LTX2 - Resolution - Quality
This version is quite honestly the fastest video generation i have used locally (3090), its quite unbelievable tbh....but generating anything above 1280x704 in I2V is not recommended as the image either duplicates or disintegrates. I would love to be able to do 1920x1080 I2V. Great work i hope you support and build on it further since its very very promising. I even tried doing the recommended 4K resolutions just to test it out but the image was distorted and destroyed.
Face consistency is an issue for I2V.
I disagree, it's 3-4 times as slow as the LTX team's workflow, because of their native custom nodes.
It's better to use the native LTX's worflow and plug the NSFW loras into the WF. Much better results and very much faster.
For the record, my workflow does include native LTX nodes and KJNodes which provide additional optimizations. Some of the VRAM optimization nodes may slightly decrease speed, but I wouldn't expect a 3-4x slowdown. However, as with all things in AI, your mileage may vary and you are free to use whatever workflow fits your use cases the best.
You may be able to do some nsfw at 16:9 or even 4:3 ratio 1080p, you mostly have to watch the nsfw loras when you go around 1200 pixels vertical or beyond.
For me, it takes about 2 minutes to generate a 5‑second 1024x1400 video on a 16 GB VRAM GPU.
I’m using a combination of Phr00t’s and Lightricks’ workflows, and I simply removed everything I didn’t need.
The pipeline is basically:
Phr00t models → Lightricks upscaling → RIFE interpolation → Save.
These are great models and an excellent workflow to start with. It would have taken me much longer to figure out how everything works without them.
Anyone can customize the workflow to fit their own needs, so it feels a bit unfair to complain about it.
Thank you guys for the support.