The details of this LoRa are quite impressive. Could you consider making it into an enlarged model for use,
#8
by liuxingsyue - opened
Train this LoRa as an upscaling model, similar to the Minimax H3 Latent Upscaler, specifically for secondary upscaling output in video applications, and compute the output using a block node mode. This will enhance its practical value. I tested using this LoRa to add details to a 10-second 1080p video, and it immediately stalled. I have a graphics card with 16GB of VRAM,
Just load it specifically in your upscale pass (keep it away from your main pass) using a tile based based upscaler like H3UltimateUpscale.
It takes so long to refine a video even with 3-4 steps lora.
I have 24gb vram and It's impossible to do any clip that has > 5 seconds length, > 0.5Megapixels.
Upscale model does seem to be a better approach