: {
    sortRadial: false
  }
})