Step 5: Load image/video and select Model
You can pick which model you perfer.
3b, 7b, 7b_sharp. All of them have fp8, fp16 and Q4/8 gguf version.
To make everytime simple, 3b is the smaller size model and 7b_sharp is the sharper version of 7b. The default model for video upscale is 3b_fp8.
The model will be download automatically when you first run the workflow.
[15:12:39.137]
[15:12:39.138] ╔══════════════════════════════════════════════════════════╗
[15:12:39.138] ║ ███████ ███████ ███████ ██████ ██ ██ ██████ ███████ ║
[15:12:39.138] ║ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ║
[15:12:39.138] ║ ███████ █████ █████ ██ ██ ██ ██ ██████ █████ ║
[15:12:39.138] ║ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ║
[15:12:39.138] ║ ███████ ███████ ███████ ██████ ████ ██ ██ ███████ ║
[15:12:39.138] ║ © ByteDance Seed · NumZ · AInVFX ║
[15:12:39.138] ╚══════════════════════════════════════════════════════════╝
[15:12:39.138]
[15:12:39.139] 📥 Downloading D:\ComfyUI_SeedVR2\ComfyUI\models\SEEDVR2\seedvr2_ema_3b-Q8_0.gguf from https://huggingface.co/AInVFX/SeedVR2_comfyUI/resolve/main/seedvr2_ema_3b-Q8_0.gguf...
seedvr2_ema_3b-Q8_0.gguf: 100%|████████████████████████| 3.41G/3.41G [00:56<00:00, 64.6MB/s]
[15:13:38.767] ✅ Downloaded and validated: seedvr2_ema_3b-Q8_0.gguf
[15:13:38.768] 📥 Downloading D:\ComfyUI_SeedVR2\ComfyUI\models\SEEDVR2\ema_vae_fp16.safetensors from https://huggingface.co/numz/SeedVR2_comfyUI/resolve/main/ema_vae_fp16.safetensors...
ema_vae_fp16.safetensors: 100%|██████████████████████████| 478M/478M [00:07<00:00, 64.9MB/s]
[15:13:47.133] ✅ Downloaded and validated: ema_vae_fp16.safetensors
[15:13:47.133] 🏃 Creating new runner: DiT=seedvr2_ema_3b-Q8_0.gguf, VAE=ema_vae_fp16.safetensors
[15:13:47.148] 🚀 Creating DiT model structure on meta device
[15:13:47.212] 🎨 Creating VAE model structure on meta device
[15:13:47.402]
Note that the new Version 2.5.0 
that just release today has make lots of change to the perviouse version. Old workflow won’t work on current version without modification.
They have split the node to 4 nodes.
(DiT model, VAE model, torch.compile settings, and main upscaler for granular control)
