Better motion LoRA for LTX 2.3
Add more motion to your videos
Trained on T2V but supports I2V
720x1280 resolution with 15-30 steps
Weight 0.2-0.4 is recommended
⚠️ Avoid prompt bleeding using shorter sentences or words
Description
Initial first release.
FAQ
Comments (20)
30 steps? wtf
i use 4 to 6
I tend to run 15-20 steps depending on how well it actually handles the prompt.
I guess they're running on the dev checkpoint without the distilled lora (?)
@bennyboy_77 There's not really a way to run without distilled, but you can turn the strength down to 0.2 to better effect.
@boobkake22 what
@gambikules858 If you've got a confirmed good setting for doing a full gen at high quality without the distilled LoRA, I'd love to know it.
@boobkake22 this was possible since day 1, try the ltx 2.3 dev checkpoint with multimodal guiders, cfg 7 for audio and 3-4 for video, 30 steps and res2s. no distill lora for the first stage, str 1.0 for the upscaling stage. runexx has a dev workflow. this is how you actually unlock the motion amplitude since distillation takes it away. but it's waay slower with the multimodal guiders and res2s on top of that.
@JouzaDouza7135 Can you point to a specific workflow? I've been testing a multitude of single stage workflows, including the runexx versions, and I haven't seen a version that manages single stage no distill with a coherent result.
@boobkake22 what is probably meant by 30 steps is actually eg 30+6. i wouldn't personally recommend running a single stage on the dev version since ltx has no official workflow for that either, only for distilled.
https://github.com/Lightricks/ComfyUI-LTXVideo/issues/435 check out the last post
@JouzaDouza7135 That's what I was referring to. My best generations are single stage. The V2V upscaling seems to always entail a good amount of fine detail loss, and that's what I've been trying to solve. Single stage seems to have its own struggles, but I'm still exploring things. Really want an official word on this.
i use 60 for ltxv. Otherwise fast motions get blurry for me
Here is the workflow, it has both distilled and full: https://github.com/Lightricks/ComfyUI-LTXVideo/blob/master/example_workflows/2.3/LTX-2.3_T2V_I2V_Single_Stage_Distilled_Full.json
Dev 1.1 really is something else, 15 steps minimum beats anything distilled can do. It just takes so damn long, cant wait to see what future cards can do.
@jd666 there is no official 1.1 dev model, only distilled 1.1
I'm not entirely sure how this LoRA works, but I have it set to 0.20 with a LTX2.3 GGUF model, and it seems to have stopped those exaggerated movements and broken anatomy issues. Thank you!
No problem, this LoRA is best to use for the issues like you had.
I did like the other comment said. At strength 0.3, this lora helps a lot with anatomy. A lot better results than without.
That's good, it depends about what LoRA you use but 0.2-0.3 should be fine.
vrey good！❤️👍😊
Thank you, have fun!