LTX-2.3 generation speed dropped significantly after upgrading to the new ComfyUI version

[wei01127]

** MOD NOTE: Please describe what disk and filesystem you are using if reporting this issue. **

I'm experiencing a massive slowdown with LTX-2.3 after updating to the latest ComfyUI Desktop version.

Previously on the older version, generating a 10-second video took around 8 minutes. However, after the update, the generation time has significantly increased to about 20 minutes for the exact same settings.

My Specs:

GPU: RTX 3060 (12GB VRAM)

RAM: 32GB

CPU: AMD Ryzen 7 5800X

I suspect it might be related to the new VRAM/RAM offloading behavior, CUDA cu130 compatibility issues, or VAE tiled decoding changes in the new version.

Has anyone else run into this issue, or does anyone know how to tweak the settings/arguments to get the old performance back? Any help would be greatly appreciated!

[marius-sucan]

Similarly for me as well. But with Wan animate. Generating a video of 15 seconds at 720 x 1280 at 30 fps took one hour. Now, it takes 1:30.

GPU is RTX 5060 TI with 16 GB of VRAM. Windows 10, latest Nvidia drivers.... System RAM: 64 GB. Intel 11th generation CPU.

[rattus128]

which part is slow? the steps? the first iteration? can we get the logs? If you have a screenshot or video of the memory visualizer we can see what it might be trying to do wrong or where its slow.

github.com/kijai/ComfyUI-MemoryVisualization

[wei01127]

@rattus128 I am using the official workflow LTX-2.3 I2V,Whenever the (SamplerCustomAdvanced) node is encountered, it gets stuck for a very long time; in the old version, this step ran very quickly.

[TxLysy]

Here are some additional findings from testing on AMD hardware showing a similar performance regression on Windows 10.(ComfyUI master branch, commit 2cdaaf).

• Test environment:
  [INFO] Total VRAM 16304 MB, total RAM 65231 MB
  [INFO] pytorch version: 2.12.0+rocm7.14.0a20260604
  [INFO] AMD arch: gfx1201
  [INFO] ROCm version: (7, 14)
  [INFO] Set vram state to: NORMAL_VRAM
  [INFO] Disabling smart memory management
  [INFO] Device: cuda:0 AMD Radeon RX 9070 : native
  [INFO] Using async weight offloading with 2 streams
  [INFO] Enabled pinned memory 26092.0
  [INFO] Using sage attention
  aimdo: src-win/cuda-detour.c:38:INFO:aimdo_setup_hooks: installing 6 hooks
  aimdo: src-win/cuda-detour.c:28:DEBUG:install_hook_entries: hooks successfully installed
  aimdo: src-win/shmem-detect.c:80:INFO:comfy-aimdo WDDM adapter match: AMD Radeon RX 9070 runtime_luid=00000000:000132dc dxgi_luid=00000000:000132dc
  aimdo: src/control.c:241:INFO:comfy-aimdo inited for GPU: AMD Radeon RX 9070 (VRAM: 16304 MB)
  [INFO] DynamicVRAM support detected and enabled
  [INFO] Python version: 3.13.0 (main, Oct 16 2024, 00:33:24) [MSC v.1929 64 bit (AMD64)]
  [INFO] ComfyUI version: 0.24.0
  [INFO] comfy-aimdo version: 0.4.9
  [INFO] comfy-kitchen version: 0.2.10
  [INFO] comfyui-frontend-package version: 1.45.15
  [INFO] comfyui-workflow-templates version: 0.9.98
  [INFO] comfyui-embedded-docs version: 0.5.2

• Test workflows
  ltx2.3 ia2v offical workflow
  1024*1024 25fps 5s

• Environment variables:
  export HIP_VISIBLE_DEVICES=0
  export TORCH_BLAS_PREFER_HIPBLASLT=1
  export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
  export USE_MEM_EFF_ATTENTION=1
  export COMFYUI_ENABLE_MIOPEN=1
  export MIOPEN_FIND_MODE=2
  export PYTORCH_MIOPEN_SUGGEST_NHWC=1

• Baseline (enabled pinned memory):
  uv run python ComfyUI/main.py
  --extra-model-paths-config extra_model_paths.yaml
  --listen
  --enable-triton-backend
  --use-sage-attention
  --auto-launch
  --enable-dynamic-vram
  --disable-smart-memory
  --fast-disk

• Treatment (disabled pinned memory):
  uv run python ComfyUI/main.py
  --extra-model-paths-config extra_model_paths.yaml
  --listen
  --enable-triton-backend
  --use-sage-attention
  --auto-launch
  --enable-dynamic-vram
  --disable-smart-memory
  --disable-pinned-memory
  --fast-disk

The result reflects the performance regression happened in first stage(8 steps) in ltx2.3(6.78s/it→10.67s/it), and that may be related to pinned memory being enabled or not, which was not an issue in previous versions.

[Crinale13]

I'm having a similar issue whereby I cannot complete previous workflows that worked fine since the update to v0.24.1 i keep getting OOM errors.

I use a simple text to image workflow (Flux1-dev-fp8) with two LORA's its a very simple workflow and i could generate images no problem at 1920x1080 or even higher on my 12GB RTX 3060 but now only 512x512 will complete on KSampler but it then fails on VAE Encode stage. Whats happened??

Is there anyway i can revert back? Perhaps i can keep this machine 'as is' and just never update it and build a newer machine if the system requirements are going to increase but i need my current machine to work s i cannot use anything now.

Though, oddly, WAN 2.2 seems to be completely uneffected and works exactly as it did.

Error log:

torch.OutOfMemoryError: Allocation on device
This error means you ran out of memory on your GPU.

TIPS: If the workflow worked before you might have accidentally set the batch_size to a large number.

MOD NOTE: log truncated. This report does not match the OP and was already pasted elsewhere.

[marius-sucan]

which part is slow? the steps? the first iteration? can we get the logs? If you have a screenshot or video of the memory visualizer we can see what it might be trying to do wrong or where its slow.

github.com/kijai/ComfyUI-MemoryVisualization

When comfyui starts and when the workflow begins running, I get no warnings and no errors. I use sage attention.

The Wan animate steps are running slower. Each sampler step is slower. I was using a bf16 safe tensor file. To regain the previous performance I had to switch to the equivalent Q8 gguf models.

On a similar note, Bernini, the official workflow, never finishes the first step with the suggested fp8 saf tensor files. To have it run, I had to use the - equivalent high and low gguf models.

Best regards, Marius.

[ooofest]

@rattus128 I am using the official workflow LTX-2.3 I2V,Whenever the (SamplerCustomAdvanced) node is encountered, it gets stuck for a very long time; in the old version, this step ran very quickly.

You're not alone with that symptom.

Prior runs of the official T2V and I2V workflow would get through SamplerCustomAdvanced nodes - i.e., the generation and upscale passes - in less than 10 total minutes on an RTX-3090.

Now it's taking over 20 minutes per sampler node to even start seeing a progress bar, then more time to fully complete the processing. Overall flows that took less than 20 minutes are now taking over 40 minutes.

EDIT:
I went back to a build from 2026-05-28 and also readjusted all the Python packages to match the environment at that time. It turns out xformers somehow got dropped during the recent update, but a number of other packages were also changed and many were added (presumably with other custom node updates).

Going back to that older environment snapshot saw the same LTX-2.3 T2V workflow which earlier took well over 30 minutes, now completes in 7 minutes for a 12 second video.

[rattus128]

I was able to build something very close to @wei01127 system and I cant reproduce a slow-down.

RTX3060 (12GBVRAM), 32GB RAM, ryzen 5600XT
PCIE 1.0 x4 NVME (bios downgraded) (disk unknown in report)

I dont have the video resolution so im assuming 1280x720x10s @25fps LTX2.3 two stage template with upscaler.

Anyways here is the full logs:

Current master (5aa71b9):

C:\Users\rattus\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable>.\python_embeded\python.exe -s ComfyUI\main.py --base-directory C:\Users\rattus\comfy-base
[INFO] setup plugin alembic.autogenerate.schemas
[INFO] setup plugin alembic.autogenerate.tables
[INFO] setup plugin alembic.autogenerate.types
[INFO] setup plugin alembic.autogenerate.constraints
[INFO] setup plugin alembic.autogenerate.defaults
[INFO] setup plugin alembic.autogenerate.comments
[INFO] Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': "ImportError: No module named 'triton'", 'capabilities': []}
[INFO] Found comfy_kitchen backend cuda: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[INFO] Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'dequantize_mxfp8', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'scaled_mm_mxfp8', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[INFO] Checkpoint files will always be loaded safely.
[INFO] Total VRAM 12287 MB, total RAM 32691 MB
[INFO] pytorch version: 2.11.0+cu130
[INFO] Set vram state to: NORMAL_VRAM
[INFO] Device: cuda:0 NVIDIA GeForce RTX 3060 : cudaMallocAsync
[INFO] Using async weight offloading with 2 streams
[INFO] Enabled pinned memory 13076.0
[INFO] Using pytorch attention
aimdo: src-win/cuda-detour.c:38:INFO:aimdo_setup_hooks: installing 6 hooks
aimdo: src-win/cuda-detour.c:28:DEBUG:install_hook_entries: hooks successfully installed
aimdo: src-win/shmem-detect.c:80:INFO:comfy-aimdo WDDM adapter match: NVIDIA GeForce RTX 3060 runtime_luid=00000000:0000a104 dxgi_luid=00000000:0000a104
aimdo: src/control.c:241:INFO:comfy-aimdo inited for GPU: NVIDIA GeForce RTX 3060 (VRAM: 12287 MB)
[INFO] DynamicVRAM support detected and enabled
[INFO] Python version: 3.13.12 (tags/v3.13.12:1cbe481, Feb  3 2026, 18:22:25) [MSC v.1944 64 bit (AMD64)]
[INFO] ComfyUI version: 0.24.0
[INFO] comfy-aimdo version: 0.4.8
[INFO] comfy-kitchen version: 0.2.10
[INFO] comfyui-frontend-package version: 1.45.15
[INFO] comfyui-workflow-templates version: 0.9.98
[INFO] comfyui-embedded-docs version: 0.5.2
[INFO] comfy-kitchen version: 0.2.10
[INFO] comfy-aimdo version: 0.4.8
[INFO] [Prompt Server] web root: C:\Users\rattus\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\comfyui_frontend_package\static
[INFO] Asset seeder disabled
[INFO] ComfyUI-GGUF: Allowing full torch compile
[INFO] ComfyUI-QuantOps: comfy-kitchen triton unavailable: ImportError: No module named 'triton'
[INFO] FP8 kernels: Triton not available, will use dequantize fallback
[INFO] ComfyUI-QuantOps: Registered TensorCoreMXFP8Layout
[INFO] ComfyUI-QuantOps: Registered layouts: ['BlockWiseINT8Layout', 'TensorWiseINT8Layout', 'RowWiseFP8Layout', 'BlockWiseFP8Layout', 'TensorCoreMXFP8Layout']
[INFO]
Import times for custom nodes:
[INFO]    0.0 seconds: C:\Users\rattus\comfy-base\custom_nodes\ComfyUI-MemoryVisualization
[INFO]    0.0 seconds: C:\Users\rattus\comfy-base\custom_nodes\ComfyUI-QuantOps
[INFO]    0.0 seconds: C:\Users\rattus\comfy-base\custom_nodes\ComfyUI-GGUF
[INFO]    0.3 seconds: C:\Users\rattus\comfy-base\custom_nodes\ComfyUI-LTXVideo
[INFO]
[INFO] Context impl SQLiteImpl.
[INFO] Will assume non-transactional DDL.
[INFO] Using RAM pressure cache.
[INFO] Starting server

[INFO] To see the GUI go to: http://127.0.0.1:8188
[INFO] got prompt
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
[INFO] Found quantization metadata version 1
[INFO] Detected mixed precision quantization
[INFO] Using mixed precision operations
[INFO] Native ops: int8_tensorwise, int8_blockwise, float8_e4m3fn_blockwise, hybrid_mxfp8, float8_e4m3fn_rowwise , emulated ops: nvfp4, float8_e4m3fn, float8_e5m2, mxfp8
[INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16
[INFO] model_type FLUX
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[WARNING] no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.
[INFO] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
[INFO] Requested to load LTXAVTEModel_
[INFO] Model LTXAVTEModel_ prepared for dynamic VRAM loading. 25440MB Staged. 0 patches attached. Force pre-loaded 400 weights: 1745 KB.
[INFO] Model LTXAVTEModel_ prepared for dynamic VRAM loading. 25440MB Staged. 0 patches attached. Force pre-loaded 400 weights: 1745 KB.
[INFO] Requested to load LTXAV
[INFO] Model LTXAV prepared for dynamic VRAM loading. 23835MB Staged. 1660 patches attached. Force pre-loaded 2104 weights: 3308 KB.
100%|████████████████████████████████████████████████████████████████████████████████████| 8/8 [03:02<00:00, 22.80s/it]
[INFO] 0 models unloaded.
[INFO] Model LTXAV prepared for dynamic VRAM loading. 23835MB Staged. 1660 patches attached. Force pre-loaded 2104 weights: 3308 KB.
100%|███████████████████████████████████████████████████████████████████████████████████| 3/3 [06:17<00:00, 125.99s/it]
[INFO] Requested to load AudioVAE
[INFO] loaded completely;  693.46 MB loaded, full load: True
[INFO] Requested to load VideoVAE
[INFO] 0 models unloaded.
[INFO] Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
[INFO] Prompt executed in 00:11:28

Comparing to v0.22.0 (2026-05-20):

C:\Users\rattus\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable>.\python_embeded\python.exe -s ComfyUI\main.py --base-directory C:\Users\rattus\comfy-base
setup plugin alembic.autogenerate.schemas
setup plugin alembic.autogenerate.tables
setup plugin alembic.autogenerate.types
setup plugin alembic.autogenerate.constraints
setup plugin alembic.autogenerate.defaults
setup plugin alembic.autogenerate.comments
Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'dequantize_mxfp8', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'scaled_mm_mxfp8', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
Found comfy_kitchen backend cuda: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': "ImportError: No module named 'triton'", 'capabilities': []}
Checkpoint files will always be loaded safely.
Total VRAM 12287 MB, total RAM 32691 MB
pytorch version: 2.11.0+cu130
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 3060 : cudaMallocAsync
Using async weight offloading with 2 streams
Enabled pinned memory 13076.0
Using pytorch attention
aimdo: src-win/cuda-detour.c:38:INFO:aimdo_setup_hooks: installing 6 hooks
aimdo: src-win/cuda-detour.c:28:DEBUG:install_hook_entries: hooks successfully installed
aimdo: src-win/shmem-detect.c:80:INFO:comfy-aimdo WDDM adapter match: NVIDIA GeForce RTX 3060 runtime_luid=00000000:0000a104 dxgi_luid=00000000:0000a104
aimdo: src/control.c:152:INFO:comfy-aimdo inited for GPU: NVIDIA GeForce RTX 3060 (VRAM: 12287 MB)
DynamicVRAM support detected and enabled
Python version: 3.13.12 (tags/v3.13.12:1cbe481, Feb  3 2026, 18:22:25) [MSC v.1944 64 bit (AMD64)]
ComfyUI version: 0.22.0
comfy-aimdo version: 0.3.0
comfy-kitchen version: 0.2.10
comfyui-frontend-package version: 1.45.15
comfyui-workflow-templates version: 0.9.98
comfyui-embedded-docs version: 0.5.2
comfy-kitchen version: 0.2.10
comfy-aimdo version: 0.3.0
[Prompt Server] web root: C:\Users\rattus\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\comfyui_frontend_package\static
Asset seeder disabled
ComfyUI-GGUF: Allowing full torch compile
ComfyUI-QuantOps: comfy-kitchen triton unavailable: ImportError: No module named 'triton'
FP8 kernels: Triton not available, will use dequantize fallback
ComfyUI-QuantOps: Registered TensorCoreMXFP8Layout
ComfyUI-QuantOps: Registered layouts: ['BlockWiseINT8Layout', 'TensorWiseINT8Layout', 'RowWiseFP8Layout', 'BlockWiseFP8Layout', 'TensorCoreMXFP8Layout']

Import times for custom nodes:
   0.0 seconds: C:\Users\rattus\comfy-base\custom_nodes\ComfyUI-MemoryVisualization
   0.0 seconds: C:\Users\rattus\comfy-base\custom_nodes\ComfyUI-QuantOps
   0.0 seconds: C:\Users\rattus\comfy-base\custom_nodes\ComfyUI-GGUF
   0.3 seconds: C:\Users\rattus\comfy-base\custom_nodes\ComfyUI-LTXVideo

Context impl SQLiteImpl.
Will assume non-transactional DDL.
Starting server

To see the GUI go to: http://127.0.0.1:8188
got prompt
VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
Found quantization metadata version 1
Detected mixed precision quantization
Using mixed precision operations
Native ops: int8_blockwise, int8_tensorwise, hybrid_mxfp8, float8_e4m3fn_blockwise, float8_e4m3fn_rowwise , emulated ops: mxfp8, float8_e4m3fn, float8_e5m2, nvfp4
model weight dtype torch.bfloat16, manual cast: torch.bfloat16
model_type FLUX
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load LTXAVTEModel_
Model LTXAVTEModel_ prepared for dynamic VRAM loading. 25440MB Staged. 0 patches attached. Force pre-loaded 290 weights: 1497 KB.
0 models unloaded.
Model LTXAVTEModel_ prepared for dynamic VRAM loading. 25440MB Staged. 0 patches attached. Force pre-loaded 290 weights: 1497 KB.
Requested to load LTXAV
Model LTXAV prepared for dynamic VRAM loading. 23838MB Staged. 1660 patches attached. Force pre-loaded 1496 weights: 44 KB.
100%|████████████████████████████████████████████████████████████████████████████████████| 8/8 [03:03<00:00, 22.89s/it]
0 models unloaded.
Model LTXAV prepared for dynamic VRAM loading. 23838MB Staged. 1660 patches attached. Force pre-loaded 1496 weights: 44 KB.
100%|███████████████████████████████████████████████████████████████████████████████████| 3/3 [05:57<00:00, 119.23s/it]
Requested to load AudioVAE
loaded completely;  693.46 MB loaded, full load: True
Requested to load VideoVAE
0 models unloaded.
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
Prompt executed in 00:12:31

Its a little faster with the latest stuff. @wei01127 if you can compare notes or send comparative logs that will help a lot.

[rattus128]

@ooofest 20 mins before seeing the step counter is very wrong, but I dont have a reproducer. What is your filesystem etc for that model?

To see where its slow, I made this:

#14371

If you can grab this change, run up to the delay and hit ctrl-C and paste I can see where you are frozen up.

[mjh02111964-netizen]

same here; ltx gens definitely decreased after latest update, but it seems to happen randomly at least for me. Generations that used to take 300-400 seconds are now well over 900-1000 seconds.

[rattus128]

@mjh02111964-netizen please see above for test data for comparisons, some possible follow up tests or post the usual logs.

[rattus128]

Can anyone reporting this please list your disk, filesystem, OS and anything form of remoteness of your filesystem (e.g. removable media or NAS etc