Describe the bug
When generating 16:9 videos (1080x1920 approximately), the borders of the videos show degradation. As the video proceeds, it becomes worse and worse. Here is a sample:
https://streamable.com/50hhtw
This artifact does not appear using the base version of LTX 2.3 (or distilled). Oddly enough, I have encountered degradation in vertical videos within an older version of LTX.
Reproduction
- Model:
FastVideo/LTX-2.3-Distilled-Diffusers (Dreamverse)
- Prompt:
A red F1 car racing down a track in Monaco
- Seed:
10
- Orientation:
1088x1920
- Generated frames per segment:
121
- Frame rate:
24 fps
- Inference steps:
5
- Video conditioning overlap:
9 frames
- Audio conditioning overlap:
49 video-frame equivalents
- Audio conditioning: enabled
- Compile mode: TorchInductor with max autotune, FX graph cache, AOTAutograd cache, Triton cache, and CUDA cache
Environment
PyTorch version: 2.12.0+cu130
Is debug build: False
CUDA used to build PyTorch: 13.0
ROCM used to build PyTorch: N/A
OS: Ubuntu 24.04.4 LTS (x86_64)
Kernel: Linux 6.8.0-136-generic
GCC: 13.3.0
Clang: Could not collect
CMake: 4.4.2
glibc: 2.39
Python: 3.12.3 (64-bit)
Is CUDA available: True
CUDA runtime: 13.0.88
GPU 0: NVIDIA B200
NVIDIA driver: 580.173.02
XNNPACK available: True
CPU: 2 × Intel Xeon Platinum 8568Y+
Logical CPUs: 192
Physical cores: 96
NUMA nodes: 2
Relevant packages:
torch==2.12.0+cu130
torchvision==0.27.0+cu130
torchaudio==2.11.0+cu130
torchcodec==0.15.0+cu130
triton==3.7.0
numpy==2.5.2
transformers==5.15.0
nvidia-nccl-cu13==2.29.7
nvidia-cudnn-cu13==9.20.0.48
FastVideo Build Flags:
CUDA Archs: Not Set
ROCm: Disabled
Neuron: Disabled
Describe the bug
When generating 16:9 videos (1080x1920 approximately), the borders of the videos show degradation. As the video proceeds, it becomes worse and worse. Here is a sample:
https://streamable.com/50hhtw
This artifact does not appear using the base version of LTX 2.3 (or distilled). Oddly enough, I have encountered degradation in vertical videos within an older version of LTX.
Reproduction
FastVideo/LTX-2.3-Distilled-Diffusers(Dreamverse)A red F1 car racing down a track in Monaco101088x192012124 fps59frames49video-frame equivalentsEnvironment
PyTorch version: 2.12.0+cu130
Is debug build: False
CUDA used to build PyTorch: 13.0
ROCM used to build PyTorch: N/A
OS: Ubuntu 24.04.4 LTS (x86_64)
Kernel: Linux 6.8.0-136-generic
GCC: 13.3.0
Clang: Could not collect
CMake: 4.4.2
glibc: 2.39
Python: 3.12.3 (64-bit)
Is CUDA available: True
CUDA runtime: 13.0.88
GPU 0: NVIDIA B200
NVIDIA driver: 580.173.02
XNNPACK available: True
CPU: 2 × Intel Xeon Platinum 8568Y+
Logical CPUs: 192
Physical cores: 96
NUMA nodes: 2
Relevant packages:
torch==2.12.0+cu130
torchvision==0.27.0+cu130
torchaudio==2.11.0+cu130
torchcodec==0.15.0+cu130
triton==3.7.0
numpy==2.5.2
transformers==5.15.0
nvidia-nccl-cu13==2.29.7
nvidia-cudnn-cu13==9.20.0.48
FastVideo Build Flags:
CUDA Archs: Not Set
ROCm: Disabled
Neuron: Disabled