SeedVR2 VAE Decode Distributed (Ray)

SeedVR2 is ByteDance's one-step video restoration model, and it became famous for the wrong reason: as an image upscaler it's arguably the best thing in the ecosystem, adding detail without re-inventing the subject. The catch is that it's a heavy VAE-style model, and decoding big latents through it OOMs single cards. RaySeedVR2VAEDecodeDistributed is Raylight's answer: it tiles the spatial dimensions and decodes across your Ray workers, keeping the full temporal sequence on a single worker so the video doesn't come apart at the seams.

How it works, from the source: every Ray worker loads the SeedVR2 VAE (ray_vae_loader), then each worker decodes a spatial partial of the latent (ray_seedvr2_vae_decode_partial, with the job split by rank and world size), and finally the partials are recombined and finalized on worker zero. The tooltips are unusually honest about the design: tile_size is "spatial tile size in output pixels" and the explicit note is "the complete temporal sequence stays on one worker" - so this node scales across GPUs by splitting frames' spatial area, not by splitting time.

The inputs that matter:

• samples - the SeedVR2 latent to decode.
• vae_name - pick the SeedVR2 VAE from your ComfyUI/models/vae folder. No file there, no dropdown options; this is the one model you have to download yourself. (ByteDance ships SeedVR2 under Apache 2.0, which is also why the KB's upscaling essay recommends it even for commercial work.)
• tile_size (default 512) - spatial tile size in output pixels. Larger tiles = fewer seams but more VRAM per worker; this is your main VRAM knob.
• overlap (default 64) - how much neighboring tiles overlap, so seams blend instead of slicing. If you see tile boundaries in the decoded output, raise this.

Output is a plain IMAGE, ready for a preview or save node.

Where it fits

The natural pairing is the pack's own video upscaling flow: generate a latent in Raylight, then decode it with this node instead of the normal VAE decode when you're using a SeedVR2-based restoration path. The KB's upscaling essay is unambiguous that SeedVR2 is the incumbent detail-adder - on clean images use it straight, on genuinely damaged sources pair it with something like Qwen-Edit - and this node is how you run it big without a 24GB card.

Install

Standard raylight, plus one model file:

cd ComfyUI/custom_nodes
git clone https://github.com/komikndr/raylight
cd raylight
pip install -r requirements.txt

Then put the SeedVR2 VAE into ComfyUI/models/vae/. Restart, or install "raylight" via ComfyUI Manager.

Gotchas

• No model file = no dropdown. The vae_name enum is populated from your vae folder at load time; if the field is empty, that's a download problem, not a node problem.
• It needs at least one Ray worker and will raise a ValueError if ray_actors has none - make sure Ray Init Actor is upstream and healthy.
• Tiling is spatial-only. Very long temporal sequences won't get the memory relief you might expect from "multi-GPU VAE decode," because time stays on one worker by design. That's a correctness choice, not an oversight - splitting time across workers is how you get temporal seams.