LTX-2.3_-_I2V_multi-subject-reference_Licon-MSR-lora ValueError: guide pre_filter_counts (1320) != keyframe grid mask length (1980)
The default workflow will rise "ValueError: guide pre_filter_counts (1320) != keyframe grid mask length (1980)", I found reconnect the nodes can run the workflow at least, please consider fix this.
Will check...
Was intending to update that workflow also, since there is a new lora out now, that works even better..
Yes please update the workflow to include the new LTX ingredient lora, which I suppose is something similar to the Licon MSR lora but hopefully even better since it is coming from lighttricks themselves.
Yes please update the workflow to include the new LTX ingredient lora, which I suppose is something similar to the Licon MSR lora but hopefully even better since it is coming from lighttricks themselves.
already on it.. they made it for single pass, and for 768×448, 121 frames, 24 fps (as training data, but that might be all ok)
So I am testing a few different runs, to see if it can work for other resolutions and longer videos as well.. and 2nd pass sampler for faster, to see if its still ok .. but will upload the wf in a not too long ;-)
(Also exploring ways to make the character sheet easily.. it requires a character sheet to work)
I just did a test of their ingredient workflow with one of their on ingredient sheet and it worked really well. the resolution they had set up on the workflow is 960x544 and at 241 frames. It came it out good. Now I just ran another test at 1280x 704 and it worked fine but it was a differnt generation. but it did have the same girl from their ingredient sheet and the yak.
Yes i noticed their own wf had 960x544, so it seems going away from the training data is ok.
Despite their warning:
out of scope: trained at a single resolution / length bucket (768×448, 121 frames, 24 fps); other resolutions, much longer clips, or use without a reference sheet are out of distribution
The "warning" is probably just there to let people know the ideal resolution, seems fine to use other settings.
Another thing they mentioned is that you have to use a static video of the ingredient sheet, but in their worflow they only have a load image node and not a load video node. So they are contradicting themselves in various things.
The biggest challenge is perhaps making the reference sheet.
I got a template for Ideogram, so I guess that can work for those that dont know how to make
Another thing they mentioned is that you have to use a static video of the ingredient sheet, but in their worflow they only have a load image node and not a load video node. So they are contradicting themselves in various things.
Yeah i was hessitating about that as well. It seems to just be an image batch (but encoded at video first, probably to get "video artifacts" that seems to help LTX models since version 0.9. As well as image compression for i2v). But a little strange, why they didnt just use the image compress also for the reference sheet, and batch that.
The biggest challenge is perhaps making the reference sheet.
I got a template for Ideogram, so I guess that can work for those that dont know how to make
I just made an ingredient sheet copying what they show in their examples. I used Klein 9b to create the individual images and photoshop to assemble it. I am running a test right now, but I didnt copy their prompt structure so I will see how it comes out. It didnt work. I guess it needs that very structured prompt as shown in their workflow