SeedVR2 Video Upscaler (v2.5.24)

This is the business end of SeedVR2 - the node that takes your frames, the model, and the VAE, and gives you back the upscaled result. SeedVR2 is ByteDance's one-step video restoration model that the ComfyUI crowd adopted as the best "add real detail" upscaler going: Apache 2.0, no non-commercial rider, and it preserves the picture instead of reinventing it the way SUPIR does. This node is the current v2.5 main node - despite the name it's what most people run on single images, because the video path really does want a datacenter GPU.

The important thing to understand up front: this node doesn't hold a model. It's the assembly point. You feed it a DiT bundle from the SeedVR2 (Down)Load DiT Model node and a VAE bundle from the SeedVR2 (Down)Load VAE Model node, and it orchestrates the four-phase pipeline (encode → upscale → decode → color-correct) across your frames.

What wires in

• image - your input frames, as an image batch. One image or a whole video's worth; both work.
• dit - the SEEDVR2_DIT output from the DiT loader. This is where 3B-vs-7B and the FP16/FP8/GGUF choice actually got made.
• vae - the SEEDVR2_VAE output from the VAE loader.

No dit, no vae, no upscale. That's the shape of it.

The settings a beginner actually touches

• resolution - target for the shortest edge, default 1080. Aspect ratio is preserved automatically, so this is your "how big" dial.
• batch_size - default 5, and this one has a rule that bites: it must follow the 4n+1 formula - 1, 5, 9, 13, 17, 21… The model processes each batch together for temporal consistency, so it needs that shape. Set it to 1 for a single still; for video, match your shot length if you can (a 21-frame shot likes batch_size: 21). Bigger batches are actually faster per frame and more temporally coherent - but they cost VRAM.
• color_correction - defaults to lab, which is the right default. LAB does full perceptual color matching and keeps output faithful to the source. The alternatives (wavelet, wavelet_adaptive, hsv, adain, none) exist if you want to tune, but lab is where to start.

That's genuinely it for a first run. seed (default 42) makes results reproducible; max_resolution (0 = no cap) auto-scales down if any edge blows past a limit you set.

The VRAM and video knobs

The optional inputs are where you fight OOM and video seams:

• offload_device - defaults to cpu. Where intermediate tensors get parked between phases. cpu is the safe choice for long videos; none keeps everything on the GPU (fastest, hungriest).
• enable_debug - turn this on the moment anything goes wrong. It prints per-phase memory and timing, which tells you which phase OOMed - and that's what decides your fix (encode phase → VAE encode tiling; upscale phase → BlockSwap on the DiT loader; decode phase → VAE decode tiling).
• temporal_overlap (0–16) and prepend_frames (0–32) - video tools. Overlap blends between batches to kill the seam where one batch meets the next; prepend feeds reversed frames at the start to reduce the artifact you'd otherwise get on frame one. Leave both at 0 for stills.
• input_noise_scale / latent_noise_scale - small nudges (try 0.1–0.3 and 0.05–0.15 respectively) if you see artifacts at very high output resolutions. Most people never touch them.
• uniform_batch_size - pads the final batch so it matches the others, avoiding a temporal hiccup when your frame count isn't a clean multiple of batch_size.

Install and the traps worth knowing

Install the pack via ComfyUI Manager (search "ComfyUI-SeedVR2_VideoUpscaler," install, restart) or clone it: cd ComfyUI && git clone https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler.git custom_nodes/seedvr2_videoupscaler, install requirements.txt, restart. Models auto-download on first use.

Two things to know before you judge the output. SeedVR2 lifts contrast - enough that some people feel the result reads as more AI-generated than the input, so eyeball it. And on close-up faces it can wreck skin texture (the corpus's bluntest review was "human to lizard upscaler"). It's not a settings bug, it's the model - give recognizable faces their own pass with ADetailer or inpainting rather than trusting one run. For video specifically, fine repeating textures (patterned wallpaper, fabric weave) can shimmer because consecutive frames resolve them differently; that's the temporal problem, and temporal_overlap only softens it.

The output is a plain IMAGE batch, normalized for ComfyUI, matching your input's RGB/RGBA format - wire it into a Save Image or a video combine node and you're done.