Image 2 video - seems to fail 90% of the time.
Not sure if anyone has got i2v working..
This is so unreliable.. I hope someone might have an enhancement coming.
Probably 2-4 people out of 100 000 people trying LTX 2.0 workflows.
Great model, but its debut is a TOTAL DISASTER.
This needs to be fixed or it will end like LTX 1 - people will again switch back to crappy Wan 2.2
Works all the time here. So you must be doing something wrong
Maybe watch some tutorials, read up on the LTX promoting guide etc.
lol... ok if you say so
Works all the time here. So you must be doing something wrong
Maybe watch some tutorials, read up on the LTX promoting guide etc.
Its an issue of the textencoders. I am currently testing my fp8 uncensored variants and with the same prompt, seed etc. the image can be static..guess thats the same when using official gemma TE-saftensors....might be different if using the "pipeline" gemmaloader instead of the TE node.
I am using the official comfy workflow.. and get static videos almost all the time... but I will say text2video works nicely.. so go figure.
I am using the official comfy workflow.. and get static videos almost all the time... but I will say text2video works nicely.. so go figure.
comfy workflow means not the gemma loader=not the pipline! Try the worklow from here https://github.com/Lightricks/ComfyUI-LTXVideo/tree/master/example_workflows and for the gemmaloader https://huggingface.co/unsloth/gemma-3-12b-it-bnb-4bit/discussions
Update: also frozen images with gemmaloader very strange....some prompts working out of the box but fe when I change only a single phrase exp working prompt has "blue eyes" into "striking blue eyes" or "light blue eyes" the video is frozen/static.
ok, the workfow from lightricks is actually working ... so I have apologize.
LTX-2_I2V_Full_wLora.json
Only issue is that I have to reduce the resolution to around 800 x 600 to avoid Out of Memory errors.
But otherwise it is working.
Tests done here https://app.ltx.studio/ltx-2-playground/i2v showing that there is a problem with the comfy generation procedure or the fp8 model. Api model doesnt generate a single static video no matter what res the input image has had!
Probably something from the early adaption. I had this issue first day, but i downloaded a few model variants and workflows, and i never had this issue ever since.
As to what exact model or workflow, i dont know, will try figure it out
But everything works my end
( currently i use my own adapted version of LTX-Video workflow (LTX-Video nodes) which also seems to be the one that fixed it for the user in this thread, with KJ model loaders (KJNodes) )