I didn't bother pixel-perfecting the nodes or lining them up neatly—as long as it’s clear what goes where and why, the rest doesn't really matter. Wasn't aiming for aesthetic perfection, and honestly, didn't have the time either! =)). Recommended page file size: at least 128 GB. Generating 5 seconds of 1920 by 1088 at 48 fps on 5060ti 16 + 64 ram takes about 500 seconds, 490-530 seconds of 1920 by 1088 at 24 fps on 5060ti 16 + 64 ram takes about 290-320 seconds =)) For optimal performance, NVMe speeds of 3000 MB/s to 3500 MB/s + . The best choice in terms of speed and quality | 0.5 | 16:9 | 1920 x 1088 | 9 | 3 |, perhaps it is still possible to optimize the process and finalize it, but there is no free time yet, it is even better to wait for the optimizations of the model and the associated nodes =))
v3.0 I2V MiniMax H3 + LTX 2.5
The upscaler has been ported to LTX 2.5
Certain settings have been changed and optimizations added to accelerate the overall process
Toggles are divided into groups, and options to quickly change step settings for upscale have been added
for optimal processing, tailored for smooth video or with a higher tendency toward dynamics
Settings for working with Lora for 4 steps have been factored in, with descriptions added for this specific Lora
An option to crop and fit images directly inside the workflow has been added "for the lazy", with the new Crop (OreX) cropping functionality, aspect ratio tests were conducted in various options, such as 34:9 and 9:34, and different combinations rounded up to 32 pixels; in all options, the result was good.
An example of an alternative model that works well in Ref2VA mode is provided
as well as a toggle between I2VA and Ref2VA.
Ref2VA uses a lightweight model that also performs quite well in I2VA
it was tested with up to 3 references (image + audio + video) and the model handled it quite well.
Description
The T2V update is aimed at optimizing performance, reducing resources, and improving workflow control. 4-step generation: High-quality results in just 4 steps of the basic Minimax H3 model, starting at 0.2 megapixels.LoRa Acceleration: Integrated to significantly increase generation speed. New VAE for video: replaced Minimax h3 with conVrot, which significantly reduced the amount of video memory and resource usage.T2V Settings: Optimized text-to-video conversion nodes to increase stability. Improved lip sync accuracy: Lip sync performance has been significantly improved, now it works well with 0.3+ megapixel resolution. The workflow and user interface have been improved: special control units and switches have been added for better modes. Real-time preview: A new preview mode with separate switches has been integrated.Documentation has been improved: Information blocks within the workflow have been expanded and updated.Node layouts have been optimized, which simplifies the use of the workflow.
FAQ
Comments (5)
Could we just swap the LTX2.3 upscaler for these ones?
https://huggingface.co/Lightricks/LTX-2.5/tree/main/latent_upscale_models
Thanks for the information, I'm not sure what will work fine, most likely it will require additional settings, I also assume that it may affect the generation speed, I will test how time will be....
They are exactly the same btw
Yes, I conducted quick tests, it is possible to use the version from LTX 2.5 upscale, although this does not give much advantage over the previous version, in the future build it will already be updated to 2.5, the sigma value will also be slightly, presumably it will differ from 2.3, of the disadvantages at the moment if you try to increase the resolution from an extremely low 0.1 - 0.2 megapixels in NSFW you can get anatomical distortions, this applies to T2V for high-quality I2V work, it already starts from a minimum of 0.4 megapixels, maybe 0.3 if the character's face is close to the camera.
@Nikolos747 Good to know, thanks for looking into it!