FL SeedVR2 1.4B Loader

If you've spent any time upscaling in ComfyUI in the last year you already know SeedVR2 - the ByteDance model the community adopted as the default "add real detail" upscaler. What you might not know is the annoying part: pointing ComfyUI at the right checkpoint, the right VAE, and hoping the two actually match. FL SeedVR2 1.4B Loader exists to make that whole step disappear. Drop it in, wire its two outputs, and on first queue it quietly downloads, verifies, and loads both model files for you.

It's from the filliptm FL pack family (same author as the FL-Trainer LoRA training packs, Patreon "Machinedelusions"). Where most SeedVR2 installs reach for the community ComfyUI-SeedVR2_VideoUpscaler nodes, this pack rides ComfyUI's built-in SeedVR2 support instead - no second implementation bundled, nothing patched at import. That's the honest, low-magic design and it's refreshing.

What it does

The loader has a single job and a single input. On execute it:

1. Checks your ComfyUI is new enough (0.28.0 or newer, which is when native SeedVR2 landed).
2. Checks whether the pinned 1.4B transformer and VAE are already registered in the standard model folders - FL SeedVR2 1.4B Upscale (the sibling node) needs both.
3. Downloads anything missing: a 2.89 GB transformer into models/diffusion_models/ and a 501 MB VAE into models/vae/. That's about 3.4 GB of storage you didn't have before.
4. Loads them through ComfyUI's own SeedVR2 path and hands you a normal MODEL and VAE.

The downloads are worth a closer look because they're unusually careful: resumable transfers (restart ComfyUI mid-download and it picks up where it left off), file-size checks, model-header checks, and SHA-256 validation before a file is moved into place. It also validates that the transformer is actually the six-block ComfyUI variant - the plain SeedVR2-1.4B checkpoint from HuggingFace won't work because it lacks the fixed conditioning tensors ComfyUI expects, and the loader will tell you so instead of silently giving you garbage. Model paths from extra_model_paths.yaml are respected too.

One gotcha: downloads only start when the node executes. Install the pack, restart, and nothing happens until you queue a prompt with the loader in the graph. That's a feature (no surprise 3 GB fetch on install) but it means your first queue after setup will sit on the download progress bar for a while.

The input that matters

• download_if_missing (boolean, default true) - the whole point. Leave it on unless you've manually placed the model files yourself. Flip it off and the loader expects you to have handled it, which is a real trap if you're on a machine without internet on first run.

That's it. No model dropdown, no filename to type. The pinned files are the only ones it will ever look for.

What comes out

• model (MODEL) - the SeedVR2 1.4B transformer, wrapped for ComfyUI's device/offload handling. Wires into the model input of FL SeedVR2 1.4B Upscale.
• vae (VAE) - the matching VAE. Wires into vae on the same node.

Both outputs are standard ComfyUI types, so nothing else in your graph needs to know this pack exists.

Installing it

Via ComfyUI Manager: search for ComfyUI-FL-SeedVR2 and install, then restart ComfyUI. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-SeedVR2.git

Then restart. There's no requirements.txt to worry about - the pack deliberately uses only dependencies ComfyUI already ships. What it does require is ComfyUI 0.28.0 or newer; anything older and you'll hit the error below.

The one error everyone hits

unexpected keys, starting with blocks.0.attn.proj_out.txt.bias

This scares people into deleting and re-downloading the model. Don't. It means your ComfyUI is older than 0.28.0 and built an incompatible SeedVR2 architecture - update ComfyUI, restart, and queue again. The downloaded files are valid and stay put.