Skip to content

Fix critical bugs in Skyfall-GS codebase #12

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Hunter5Thompson wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix critical bugs in Skyfall-GS codebase #12

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 6 additions & 1 deletion render_video.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -156,7 +156,12 @@ def colorize_depth_torch(depth_tensor, mask=None, normalize=True, cmap='Spectral
if normalize:
min_disp = np.nanquantile(disp, 0.01)
max_disp = np.nanquantile(disp, 0.99)
disp = (disp - min_disp) / (max_disp - min_disp)
# Avoid division by zero for constant depth
disp_range = max_disp - min_disp
if disp_range > 1e-6:
disp = (disp - min_disp) / disp_range
else:
disp = np.zeros_like(disp)

# Apply colormap
colored = plt.get_cmap(cmap)(1.0 - disp)
Expand Down
7 changes: 6 additions & 1 deletion render_video_from_ply.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,7 +153,12 @@ def colorize_depth_torch(depth_tensor, mask=None, normalize=True, cmap='Spectral
if normalize:
min_disp = np.nanquantile(disp, 0.01)
max_disp = np.nanquantile(disp, 0.99)
disp = (disp - min_disp) / (max_disp - min_disp)
# Avoid division by zero for constant depth
disp_range = max_disp - min_disp
if disp_range > 1e-6:
disp = (disp - min_disp) / disp_range
else:
disp = np.zeros_like(disp)

# Apply colormap
colored = plt.get_cmap(cmap)(1.0 - disp)
Expand Down
49 changes: 38 additions & 11 deletions scene/dataset_readers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,13 +111,19 @@ def readColmapCameras(cam_extrinsics, cam_intrinsics, images_folder):
image = copy.deepcopy(image)
cx = (cx - width / 2) / width * 2
cy = (cy - height / 2) / height * 2

cam_info = CameraInfo(uid=uid, R=R, T=T, FovY=FovY, FovX=FovX,

# For COLMAP data, depth and mask are not available
depth = None
mask = None

cam_info = CameraInfo(uid=uid, R=R, T=T, FovY=FovY, FovX=FovX,
cx=cx, cy=cy,
image=image,
image_path=image_path,
image_name=image_name,
width=width,
image_path=image_path,
image_name=image_name,
depth=depth,
mask=mask,
width=width,
height=height)
cam_infos.append(cam_info)
sys.stdout.write('\n')
Expand Down Expand Up @@ -229,11 +235,22 @@ def readCamerasFromTransforms(path, transformsfile, white_background, extension=
image = Image.fromarray(np.array(arr*255.0, dtype=np.byte), "RGB")

fovy = focal2fov(fov2focal(fovx, image.size[0]), image.size[1])
FovY = fovy
FovY = fovy
FovX = fovx

cam_infos.append(CameraInfo(uid=idx, R=R, T=T, FovY=FovY, FovX=FovX, image=image,
image_path=image_path, image_name=image_name, width=image.size[0], height=image.size[1]))
# For NeRF synthetic data, we don't have cx, cy, depth, mask in the JSON
# Assume centered principal point and no depth/mask data
cx = 0.0
cy = 0.0
depth = None
mask = None

cam_infos.append(CameraInfo(uid=idx, R=R, T=T, FovY=FovY, FovX=FovX,
cx=cx, cy=cy,
image=image,
image_path=image_path, image_name=image_name,
depth=depth, mask=mask,
width=image.size[0], height=image.size[1]))

return cam_infos

Expand Down Expand Up @@ -311,11 +328,21 @@ def readMultiScale(path, white_background,split, only_highres=False):

fovx = focal2fov(meta["focal"][idx], image.size[0])
fovy = focal2fov(meta["focal"][idx], image.size[1])
FovY = fovy
FovY = fovy
FovX = fovx

cam_infos.append(CameraInfo(uid=idx, R=R, T=T, FovY=FovY, FovX=FovX, image=image,
image_path=image_path, image_name=image_name, width=image.size[0], height=image.size[1]))
# For LLFF data, assume centered principal point and no depth/mask data
cx = 0.0
cy = 0.0
depth = None
mask = None

cam_infos.append(CameraInfo(uid=idx, R=R, T=T, FovY=FovY, FovX=FovX,
cx=cx, cy=cy,
image=image,
image_path=image_path, image_name=image_name,
depth=depth, mask=mask,
width=image.size[0], height=image.size[1]))
return cam_infos


Expand Down
47 changes: 16 additions & 31 deletions train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,8 +69,9 @@ def create_offset_gt(image, offset):
id_coords = torch.from_numpy(id_coords).cuda()

id_coords = id_coords.permute(1, 2, 0) + offset
id_coords[..., 0] /= (width - 1)
id_coords[..., 1] /= (height - 1)
# Avoid division by zero for single-pixel dimensions
id_coords[..., 0] /= max(width - 1, 1)
id_coords[..., 1] /= max(height - 1, 1)
id_coords = id_coords * 2 - 1

image = torch.nn.functional.grid_sample(image[None], id_coords[None], align_corners=True, padding_mode="border")[0]
Expand Down Expand Up @@ -235,10 +236,9 @@ def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoi

opacity_loss = 0.0
if opt.lambda_opacity > 0:
# Get each gaussians' opacity and use cross entropy loss
# Get each gaussians' opacity and use entropy loss to encourage binary values
opacity = gaussians.get_opacity.clamp(1.0e-3, 1.0 - 1.0e-3)
opacity_loss = torch.nn.functional.binary_cross_entropy(opacity, opacity)
# opacity_loss = torch.mean(-opacity * torch.log(opacity + 1e-6))
opacity_loss = torch.mean(-opacity * torch.log(opacity + 1e-6))
Copy link

Copilot AI Nov 24, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The + 1e-6 epsilon in the log is redundant since opacity is already clamped to a minimum of 1.0e-3 on the previous line. The epsilon should either be removed from the log calculation, or the clamping bounds should be adjusted to use the same epsilon value for consistency (e.g., clamp to 1e-6 instead of 1.0e-3).

Suggested change
opacity_loss = torch.mean(-opacity * torch.log(opacity + 1e-6))
opacity_loss = torch.mean(-opacity * torch.log(opacity))

Copilot uses AI. Check for mistakes.
loss += opt.lambda_opacity * opacity_loss


Expand Down Expand Up @@ -272,7 +272,7 @@ def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoi
depth_loss_pseudo = depth_loss_func(gt_depth, render_depth)

if torch.isnan(depth_loss_pseudo).sum() == 0:
loss_scale = min((iteration - args.start_sample_pseudo) / 500., 1)
loss_scale = min((iteration - opt.start_sample_pseudo) / 500., 1)
loss += loss_scale * opt.lambda_pseudo_depth * depth_loss_pseudo
depth_loss += depth_loss_pseudo

Expand Down Expand Up @@ -459,28 +459,9 @@ def generate_idu_training_set(
n_avg=flow_edit_n_avg
)
elif use_difix3d:
refine_pipe = Difix3DRefineIDU(
save_path=refine_path,
device="cuda:0",
model_name=difix3d_model,
use_reference=difix3d_use_reference
)
final_imgs = refine_pipe.run(
imgs,
prompt=difix3d_prompt,
num_inference_steps=difix3d_steps,
timesteps=difix3d_timesteps,
guidance_scale=difix3d_guidance
)
raise NotImplementedError("Difix3D refine is not yet implemented. Please use FlowEdit refine instead.")
elif use_dreamscene:
refine_pipe = DreamSceneRefineIDU(
save_path=refine_path,
device="cuda:0",
model="sd21" if use_sd21 else "diffusionsat",
)
final_imgs = refine_pipe.run(
imgs,
)
raise NotImplementedError("DreamScene refine is not yet implemented. Please use FlowEdit refine instead.")
else:
raise NotImplementedError("DiffusionSat refine is deprecated")
if refine_pipe:
Expand Down Expand Up @@ -833,10 +814,9 @@ def training_idu_episode(

opacity_loss = 0.0
if opt.lambda_opacity > 0:
# Get each gaussians' opacity and use cross entropy loss
# Get each gaussians' opacity and use entropy loss to encourage binary values
opacity = gaussians.get_opacity.clamp(1.0e-3, 1.0 - 1.0e-3)
opacity_loss = torch.nn.functional.binary_cross_entropy(opacity, opacity)
# opacity_loss = torch.mean(-opacity * torch.log(opacity + 1e-6))
opacity_loss = torch.mean(-opacity * torch.log(opacity + 1e-6))
Copy link

Copilot AI Nov 24, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The + 1e-6 epsilon in the log is redundant since opacity is already clamped to a minimum of 1.0e-3 on the previous line. The epsilon should either be removed from the log calculation, or the clamping bounds should be adjusted to use the same epsilon value for consistency (e.g., clamp to 1e-6 instead of 1.0e-3).

Suggested change
opacity_loss = torch.mean(-opacity * torch.log(opacity + 1e-6))
opacity_loss = torch.mean(-opacity * torch.log(opacity))

Copilot uses AI. Check for mistakes.
if loss:
loss += opt.lambda_opacity * opacity_loss
else:
Expand Down Expand Up @@ -1027,7 +1007,12 @@ def colorize_depth_torch(depth_tensor, mask=None, normalize=True, cmap='Spectral
if normalize:
min_disp = np.nanquantile(disp, 0.01)
max_disp = np.nanquantile(disp, 0.99)
disp = (disp - min_disp) / (max_disp - min_disp)
# Avoid division by zero for constant depth
disp_range = max_disp - min_disp
if disp_range > 1e-6:
disp = (disp - min_disp) / disp_range
else:
disp = np.zeros_like(disp)

# Apply colormap
colored = plt.get_cmap(cmap)(1.0 - disp)
Expand Down
3 changes: 2 additions & 1 deletion utils/image_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,4 +16,5 @@ def mse(img1, img2):

def psnr(img1, img2):
mse = (((img1 - img2)) ** 2).view(img1.shape[0], -1).mean(1, keepdim=True)
return 20 * torch.log10(1.0 / torch.sqrt(mse))
# Add epsilon to avoid division by zero when images are identical
return 20 * torch.log10(1.0 / torch.sqrt(mse + 1e-10))
11 changes: 10 additions & 1 deletion utils/system_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,5 +24,14 @@ def mkdir_p(folder_path):
raise

def searchForMaxIteration(folder):
saved_iters = [int(fname.split("_")[-1]) for fname in os.listdir(folder)]
saved_iters = []
for fname in os.listdir(folder):
try:
# Try to parse the iteration number from the filename
saved_iters.append(int(fname.split("_")[-1]))
except (ValueError, IndexError):
# Skip files that don't match the expected pattern
continue
if not saved_iters:
return 0
return max(saved_iters)
Comment on lines 26 to 37
Copy link

Copilot AI Nov 24, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The error handling doesn't address the case where the folder doesn't exist. If folder doesn't exist, os.listdir(folder) will raise FileNotFoundError. Consider wrapping the entire function body in a try-except block to handle this case and return 0, or check if the folder exists using os.path.exists(folder) before calling os.listdir().

Copilot uses AI. Check for mistakes.
Loading