Superside SeedVR2 Upscale Image

SeedVR2 is ByteDance's one-step restoration model that the community quietly adopted as the default single-image upscaler - it adds detail without re-inventing the subject, which is why it took over portrait and product restoration. The catch with the local version is real: it's quantization-sensitive, it wants specific VRAM, and getting it running is a small project of its own. This node skips all of that - it calls SeedVR2 over fal.ai's seedvr/upscale/image/seamless endpoint and hands you the result back in the graph. No local weights, no quant headaches, no GPU in the loop.

How it works

SeedVR2's mechanism is worth knowing because it explains why people use it: one-step diffusion restoration with adaptive window attention, trained with diffusion adversarial post-training. In plain terms - it restores while adding plausible detail, and it's fast enough to sit inside a normal workflow rather than being an overnight job. The hosted version has a couple of nice properties the local one doesn't: no quantization sensitivity (fal runs it at full precision) and no OOM roulette.

The inputs that matter

• image + api_key - the required pair.
• upscale_mode - factor (default) scales by a multiplier; target scales to a fixed resolution.
• upscale_factor - 1–10, default 2. The one you'll set most.
• target_resolution - 720p / 1080p (default) / 1440p / 2160p, used in target mode.
• noise_scale - 0–1, default 0.1. How much noise tolerance the restoration gets; leave near the floor for clean sources.
• seed - reproducibility for the restoration pass.

Output is a single IMAGE - the upscaled result, straight back into your graph to composite or save.

Installing it

It's in the comfyui-superside-nodes pack - clone, install requirements, restart:

cd ComfyUI/custom_nodes
git clone https://github.com/Superside/comfyui-superside-nodes.git
cd comfyui-superside-nodes
pip install -r requirements.txt

Paste your fal api_key. No model downloads - the weight is on fal's side, and this is a pay-per-call upscale.

The honest take

Two caveats carry over from the model's local reputation, because they're properties of SeedVR2 itself rather than of the node. First, the "contrast lift": SeedVR2 tends to brighten and lift contrast, which can make a result read as more AI-generated than the input - if you're matching a batch of catalogue images, be ready to equalize brightness downstream. Second, the "more detail" promise is real but it's still generation: it adds plausible texture, and on an already-sharp image the difference can be subtle while the bill is not. Where this node wins over the local route is pure convenience - if you don't have the VRAM, don't want the quant setup, or just need the occasional clean upscale without a local install, the hosted version is the friction-free door to the same model.