🚦 TA SeedVR2 Gate

SeedVR2 is currently the best-regarded upscaler in the ecosystem - the KB's panel calls it the top rung since late 2025 - but "best" comes with a price: its model loaders pull down on the order of 14GB of weights. If upscaling is the last optional step of your workflow, you don't want those weights loaded on runs where you don't need the upscale. TA SeedVR2 Gate is the on/off valve for exactly that, built on ComfyUI's lazy evaluation so a disabled gate means the heavy loaders never even run.

The design is clean. One required input, enabled (default false), and three optional inputs: image, dit (type SEEDVR2_DIT), and vae (type SEEDVR2_VAE) - the model objects emitted by the SeedVR2 loader/upscaler nodes. All three optional inputs are marked lazy. When enabled is on, the gate requests image/dit/vae and passes them through unchanged on its three outputs. When it's off, lazy evaluation means those inputs are never requested at all - ComfyUI skips the entire upstream chain, including the model loaders, so a disabled run costs you nothing.

That's the whole trick, and it's the right trick: lazy inputs are ComfyUI's built-in mechanism for "don't execute upstream unless this node actually needs the data," and the gate leans on it instead of hand-rolling bypass logic. The workflow implication is that the SeedVR2 loaders sit upstream of the gate, and the gate sits between them and the upscaler application - so flipping one boolean turns the whole 14GB pipeline on or off.

Where this fits in the wider picture: SeedVR2's community consensus is strong but conditional - FP16 is the clean path (GGUF can damage skin detail, FP8 can add tiling artifacts), it runs fine on 6GB for images, and its known failure modes are a contrast lift and occasionally plasticky skin on faces. So this gate is also your "I'm not upscaling this batch" decision point: iterate fast on raw outputs, flip the gate on only for the runs that actually deserve the upscale pass.

Install. Pack ships dependency-free, but the gate is only useful if you have the SeedVR2 nodes installed too - it expects their SEEDVR2_DIT/SEEDVR2_VAE types:

cd ComfyUI/custom_nodes
git clone https://github.com/tmode-1960/TA-ComfyUI-Nodes-Pack

or ComfyUI Manager (search "TA ComfyUI Nodes Pack"), restart, under TA Nodes. Install the SeedVR2 node pack separately and have its model files available.

Gotchas. If the dit/vae inputs show red or can't connect, you don't have the SeedVR2 nodes installed - the gate is a gate, not a loader. enabled defaults to false, which is intentional: it's a "turn on upscaling" switch, and forgetting it's off just means you get un-upscaled output, not an error. And if the upscale itself looks bad, the KB's advice is to check which quantization you ran before blaming the pipeline. Plus the usual v2.x pack-rewrite note for old TA v1 workflows.