I have a question regarding the workflow "LTX-2.3_-_FML2V_First_Middle_Last_Frame_guider"
Hi, first, I very much apreciate you share your warkflow! I only wish it could work for me. See, I I try to run it as is and run into an error. It says: "The size of tensor a (1360) must match the size of tensor b (206336) at non-singleton dimension 2". It gives me the same error with full model. Maybe you could help me with that?
Hmm little unsure what that could be.
Try this, as its very important to have latest (for supported features)
- Update KJNodes to the very latest version
- Updated ComfyUI to the very latest version
Can you see at what node it fails? usually it turns red-ish where it fails.
It could also be a mismatch with the frames count or the size of the input images (but that should auto resize in the workflow)
LTX-2 requires width & height to be divisible by 32 .Frame Count: Must be divisible by 8 (+1 .. add one frame)
(but this part should be automatically adjusted in the workflow)
Yeah, I updated comfy and nodes. It fails at SamplerCustomAdvanced node
Yeah, I updated comfy and nodes. It fails at SamplerCustomAdvanced node
check if the model loaders are correct also. That audio vae and video vae is not swapped
No, it is not swapped. I actually receive the same error from other ltx workflows, including the one in comfyui templates...
No, it is not swapped. I actually receive the same error from other ltx workflows, including the one in comfyui templates...
Could be something with your install .. perhaps some custom node that breaks things.
But will try here, sometimes ComfyUI breaks themselves also, but they are usually very quick at fixing if so
btw, where did you download the models from? maybe its a combo that doesnt work.
Try use all the models from here: https://huggingface.co/Kijai/LTX2.3_comfy
- the upscaler model from official ltx repro.
And if you are using GGUF try the ones from Quantstack or Unsloth
I believe I downloaded it from Kijai's hugginface
Strange. Should work fine.
I'm a little unsure what that python message could be. Check if the Clip encoder loader is set to LTX.
And if you used LTX in the past, make sure all is LTX2.3, and not some VAE or other part that is from LTX2.0
(and if using any lora, try without the lora)
Doing a little google search, some report that ComfyUI_smZNodes makes such errors. If you have this node, try remove it (easy to add back in later)
And try run the ComfyUI Manager "Update All" commands so all is up to date.
https://github.com/Comfy-Org/ComfyUI/issues/11653
You were right! It was the smZNodes after all. Even though I wasn't using them in workflow, they somehow hijack the KSampler logic globally, causing a tensor size mismatch error with the LTX-Video model.
I physically deleted them from the custom_nodes folder and it works now