SeedVR2 Workflows

This includes both my modified version of the numz workflow from his custom nodepack, as well as a workflow I made from scratch after SeedVR2 received native support for ComfyUI. For regular image upscaling, I recommend using the native workflow since it's able to skip the vae encode step when testing different settings.

Video upscaling has been tested and working for both workflows. I still recommend video upscaling with the custom nodes if you want the highest possible resolution. I usually upscale videos that are a maximum of 20s to 2k resolution, and with the native workflow I tend to OOM because of the way the native workflow utilizes dynamic VRAM loading/offloading as opposed to the traditional block-swap method. The native workflow probably is faster when it comes to video upscaling, but nothing sucks more than getting to the very end of the upscale process and getting an OOM error. To be fair, this will inevitably happen on both workflows as you experiment, but I've been able to push the limits farther with the custom nodes workflow.

If you're not happy with the upscale results, I recommend experimenting with the latent_noise_scale (Anything from 0.01-0.15 can bring meaningful change to fine details). The input_noise_scale can also be helpful at times, but is more subtle. Another practical way to get better results is by either a) Downscaling the image to a certain resolution or b) adjusting the final output resolution to something like 2000px or 3000px.

Link to int8_convrot model: https://huggingface.co/Comfy-Org/SeedVR2/blob/main/diffusion_models/seedvr2_7b_int8_convrot.safetensors

This workflow is compatible with the 3b and sharp variants of the model as well.