Fix #22574: fix volumetric fog back-face intersection artifacts#23588
Fix #22574: fix volumetric fog back-face intersection artifacts#23588joaoconceicao12 wants to merge 5 commits into
Conversation
|
Follow up target branch correction from PR #23406 |
alice-i-cecile
added
C-Bug
alice-i-cecile
added
X-Uncontroversial
mate-h
left a comment
mate-h
left a comment
There was a problem hiding this comment.
This change breaks the fogvolume during my testing (no longer visible.)
|
The issue was happening on the release-0.18.0 branch(that is why my target branch was this branch in the other PR) and so was my solution for this based on it. In the older branch the atmosphere display a fog box volume that made the bug appear clearly. |
joaoconceicao12
force-pushed
the
fix-fogvolume-22574-main
branch
from
8486bdb to
56aead8
Compare
|
Thanks for the due dilligence - we never modified the example to remove the fog. Here are the lines in the example that add it to the scene: https://github.com/joaoconceicao12/bevy/blob/fix-fogvolume-22574-main/examples/3d/atmosphere.rs#L231-L235 So there must be something with this change set in this pull requests that makes it not work anymore. The render pass executes (checked this with xcode debugger) but the location of the rendered fogvolume is wrong. |
|
Thanks for the heads up you were right about the issue in rendering. After tracing the issue, i found the problem was in the volumetric fog intersection math, not in scene setup:
The fix updates both paths:
I also recorded a video showing the example running with this patch and camera rotation, to demonstrate that the issue no longer happens. (increased density_factor for better visualization) video2.movSorry for the delay im currently finishing my bachelor's degree at IST and exam week is coming up. |
| fn camera_is_inside_fog_volume_hysteresis(local_from_view: &Affine3A, was_inside: bool) -> bool { | ||
| // Entering interior requires being slightly deeper than the boundary, | ||
| // while leaving interior requires being slightly further outside. | ||
| const ENTER_THRESHOLD: f32 = 0.495; |
There was a problem hiding this comment.
These threshold values seem arbitrary. I think fundamentally we shouldn't distinguish between is inside and is outside volume, instead just fix the ray-plane intersection tests in the shader.
|
I appreciate all your efforts. Tested the example again, while I can see it from outside the fog volume, the fog seems to fade out when I get close with the camera, and when the camera is inside the volume it's not visible at all. This change/fix seems to be heading in the wrong direction. If you're busy - I could look into fixing it if you want to prioritize this issue! Also make sure to review our AI policy. |
|
Thanks for the feedback. Being honest i wasn't really sure on how FogVolume should behave when the camera is inside it or when approaching it that's my mistake for not expressing my doubt. I don't like delivering a half-done job so i will continue working to fix the problems you mentioned. |
|
Im having some doubts about the expected behaviour for the FogVolume seeing it from the outside. Could you explain its behaviour? |
|
That depends but the borders of the fogvolume should generally remain visible when not using a custom density texture. Any camera ray that hits the fog volumes bounding box will travel a non zero distance that is then attenuated with the beer lambert law to render it. Any kind of fading with distance is a rendering artifact. The only fading that should happen is when the ray travels less in the volume that is the camera is inside looking at and approaching the end of the bounding volume. Hope this clears things up ! |
|
@JeroenHoogers, your expertise from #23887 would also be welcome here :) |
…acts Fixes angle-dependent volumetric fog artifacts in the atmosphere example (shadow flipping and hard edges while rotating the camera). Root cause was inconsistent back-face intersection setup: far-plane construction in CPU code and ray-plane intersection in WGSL shader code. This patch corrects both calculations and adds a regression test for far-plane positioning in volumetric fog (get_far_planes). Signed-off-by: João Conceição <joao.conceicao@tecnico.ulisboa.pt>
…acts Fixes angle-dependent volumetric fog artifacts in the atmosphere example (shadow flipping and hard edges while rotating the camera). Root cause was inconsistent back-face intersection setup across CPU and shader paths: far-plane selection instability in CPU code and ray marching/view-space sampling inconsistencies in WGSL shader code. This patch makes far-plane selection deterministic, adds inside/outside hysteresis near fog boundaries, and corrects shader ray-plane intersection usage and view-space sampling origin. Signed-off-by: João Conceição <joao.conceicao@tecnico.ulisboa.pt>
This commit fixes the volumetric fog fading when aproaching and when inside of it. It also removes threshold values. Signed-off-by: João Conceição <joao.conceicao@tecnico.ulisboa.pt>
Fixes fog fading when leaving the Volume looking toward its end boundary. Also fixed black sky inside the FogVolume when leaving(view_end_depth_from_buffer was very far making it over-attenuated making the sky dark) Signed-off-by: João Conceição <joao.conceicao@tecnico.ulisboa.pt>
joaoconceicao12
force-pushed
the
fix-fogvolume-22574-main
branch
from
a52a297 to
8677ecb
Compare
|
While I really appreciate the efforts to fix this issue, I think this is not the correct approach. Especially the boundary fading logic seems undesirable and will produce non-physically accurate results. Like @mate-h suggested earlier, the fix should be limited to fixing the problems with the intersection tests. Given the back and forth on this fix already I think it is worth discussing whether this is the right path to follow. I'd be happy to look into this and explore alternatives. |
alice-i-cecile
added
S-Nominated-To-Close
|
@JeroenHoogers I'd be keen to see another PR with that approach. |
|
@alice-i-cecile @mate-h I fixed it here #24220 |
5 participants
Objective
Fixes #22574.
This PR fixes angle-dependent volumetric fog artifacts in the atmosphere example ([--features free_camera], where rotating the camera caused shadow flipping and hard edges.
Solution
The root cause was inconsistent back-face intersection setup between CPU and shader paths:
CPU side: far-plane construction for fog volume traversal.
WGSL side: ray-plane intersection math.
This PR:
Corrects far-plane setup in the volumetric fog CPU path.
Corrects the WGSL ray-plane intersection calculation.
Adds a regression unit test for [get_far_planes]) to prevent this from regressing.
Testing
cargo test -p bevy_pbr --lib passes (including the new regression test).
Manual visual test:
cargo run -r --example atmosphere --features free_camera
Rotating the camera no longer shows the previous flipping/hard-edge artifacts. (see video)
video1.mov
Platform tested:
macOS (Metal backend).
No additional untested platform-specific behavior is currently known.Fixes angle-dependent volumetric fog artifacts in the atmosphere example (shadow flipping and hard edges while rotating the camera).