Seedvr Upscaler (fal)

SeedVR2 is the upscaler that took over the "add detail" job around late 2025 and hasn't let go - about one upscaling thread in three still mentions it. It's technically a ByteDance video restoration model that the community discovered does its best work on single images, and what makes it special is restraint: it adds detail without re-inventing your subject, which is exactly why it displaced SUPIR for anything involving a real face. This node runs it on fal.ai, so you skip the part where people usually get stuck - the local install and the quantization lottery.

Hosted call, per-image billing, no VRAM to manage. That last bit matters more than usual here: SeedVR2 has a well-earned OOM reputation locally, and running it on fal sidesteps it entirely.

How it works

SeedVR2 is a one-step restoration model. You hand it an image and a factor, and it rebuilds a higher-resolution version, synthesizing plausible fine detail - pores, fabric weave, lashes - rather than just interpolating pixels. That's the whole reason to use it over a plain ESRGAN pass: it's generative detail, aware of what it's looking at.

Important framing from the upscaling playbook: SeedVR2 is for clean images that need more detail, not damaged ones. If your source has heavy compression, noise, or blur, that's a restoration job and SUPIR is the tool. SeedVR2 preserves what's there; it doesn't repair wreckage.

The inputs and outputs that matter

Refreshingly short:

• image (IMAGE, required) - the thing you're upscaling.
• upscale_factor (default 2, range 1–4 in 0.5 steps) - how much bigger. 2× is the safe default; 3–4× is where it shines but also where it costs most.
• seed (optional, default -1 = random) - the detail synthesis has some randomness, so a fixed seed lets you reproduce a result.

Output is a native IMAGE, straight into Preview/Save or a further pass. Note this is the image upscaler - SeedVR2 video upscaling is a datacenter-GPU job, and this node takes an IMAGE, not a VIDEO.

How to install it

From gokayfem's ComfyUI-fal-API pack. ComfyUI Manager: search ComfyUI-fal-API, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
pip install -r ComfyUI-fal-API/requirements.txt

Add your fal key (copy config.ini.example → config.ini, paste from fal.ai/dashboard/keys, or export FAL_KEY=your_key), restart.

Common issues & troubleshooting

Close-up faces come out plasticky or scaly. This is SeedVR2's best-known weakness - its most-upvoted criticism in the community is literally "human to lizard upscaler." It's real, and it's specific to skin texture on faces. The fix is design, not settings: give the face its own pass (ADetailer or inpainting) instead of trusting one SeedVR2 run on a portrait.

The result reads as more AI-looking than the input. SeedVR2 lifts contrast, consistently enough that some images end up looking more obviously generated. If tonal fidelity matters, compare before/after and dial your expectations - it's a known trait, not your mistake.

You pointed it at a damaged photo and it looks worse. Wrong tool. SeedVR2 doesn't fix compression or heavy blur; it sharpens and details a clean source. For genuine restoration, reach for SUPIR.

Insufficient funds / billing. fal is pay-per-call and legit; top-ups have occasionally posted late. Wait for credits rather than repaying. fal's errors come through as plain text in ComfyUI.