Workflow: V2V - Remove watermark & subtitles with ICEdit-Insight lora
V2V - Remove watermark & subtitles with ICEdit-Insight lora
The subtitle and watermark removal loras are 2 of many loras from Valiant Cat/JoyFox Ai.
Get the loras from here: https://huggingface.co/joyfox/LTX2.3-ICEdit-Insight
If you want to try it out, workflows here:
https://huggingface.co/RuneXX/LTX-2.3-Workflows/tree/main/Video-2-Video
(will make a workflow for the upscale and refine loras from them also, although you can pretty much use same wf)
(Joyfox also has workflows here https://github.com/Valiant-Cat/LTX2-ICEdit-Insight)
@GuardSkill https://huggingface.co/joyfox
Thanks as always for these. I am wondering, do you think there's any way to reduce the motion smearing with V2V or just have to wait until LTX 2.5 which hopefully will not have that problem?
Thanks as always for these. I am wondering, do you think there's any way to reduce the motion smearing with V2V or just have to wait until LTX 2.5 which hopefully will not have that problem?
Probably, for now at least, just add more steps. The distilled default with 8 step is a perhaps a bit "optimistic" for fast motion. Its nice and fast for normal close up videos, but for more demanding ones, you might want to use a bit more steps. Under the sigma node (in most all of my workflow), is a "hidden" node that you can connect instead of the hardcoded 8 step sigmas to the sampler.
And set say 10-15 steps or so.
And if you want to take it even further, use the dev model, with distilled lora set low. And set CFG to 3.0 or so, and 20-30 steps at first sampler
Haven't tested too many fast motions one myself, but the above should probably help
Thanks for the tips! Increasing to 15 steps helps a lot. I'll see what else can be improved with higher CFG too.
This wf works perfectly with joyfox's video restoration lora also.
This wf works perfectly with joyfox's video restoration lora also.
nice nice ;-) i didnt try that lora yet, been meaning to...
Have you also tried their full Edit model also? I got a mismatch model error while loading that one.
Have you also tried their full Edit model also? I got a mismatch model error while loading that one.
Thats probably a checkpoint (all in one model with everything baked in).
Basically the same as using the Kijai's split models + the lora
You can replace the model loaders in the workflow here, with the checkpoint loader way.
If you download the "I2V T2V Basic Checkpoint version" workflow here https://huggingface.co/RuneXX/LTX-2.3-Workflows/tree/main you will see how the checkpoint loaders should be setup.
Its a little different since the vae, etc is baked into the model (checkpoint), but not a lot ... its just a few node changes
I'll try their full model soon though, see how it works, and maybe make its own wf