Environment
LTX Desktop: 1.2.7
OS: macOS Sequoia 15.3.1
Hardware: Mac Studio, Apple M1 Ultra, 128 GB Unified Memory
PyTorch: 2.11.0
mps_sdpa: 0.2.0
Generation: Local
Model: LTX-2.5 BF16
Issue
When generating a video locally using LTX Desktop 1.2.7 on Apple Silicon, the generated video is completely green.
This happens with both:
Text-to-Video (T2V)
Image-to-Video (I2V)
The output MP4 is successfully created and saved, but all frames appear as a solid green image.
What I observed
The model loads successfully and the denoising process completes without an apparent runtime failure.
The logs show:
Building transformer with attention backends -- self: MPS-SDPA, masked: MPS-SDPA
The denoising loop also completes successfully, for example:
Running denoising loop (8 steps, 512x288 121 frames 24 (@24) fps)
The video VAE/audio VAE are subsequently loaded and the MP4 is successfully written.
There is no obvious:
Traceback
RuntimeError
Exception
in the generation log.
I also tested MPS attention separately:
PyTorch MPS FP32 attention works
PyTorch MPS BF16 attention works
mps_sdpa.sdpa_opt() BF16 test works
No NaN/Inf was observed in these standalone tests
System memory usage also remained below the available 128 GB unified memory.
Suspected cause
The issue appears to be related to the LTX-2.5 Apple/MPS execution path, particularly the custom MPS-SDPA attention backend, rather than a model-loading or memory issue.
LTX's attention implementation automatically selects MPS-SDPA on macOS, and the transformer is reported as using:
self: MPS-SDPA
masked: MPS-SDPA
I have not yet confirmed that this is the root cause, but forcing the transformer to use the standard PyTorch attention implementation may be a useful diagnostic.
Expected behavior
The generated video should contain the requested scene/content normally.
Actual behavior
The generated MP4 is completely green.
Question
Could the maintainers confirm whether LTX-2.5/LTX Desktop 1.2.7 is currently supported on Apple Silicon/MPS for local generation?
Also, is there a recommended way to force the LTX transformer to use PyTorch SDPA / standard MPS attention instead of MPS-SDPA for testing?
I can provide the complete generation logs and additional diagnostic information if required.
LTX Desktop: 1.2.7
OS: macOS Sequoia 15.3.1
Hardware: Mac Studio, Apple M1 Ultra, 128 GB Unified Memory
PyTorch: 2.11.0
mps_sdpa: 0.2.0
Generation: Local
Model: LTX-2.5 BF16
Issue
When generating a video locally using LTX Desktop 1.2.7 on Apple Silicon, the generated video is completely green.
This happens with both:
Text-to-Video (T2V)
Image-to-Video (I2V)
The output MP4 is successfully created and saved, but all frames appear as a solid green image.
What I observed
The model loads successfully and the denoising process completes without an apparent runtime failure.
The logs show:
Building transformer with attention backends -- self: MPS-SDPA, masked: MPS-SDPA
The denoising loop also completes successfully, for example:
Running denoising loop (8 steps, 512x288 121 frames 24 (@24) fps)
The video VAE/audio VAE are subsequently loaded and the MP4 is successfully written.
There is no obvious:
Traceback
RuntimeError
Exception
in the generation log.
I also tested MPS attention separately:
PyTorch MPS FP32 attention works
PyTorch MPS BF16 attention works
mps_sdpa.sdpa_opt() BF16 test works
No NaN/Inf was observed in these standalone tests
System memory usage also remained below the available 128 GB unified memory.
Suspected cause
The issue appears to be related to the LTX-2.5 Apple/MPS execution path, particularly the custom MPS-SDPA attention backend, rather than a model-loading or memory issue.
LTX's attention implementation automatically selects MPS-SDPA on macOS, and the transformer is reported as using:
self: MPS-SDPA
masked: MPS-SDPA
I have not yet confirmed that this is the root cause, but forcing the transformer to use the standard PyTorch attention implementation may be a useful diagnostic.
Expected behavior
The generated video should contain the requested scene/content normally.
Actual behavior
The generated MP4 is completely green.
Question
Could the maintainers confirm whether LTX-2.5/LTX Desktop 1.2.7 is currently supported on Apple Silicon/MPS for local generation?
Also, is there a recommended way to force the LTX transformer to use PyTorch SDPA / standard MPS attention instead of MPS-SDPA for testing?
I can provide the complete generation logs and additional diagnostic information if required.