Don't Use LTX 2.5 For Music Videos! Stick To LTX 2.3
Is newer always better? Not this time.
If you are trying to create singing AI avatars, the older LTX 2.3 model delivers significantly better results than the new LTX 2.5. Based on my rigorous testing, including the latest update released just yesterday, the problem is still very much present. LTX 2.5 clearly struggles with its text encoder (Gemma 4) implementation. This causes major issues when trying to sync and process precise audio and lip movements for musical avatars. This video is for anyone doubting whether LTX 2.5 is actually usable for external audio. After extensive benchmark tests, I absolutely cannot recommend version 2.5 for singing avatars or music-focused workflows. Save yourself hours of frustrating experimentation that will only make you want to quit using LTX 2.5 altogether. Stick to LTX 2.3 for external audio, just like in this demo, and everything will work flawlessly. It would be a huge shame to give up on LTX 2.5 completely, though. The new 2.5 version is actually excellent and highly usable for many other use cases, just not for music video clips with external audio.
Rock on! 🤘 Cheers!
Try this at the end of your prompt: Audio: Audio1
Is newer always better? Not this time.
If you are trying to create singing AI avatars, the older LTX 2.3 model delivers significantly better results than the new LTX 2.5. Based on my rigorous testing, including the latest update released just yesterday, the problem is still very much present. LTX 2.5 clearly struggles with its text encoder (Gemma 4) implementation. This causes major issues when trying to sync and process precise audio and lip movements for musical avatars. This video is for anyone doubting whether LTX 2.5 is actually usable for external audio. After extensive benchmark tests, I absolutely cannot recommend version 2.5 for singing avatars or music-focused workflows. Save yourself hours of frustrating experimentation that will only make you want to quit using LTX 2.5 altogether. Stick to LTX 2.3 for external audio, just like in this demo, and everything will work flawlessly. It would be a huge shame to give up on LTX 2.5 completely, though. The new 2.5 version is actually excellent and highly usable for many other use cases, just not for music video clips with external audio.
Rock on! 🤘 Cheers!
i have done many videos with 2.5 and lip synch is perfect, BUT every once in a while ONE scene will mess you, yes, like you said..... chat gpt says its because of the GEMMA MODEL, INT8 and to use the BF16...... which gemma are you using?
at least my animacion came out nice with 2.5 . but yes SOME SCENES it will break,,,,
After last night’s nightly update, LTX 2.5 works with custom external audio, but it doesn’t always work consistently. I tested an older version of the text encoder projection model, and it seems to work better in some cases. Instead of the native ComfyUI workflow, use the new A2V workflow that you can find in the Custom Nodes folder after updating LTX 2.5. You can still include any custom LoRA you need. This way, LTX 2.5 should work better, with less jitter and less overreacting to external audio. Video samples will be posted on my YT channel ComfyUI Playground and on another dedicated channel specifically focused on AI music and AI avatars: https://www.youtube.com/@MusicStageAI/videos
If this tip helps you, don’t forget to subscribe for more.
Cheers!