AttributeError: 'NoneType' object has no attribute 'endswith'
Hello everyone, I am using LTX-2_I2V_T2V_Simple_single_pass.json, and the following node issues are displayed when opening the workflow:
Then I replaced it with the DualCLIPLoader (GGUF) node below, and the workflow no longer showed an error, but an error was prompted again during running it:
Of the 5 I2V workflows, only LTX-2_I2V_T2V_Basic.json has no node issues and runs perfectly. LTX-2_I2V_T2V_Basic_GGUF.json shows the same problem, but it runs perfectly after I replaced it. I haven't tested the other 2 I2V workflows yet.
Has anyone encountered the same problem, or did I do something wrong?
My configuration:
RTX 5070 12GB
Python 3.10
torch 2.9.1+cu130
ComfyUI 0.15.1
ComfyUI-GGUF 1.1.10
ComfyUI-KJNodes 1.3.2
The models are from Kijai and unsloth according to the workflow json files.
Its because you the Gemma GGUF model you have is not compatible with the prompt enhancer part.
Simply disable that part ;-)
I'll make it easier to disable that, but think you might already have a toggle in the workflow, if you look at the bottom row
Thank you. That toggle node seems very clever. I have successfully run both branches, and they work perfectly.
And I noticed:
- You used LTXV-13B-0.9.8-dev-VAE.safetensors. I replaced it with LTX2_video_vae_bf16.safetensors, and it runs successfully. But I wonder why you used LTXV-13B-0.9.8-dev-VAE.safetensors. Is it because it works better?
- The sampler you used is lcm. I tried to replace it with euler and dpmpp_2m. The effect of euler seems similar, but dpmpp_2m is obviously not good. I want to know why you chose lcm?
- There is another workflow using ksampler that I am going to try. Among basic, basic_gguf, single_pass, and single_pass_ksampler, which one do you think works best? I have used the first three, and I feel there is not much difference in their effects.
Thanks again!
Yes, its the LTX2_video_vae_bf16.safetensors for sure. ITs so many workflows so that must have been a slip ;-)
You can try other samplers. Euler and LCM are usually safe choices. The default from comfy is euler at 1st pass, and gradient_gradient estimation and 2nd pass (and LTX team euler at 1st pass, res_2s at 2nd pass)
I tend to think it burns the image a bit, so i usually use LCM or Euler also at 2nd pass (but the result is a little softer)There is little difference. The single pass was more to cater to some that prefer it a bit "Wan way" with just one sampler (same with K-Sampler, old habits die hard)
Depending on your output size, the 2pass workflows might run faster/easier on the computer though, since 1st pass is half size.
But yes, all in all, some of the workflows are very similar ;-) For LTX2.3 (came out yesterday) I will do far less.... less is more ;-)