SeedVR2 Video Path Upscaler (v2.5.24)

Most ComfyUI video nodes cheat: they unload the whole clip into memory as one big IMAGE tensor and process it in a single pass. That's exactly how SeedVR2 earned its OOM reputation - video was the "you'll need a datacenter GPU" direction for months, while the same model quietly became the default image upscaler. The SeedVR2 Video Path Upscaler is this add-on's answer. It takes a native ComfyUI VIDEO, streams it through the model in bounded chunks, and hands back a file-backed video. The complete frame sequence never becomes an IMAGE tensor in your workflow at all.

How it works

Instead of loading the clip, it decodes frames lazily with PyAV, processes them chunk_size frames at a time (default 21), writes each finished chunk to an H.264 intermediate, and carries a few context frames (chunk_overlap, default 2) from the previous chunk so temporal seams don't show. Those context frames are stripped from the output, so duration stays exactly the same. Once the clip is done it muxes the source audio back in.

Because the model is genuinely temporally aware - the 4n+1 batch pattern isn't cargo cult, the model needs those frame counts for temporal consistency - it sidesteps the shimmer you get from a per-frame image upscaler on fine repeating texture. That's the whole reason to use SeedVR2 on video rather than any good image upscaler applied 24 times a second.

The inputs that matter

• video - a native VIDEO, normally from core Load Video. File-backed and decoded only as needed.
• dit / vae - model configurations from the main package's loaders, or from the sibling Auto Configurator in this pack.
• batch_size - frames processed together. Follows 4n+1: 1, 5, 9, 13, 17, 21… Default 5. Higher = better temporal consistency and speed, more VRAM.
• chunk_size - max source frames held in RAM at once (default 21). Set 0 to disable chunking and process the complete active video in one pass.
• resolution - short-edge target in pixels, aspect ratio preserved. max_resolution caps either dimension (0 = no limit) for extreme aspect ratios.
• color_correction - lab is the default and the recommended one; none leaves the model's output uncorrected.

You can ignore most optional inputs, but temporary_video_crf (H.264 quality of the intermediate, default 18, lower is better) and offload_device (cpu default, for intermediate tensors between phases) matter if you're tuning. Connect auto_settings from the Auto Configurator and it overrides the corresponding controls above.

Outputs

• video - a file-backed upscaled H.264 video with the source audio embedded when present. Connects directly to core Save Video or any native-video save node.
• audio - the full, unchunked audio for the active source window as a separate AUDIO value, handy for alternate muxing or replacing the track.

Installing it

Same pack, same story: this is a companion add-on to the main numz/ComfyUI-SeedVR2_VideoUpscaler package - install that first, then:

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-SeedVR2_VideoPathUpscaler.git

Install its requirements into your ComfyUI venv (.venv/bin/python -m pip install -r custom_nodes/ComfyUI-SeedVR2_VideoPathUpscaler/requirements.txt), restart. ComfyUI Manager finds it under the same pack title. Models auto-download to ComfyUI/models/SEEDVR2 on first use.

Gotchas

• chunk_overlap must be smaller than chunk_size - the node raises a clear error otherwise.
• Alpha isn't preserved. The H.264 intermediate is RGB only; if you need transparency, the main package's tensor upscaler is the path.
• Quant still matters. The standing community advice: FP16 if it fits - GGUF damages skin texture and FP8 can introduce tiling on some content, so don't judge SeedVR2 on a quant it wasn't meant for.
• Budget the compute. Even memory-bounded, video upscaling is expensive; community numbers put 720p→4K at roughly twelve minutes on a rented 48 GB L40S. Test a short clip with enable_debug on before committing a long video.
• PyAV (av) does the file work here. ComfyUI bundles it so you're fine in the GUI, but the requirements.txt oddly omits it - if you ever run the pack's CLI standalone, you may need to pip install av yourself.