Reportedly 10 minutes on 4090 to upscale 5 sec video to 2k
This is Drozbay’s private fork of Impact Pack set of ComfyUI nodes, not merged into the original project:
GH:drozbay/ComfyUI-Impact-Pack.
Includes example workflows.
Allows selecting part of generated video like a vace or a complete character, upscale it to the resolution at which AI video generating models - those which take references; say WAN + Phantom - or Humo - work best - run generations, downscale and integrate result back into the original video.
Supports temporal blending across generated video fragments using VACE and context windows.
but flashvsr is not good quality wise, they made a lot of poor decisions that compromise quality for speed; tiny attention windows that prevent proper motion tracking so details swim; also the vae is worse but you can use the original vae instead, it’s just slower
Very fast, 1 step, but likely lesser quality than slow upscalers. Context options allow longer than 81 frames upscaling at the cost of minor glitches. No sageattn, use spda. Reducing strength can help with oversharpening. Upscaling from 720x512x720 was kind of ok.
Kijai:
it’s basically 4x upscaler, at least 2x or it’s bad results; would be fine if the input was tiny video, like their examples were 384x384, which is what the workflow used to test
Mads:
At least it does not try to resolve motion blur, like most upscalers does
Can use with Cinescale lora and possibly with RoPE Scaling
The new Flash VSR 1.1 upscaler is significantly better than v1;
Image Median Blursize=4;FlashVSR_Ultra_Fastmodel=FlashVSR-v1.1 mode=tiny scale=2; good afterFL RIFE Frame Interpolationrife49 from GH:filliptm/ComfyUI_Fill-Nodes multiplier=3; I useImage Median Blurto make results less sharp, I find FlashVSR to be too sharp
waiting for Kijai to update it cuz others not optimized for low vram
new vsr model works in kj wr but its bad
FlashVSR “ultra fast”: link
VSR safetensors loaded by WanVideo Extra Model Select plugged into extra_model input of WanVideo Model Loader loading Wan 2.1 T2V safetensors.
Very specific wording to be used in WanVideo TextEncode.
Can also do images, has been used on z-image-turbo produced images. …
2x spatial upscale only, works on images and video
something like LUA could be much better, but this is at least a huge improvement over bilinear/bislerp