Optimized workflow for running LTX 2.3 (22B Distilled) on 8GB VRAM GPUs (e.g., RTX 3070 Laptop), focusing on stability, memory efficiency, and high-quality video/audio output.
Description
FAQ
Comments (34)
Hello. I have just used your workflow to create about 10 videos, and it works very well. Thank you for creating such a great workflow.
How to open this file?
@akshitdhoundiyal651
1. Launch Comfyui.
2. Comfyui finishes launching.
3. Drag and drop the .json file provided by the user who created this workflow into the internet window where Comfyui is running (with Comfyui fully launched).
4. The workflow file opens.
5. Done.
what gpu and how long does a video take to finish?
@robotfromspace I am too busy with many tasks right now to remember clearly. However, based on my uncertain memory, the video generation speed was much faster than the LTX 2.3 template version provided by ComfyUI, and PC RAM usage seemed to be reduced by almost half. I am using an RTX 4080 Super graphics card and 64GB of PC RAM.
I tried your workflow and downloaded all the resorces but getting this error:
CLIPTextEncode
'Linear' object has no attribute 'weight'
Do you know how to fix this?
Upgrade all
@gio83dj Ty, I'm new to LTX workflows and I'm using a RTX 2070 so I don't know if it even should work. After the update all I got a new error after 30 minutes of rendering. I hope I can fix it.
Having the same issue. Comfyui is updated. Any ideas would be great.
I ran into several errors too. I dumped my comfyui log file and the workflow json into grok and asked it for help. It was a lot of updating comfyui and python and safetensors to get it further along.
@john05636583 I've come to the conclusion that its a bug within comfyui. Reddit has been bitching about the latest updates. I'm also having issues with other workflows that I built. I expect when that gets sorted out, this workflow will work like a champ. @robotfromspace Thanks for your work.
@dreadfulpirate "Reddit has been bitching about..." you could just stop there. That place is a dumpster fire of rage all the time. It's working for me pretty well now. Still takes forever compared to the good ol'e days of Grok Imagine for NSFW prompted videos. Let me know if there's a way for me to send my workflow json to you and I'll do it.
@john05636583 Thanks for the offer. I gave up and reinstalled Comfyui. Fresh install cleaned things up. LTX Vids are working fine now. This workflow works nicely.
What should be changed for a 5090 in your opinion ?
how to fix this issue?
RuntimeError: ERROR: VAE is invalid: None
If the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.
File "C:\Users\Admin\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 525, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Admin\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 334, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "F:\ComfyUI\custom_nodes\comfyui-lora-manager\py\metadata_collector\metadata_hook.py", line 168, in async_map_node_over_list_with_metadata
results = await original_map_node_over_list(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Admin\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 308, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "C:\Users\Admin\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 296, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "F:\ComfyUI\custom_nodes\ComfyUI-KJNodes-main\nodes\nodes.py", line 2458, in load_vae
vae.throw_exception_if_invalid()
File "C:\Users\Admin\AppData\Local\Programs\ComfyUI\resources\ComfyUI\comfy\sd.py", line 847, in throw_exception_if_invalid
raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.")
update the ComfyUI-KJNodes node
This is exactly what I was looking for thank you! Any idea on I2V version, or adding loras?
I'm very new to this and I have everything set up but I dont understand this part:
"Performance Optimization
Recommended Settings
Install xFormers for improved memory efficiency and speed
Use Sage Attention with:
fp16_triton
Enable compile
Torch Settings
Enable:
fp16 accumulation
Helps reduce VRAM usage with minimal quality loss"
How do I install this stuff, is it a custom node or do I have to do it manually in the venv somehow?
Also How do I run it, do you add some flags on the .sh-file like with --low-vram? (should I run it with low vram?)
Hey this works, I have been trying to get ltx2 or 2.2 to work for months on 16gb amd card, but this flow is the first one that actually works.
none of this stuff works on AMD cards. You might as well be running with no card.
I can't seem to find where to set the # of steps. Running the workflow shows 8 steps. Where can I set this? Thanks.
in manual sigmas
hi i cant seem to find custom node SaveLatentWithPurge
Hey, excellent workflow, thank you so much!.
Will you also be working on an I2V?
Why are you creating the T2V format? It's completely useless.
I couldn't find a better fit for my 12gb vram -32gb ram pc. Your workflow is great.
Please make an I2V workflow too!
Running on 3060 . But seems a hard road. I can't run with sage attention or Triton.
I try install that Microsoft Vstudio and cuda toolkit cuz of that node, anyone can give light how get this to work ?
Even gpt or Claude give me I mess the final result booth say is turn it off.
I use portable comfyui if that metter
That's a long hard road you're walking. Depending on version of torch, drivers, etc. you might need to compile with different versions of Visual Studio.
Hey folks, I noticed that a lot of people wanted this workflow to have I2V and LoRA feature support, but the author seemed to have abandoned the project. Don't worry, I made a fork and added those features myself, and I want to share that workflow with you!
Here's the new and improved workflow: https://pastebin.com/2SQfS8Sk
Here's an example of the I2V feature: https://civitai.com/posts/29357577
A quick explanation of how it works: if you bypass the node above the "Image Switch" or "Latent Switch" nodes, it will take priority over the other node.
- Example size: If you un-bypass the "Empty Image" node, it will take priority over the image size proportion, and you can set the size manually. (You don't have to bypass the image group).
- Example I2V or T2V: Similar concept with the size switcher. If you enable the "T2V Base Latent" node, it will take priority over the I2V node LTXVImgToVideo).
Changes I've made:
- Added support for I2V while keeping T2V support!
- Added LoRA support!
- Added VRAM and RAM clean-ups
- Added a video interpolation group RIFE VFI) to optionally double the FPS and make the video smoother.
- Switched the video VAE decode to VAEDecodeTiled to heavily reduce VRAM spikes during the final step.
- Added a ColorMatch node for I2V so the generated video perfectly matches the color grading of your input image.
- Made every node feel more organized, it was triggering my OCD.
Known issues:
- If you use a low scale in the ImageScaleBy node, it tends to not replicate the person in the image as a 1:1 match. To fix this, simply use a higher scale (0.7 to 0.8). Note: Higher scale = better detail but slower generation.
- Using I2V and any LoRAs may take 1 or 2 minutes longer than usual to create a video. I'm not sure if I can fix this... I don't think it's possible to "fix" this.
- The video may feel sluggish sometimes; it's very dependent on the prompt and the seed you use.
- If you decide to use a CFG higher than 1.0 (to make the negative prompt stronger), expect slightly longer generation times per video.
P.S. Thanks and have fun!
Ohh FFS, i just notice that some nodes are 1 pixel off...
thanks for the workflow, does it use ssd for offloading ? i dont like using ssd for loading models, i like to cook ram and gpu only while the ssd is chilling
@cabsluke2022942 You can safely remove them, i later realise that they are completely useless to this specific workflow but i'm to lazy to update the workflow...
I got rtx5070 with 12gb vram (10.6gb avail) and 32gb RAM. I got the newest comfyui (july2026), i got pytorch 12.9 (also tried 13.3), and sage attention + xformers. I got cuda 12.9 installed. (also tried with cuda 13.3). All models from WorkFlow downloaded, and I set all the suggested low ram options like fp16_trition + compute + torch_fp16_accu. I use an empty FF or LibreWolf browser with only ComfyUI open. (I even tried with dev Q2 model and it still OOMs)
Yet I cannot make a 4 second clip in 480x480, I get an OOM with such low settings, so how do you guys get the 8gb VRAM cards to make a 20 sec video with this workflow?
My python settings are also targeting low ram as much as possible: python3 main.py --preview-method taesd --bf16-vae --lowvram --async-offload --use-sage-attention
What version of ComfyUI does this work with? What version of torch with cuda (12.8, 12.9, 13.0, 13.3 ??) I have most recent xformers and sage attention.
I can however make 5-10-15 second clips with ComfyUI's workflow using LTX2.3 Dev, with larger models loaded than what this WF is using, so why do this OOM all the time? Can anyone help! I'd love to do GGUF and not the much larger SafeTensor files.