星光2.6｜视频超分（v2.6.0-rh1）

Let's get the branding out of the way first, because it will save you a confused hour. This node's Chinese display name is 星光2.6｜视频超分 - "Starlight 2.6, video super-resolution" - but open the package and every file inside says SeedVR2. What you've installed is ByteDance's one-step diffusion restoration model, forked from the community-standard numz/ComfyUI-SeedVR2_VideoUpscaler nodes and re-branded for RunningHub's GPU-cloud workflow market. Not a scam - the upscaling is real and genuinely good - it just means every bit of SeedVR2 wisdom you already have transfers directly. You're not learning a new model; you're learning a new wrapper around the best-known one.

What it's actually for

This is the detail-adding kind of upscaling, not the "make it bigger" kind. Feed it a low-resolution clip and it runs a diffusion pass over the latent space to invent plausible detail - fabric weave, pores, fine texture - the way a Lanczos or ESRGAN pass never can. It shines on video that's genuinely low-res: SeedVR2 consistently beats FlashVSR on the big lifts, like 256px to 1024px. For a source that's already decent it's overkill - a cheap pixel upscaler is faster and can't hallucinate.

The uncomfortable part: this is a slog on consumer hardware - the community figure is 12 minutes for a 720p→4K clip on a rented 48GB L40S. Keep the batch small and the clip short and a 3060 can do it; for long footage, rent a GPU or run it on RunningHub, where the weights are already mounted.

How it works

The node expects three things wired in: your frames as an image batch, plus a dit and a vae configuration from the pack's other two nodes - Starlight26LoadDiTModel and Starlight26LoadVAEModel. It runs a four-phase pipeline: VAE-encode the frames, diffuse-upscale them with the DiT, decode, then color-correct.

This fork earns its keep on video, where per-frame upscaling fails in a classic way: consecutive frames disagree on how to resolve fine repeating texture, and you get shimmering wallpaper. The node processes frames in batches shaped 4n+1 (1, 5, 9, 13…), the temporal window the model's video VAE was trained on. temporal_overlap re-feeds a few frames across batch boundaries and blends them so seams don't appear, and uniform_batch_size pads a short final batch so the tail doesn't wobble. prepend_frames runs a few reversed lead-in frames and trims them, cleaning up artifacts at the start of the clip. All of it exists to make adjacent frames agree.

The output is a single IMAGE - the upscaled frame batch in the same RGB/RGBA format as your input, normalized to [0, 1] for ComfyUI. Wire it into a save-video or preview node.

The inputs that matter

• resolution (default 1080): target for the shortest edge, aspect ratio kept. The number you'll actually change.
• batch_size (default 121): frames per batch, and it must follow 4n+1. Bigger means better temporal consistency and faster throughput but more VRAM; smaller shrinks the footprint.
• color_correction: SeedVR2 famously lifts contrast and shifts color. This pass grades the output back toward your input - lab is the tooltip's recommended pick, wavelet the shipped default, none for the raw upscale.
• max_resolution (default 3840): a safety cap on the longest edge so extreme aspect ratios don't blow your VRAM.

Seed, the two noise sliders, and offload_device are the fiddly rest. Leave them alone until something misbehaves.

Installing it

Through ComfyUI Manager, search for the pack title or starlight26-runninghub. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/284951641/ComfyUI-LHStarlight26

Then restart ComfyUI. Manager installs the pip deps from requirements.txt - diffusers, peft, omegaconf, gguf, opencv - so the first install is chunky. Two gotchas. First, this pack is written against ComfyUI's newer V3 node API, so an old ComfyUI won't register the nodes - update before you try. Second, the README's only real warning: model weights are not included. Weights go in ComfyUI/models/SEEDVR2/ (the folder is literally named SEEDVR2); the defaults expect a starlight26_dit_3b_bf16.safetensors plus a matching VAE. On RunningHub those are mounted for you; locally you have to fetch them. The standard numz/AInVFX SeedVR2 3B/7B .safetensors and .gguf files work in this fork too.

Common issues

OOM. This is SeedVR2's earned reputation. Drop batch_size to 5, set the upscaler's offload_device to cpu, and turn on the loader nodes' tiled encode/decode or block-swap options.

"It upscaled fine but looks bad." Ask which weights you ran. FP16/bf16 is clean; GGUF quant is known to damage skin texture and FP8 can leave tiling-grid artifacts - the bf16 defaults are the safe ones. And on a close-up face, don't trust a single pass: the community's harshest SeedVR2 verdict is "human to lizard upscaler" on skin. Give faces their own pass.

Jitter between batches. Bump temporal_overlap to 1–4 and keep uniform_batch_size on.

If the node errors on a missing model file, the answer is boring and always the same: models/SEEDVR2/ is empty. Put the weights in there and it runs.