*****USE NIGHTLY VERSION OF SEEDVR2 CUSTOM NODES!!!
MUST INSTALL MISSING CUSTOM NODES FOR WORKFLOW TO RUN!
Upscale Videos or Images at ANY resolution!
this is a LOWVRAM workflow for the implementation of the GGUF files for SeedVR2. workflow was tested using the 7b_Q8 GGUF model of seedvr2. there are many sizes and versions to choose from, i chose the 7b_Q8 GGUF because its only 8gb in size and isnt very vram intensive with tiled encoding so you dont have to worry about OOM errors during the encoding process.
Link to GGUF model files:
https://huggingface.co/cmeka/SeedVR2-GGUF/tree/main
run times are extremely long for video upscaling, 2+ hours per 5 seconds of video. i highly recommend setting batch size to 5 or lower. this maintains temporal consistancy between frames and also saves vram.
NO TRITON, NO SAGE ATTENTION! <3
Description
version 1 includes the base SeedVR2 workflow with the GGUF implementation. clean VRAM and clear cache nodes were added in place of the block swap and torch compile nodes to avoid triton and sage attention.