Update :
LTX 2.3/2.5 All in One : you can use for both model.
My Machine:
CPU: i5-12400
RAM: 64GB
GPU: RTX 3060 Ti 8GB
Time to Generate: (0.5MP_5s_24fps)
t2v
ltx 2.5 = 97 seconds
ltx 2.3 = 69 seconds
i2v
ltx 2.5 = 88 seconds
ltx 2.3 = 78 seconds
Minimax H3 : Workflow to FL2VA and Ref2VA
Upscaler Included (Nvidia RTX Video Super Resolution)
Interpolation also included (Rife VFI)
you can generate at 12/16fps (faster to generate) and interpolate to 24/32 fps
Time to Generate: (0.5MP/5s/16fps)
FL2VA
20 Steps with Sage Attn only = 350 seconds
20 Steps with Sage+Spectrum = 250 seconds
8 Steps with Sage+Turbo Lora = 180 seconds
Ref2VA
20 Steps with Sage Attn only = 370 seconds
20 Steps with Sage+Spectrum = 250 seconds
8 Steps with Sage+Turbo Lora = 170 seconds
Description
Minimax H3 Workflow to FL2VA and Ref2VA
Upscaler Included (Nvidia RTX Video Super Resolution)
Interpolation also included (Rife VFI)
you can generate at 12/16fps (faster to generate) and interpolate to 24/32 fps
My Machine:
CPU: i5-12400
RAM: 64GB
GPU: RTX 3060 Ti 8GB
Time to Generate: (0.5MP/5s/16fps)
FL2VA
20 Steps with Sage Attn only = 350 seconds
20 Steps with Sage+Spectrum = 250 seconds
8 Steps with Sage+Turbo Lora = 180 seconds
Ref2VA
20 Steps with Sage Attn only = 370 seconds
20 Steps with Sage+Spectrum = 250 seconds
8 Steps with Sage+Turbo Lora = 170 seconds
FAQ
Comments (2)
If you do 12 or 16 frames, the audio will run ahead of the video, so this method is not suitable. Interpolation won't even help.
16 fps is good for visual only, for audio it de-syncs.
Also, the workflow starts with "max" ref_image_size which slows down the generation which is worthy consideration, but otherwise it increases reference adherence