EEVEE-Next: Improve shadow bias #121088

Clément Foucault · 2024-04-25T19:08:02+02:00

Clément Foucault commented

2024-04-25 19:08:02 +02:00

The goal of this PR is to remove any user facing parameter
bias that fixes issues that are caused by inherent nature of
the shadow map (aliasing or discretization).

Compute shadow bias in the normal direction to avoid
both shadow leaking at certain angles or shadow
acnee.

This bias is computed automatically based on the minimum
bias amount required to remove all errors. The render
setting is removed as of no use for now.

Normal bias

We do the bias in world space instead of shadow space
for speed and simplicity. This requires us to bias using
the upper bound of biases for the same location in space
(using biggest texel world radius instead of UVZ bias).
This isn't much of an issue since the bias is still less
than 2 texel. The bias is still modulated by facing
ratio to the light so surfaces facing the light have no
biase.

This fixes both self shadowing and flat occluders aliasing
artifacts at the cost of moving the shadow a bit on the
side. This is the blue arrow in the diagram.

We always bias toward the normal direction instead of the
light direction. This is alike Cycles geometric offset for
the shadow terminator fix. This is better since it does'nt
modify the shading at all.

Slope bias

To avoid aliasing issue on zero slope receiver, we still
have to use the slope bias with a size of 1 pixel.

No bias	1px Slope Bias	1px Normal Bias	Combined

PCF filtering

We now parametrize the filter around the normal instead of
using the shadow map local space. This requires to use
a disk filter instead of box, which is also more pleasant
for most light shapes (all except rectangle lights).

Setting the filter around normal avoid overshadowing from
zero slope occluders. This cannot be fixed by more slope bias
in light space PCF. We could fix it in light space by projecting
onto the normal plane but that gives an unbounded bias when N.L
is near 0 which causes either missing shadows or self shadow if
using an arbitrary max offset value.

Light PCF	Light PCF + Slope Bias	Normal PCF

To avoid overshadowing from any surface behind the shading
point, we reflect the offset to always face the light.
Doing so instead of using the perpendicular direction
is better for very sharp geometric angles, has less
numerical precision issue, is symetrical and is cheaper.

Normal PCF	Normal PCF + Reflect

To avoid any self shadowing artifact on zero slope receivers
with angled neighbors (like a wall and the floor), we have
to increase the slope bias according to the filter size.
This might be overkill in most situation but I don't feel
this should become a setting and should be kept in sync
with the filter. If it has to become an option, it should
simply a factor between unbiased filter and best bias.

Normal PCF + Reflect	Normal PCF + Slope bias	Normal PCF + Reflect + Slope bias

Shadow terminator

The remaining artifacts are all related to shadow terminator
one way or another. It is always caused by the shading
normal we use for biasing and visibility computation not
being aligned with the geometric normal.

This is still something we need a setting for somewhere.

The goal of this PR is to remove any user facing parameter bias that fixes issues that are caused by inherent nature of the shadow map (aliasing or discretization). Compute shadow bias in the normal direction to avoid both shadow leaking at certain angles or shadow acnee. This bias is computed automatically based on the minimum bias amount required to remove all errors. The render setting is removed as of no use for now. ![](https://projects.blender.org/attachments/12998611-2b81-48f1-83fb-338dd74d66f9) #### Normal bias We do the bias in world space instead of shadow space for speed and simplicity. This requires us to bias using the upper bound of biases for the same location in space (using biggest texel world radius instead of UVZ bias). This isn't much of an issue since the bias is still less than 2 texel. The bias is still modulated by facing ratio to the light so surfaces facing the light have no biase. This fixes both self shadowing and flat occluders aliasing artifacts at the cost of moving the shadow a bit on the side. This is the blue arrow in the diagram. We always bias toward the normal direction instead of the light direction. This is alike Cycles geometric offset for the shadow terminator fix. This is better since it does'nt modify the shading at all. #### Slope bias To avoid aliasing issue on zero slope receiver, we still have to use the slope bias with a size of 1 pixel. | No bias | 1px Slope Bias | 1px Normal Bias | Combined | | -------- | -------- | -------- | -------- | | <img src="https://projects.blender.org/attachments/1bbab178-7064-4fd9-b82e-47dc11c423a4" width="200"> | <img src="https://projects.blender.org/attachments/140e4a84-4c77-4b6c-bfc6-b035efd32329" width="200"> |<img src="https://projects.blender.org/attachments/17b8e914-fa46-4639-8e46-bbd8d01d00c1" width="200"> |<img src="https://projects.blender.org/attachments/6ed2bfd4-83e1-441a-88a2-4e1ea736e1e6" width="200"> | #### PCF filtering We now parametrize the filter around the normal instead of using the shadow map local space. This requires to use a disk filter instead of box, which is also more pleasant for most light shapes (all except rectangle lights). Setting the filter around normal avoid overshadowing from zero slope occluders. This cannot be fixed by more slope bias in light space PCF. We could fix it in light space by projecting onto the normal plane but that gives an unbounded bias when `N.L` is near 0 which causes either missing shadows or self shadow if using an arbitrary max offset value. | Light PCF | Light PCF + Slope Bias | Normal PCF | | -------- | -------- | -------- | |<img src="https://projects.blender.org/attachments/2f67cc52-5479-4595-8230-c73b0aaf4266" width="200"> |<img src="https://projects.blender.org/attachments/aa6e565f-11b1-460c-8725-24f7db8c4bf4" width="200"> |<img src="https://projects.blender.org/attachments/489e7a59-16e5-4c14-adbe-9da256d31bce" width="200"> | To avoid overshadowing from any surface behind the shading point, we reflect the offset to always face the light. Doing so instead of using the perpendicular direction is better for very sharp geometric angles, has less numerical precision issue, is symetrical and is cheaper. | Normal PCF | Normal PCF + Reflect | | -------- | -------- | |<img src="https://projects.blender.org/attachments/489e7a59-16e5-4c14-adbe-9da256d31bce" width="200"> |<img src="https://projects.blender.org/attachments/18737ed8-47d2-4f3f-91ab-66e9cf570f4f" width="200"> | To avoid any self shadowing artifact on zero slope receivers with angled neighbors (like a wall and the floor), we have to increase the slope bias according to the filter size. This might be overkill in most situation but I don't feel this should become a setting and should be kept in sync with the filter. If it has to become an option, it should simply a factor between unbiased filter and best bias. | Normal PCF + Reflect | Normal PCF + Slope bias | Normal PCF + Reflect + Slope bias | | -------- | -------- | -------- | |<img src="https://projects.blender.org/attachments/18737ed8-47d2-4f3f-91ab-66e9cf570f4f" width="200"> |<img src="https://projects.blender.org/attachments/6cac3096-fc81-4754-b2e7-28d1f0d92001" width="200"> |<img src="https://projects.blender.org/attachments/8caded22-abd2-4ec4-b84a-f2ae8b5242f8" width="200"> | #### Shadow terminator The remaining artifacts are all related to shadow terminator one way or another. It is always caused by the shading normal we use for biasing and visibility computation not being aligned with the geometric normal. This is still something we need a setting for somewhere.

shadow_map_bias.drawio(2).png

72 KiB

Capture d’écran 2024-04-27 à 10.10.50.png

635 KiB

Capture d’écran 2024-04-27 à 10.06.48.png

656 KiB

Capture d’écran 2024-04-27 à 10.00.16.png

718 KiB

Capture d’écran 2024-04-27 à 09.53.08.png

614 KiB

Capture d’écran 2024-04-27 à 09.51.47.png

615 KiB