Innovate Futures @ Benji

(For Patreon Supporters) Hunyuan Custom subject2vid and LTX Upscaler

Added 2025-05-13 19:11:22 +0000 UTC

Hunyuan Custom Post : https://www.patreon.com/posts/hunyuan-custom-128857913

Video about Hunyuan Custom : https://youtu.be/4MwxgtZztOU

LTX Video 13B Post: https://www.patreon.com/posts/ltx-video-13b-in-128627105

Video about LTX Video 13B : Video : https://youtu.be/2asTwiDn0y0

------------------------------------------------------------------------

Previously, we used the SDXL tile upscaling method for both images and videos. However, when working with AI-generated videos—especially those based on DiT (Diffusion Transformer) models—using SDXL upscaling can sometimes result in a loss of frame speed or flickering, even when the denoise level is set very low.

The recent release of LTX Video 13B 0.9.7 offers a promising alternative as a video enhancer. Rather than using LTX to generate the initial video sample, we can instead utilize its STG (Spatial-Temporal Generation) and upscaling capabilities as an effective way to upscale low-resolution AI video chips—from 480px to 1280px, for example.

I followed the latest custom workflows shared in Hunyuan’s newly posted tutorial and integrated them with the LTX 13B upscaler. I’ve built this into a modular group so that it can accept frames from outputs of various AI video generation tools, making it flexible for different pipelines. Then, this upscaling method can be applied consistently.

Because LTX 13B is a dedicated video model, it differs significantly from SDXL upscaling, which processes each frame individually as if they were still images. While the SDXL method is lightweight, it can introduce minor flickering due to inconsistencies between frames. In contrast, LTX performs upscaling as a holistic video process, maintaining coherence across all frames and ensuring that objects remain stable throughout the sequence.

---------------------------------------------------------------------------------------------------

It is better to use spatial rather than temporal.

---------------------------------------------------------------------------------------------------

2-4 minutes for typical 5 or 10 seconds AI video upscaling. Way faster than SDXL Upscaling frame by frame.

------------------------------------------------------------------------------------------------------

You might caught into error like this one if the frame number is not right for LTX Upscaler.

Because LTX model train with a specific number of frame to be process. Default setting is 129 frames. So fix your frame number before you pass into LTX Upscaler.

You can always check the frame length by going to this node click the length and it will give you the range of number that pass it's encoding frames. e.g: 129 , 137 , 145,... so on...

------------------------------------------------------------------------------------------------------