SeedVR2 Video Path Auto Configurator

SeedVR2 is easy to love and annoying to set up. The model itself is ByteDance's one-step restoration network - the thing the community has effectively crowned the default "add detail" upscaler - but the settings between you and it read like a pop quiz: which quant, how many transformer blocks to swap, what batch size fits without OOMing, whether the VAE needs tiling. The SeedVR2 Video Path Auto Configurator's entire job is to run that quiz for you. Give it your GPU, and it emits a complete, working dit, vae, and auto_settings configuration that plugs straight into the SeedVR2 Video Path Upscaler.

It ships in the companion add-on that lives alongside the main numz/ComfyUI-SeedVR2_VideoUpscaler package - the one everyone actually installs. This add-on contributes just these two "Video Path" nodes, and this is the configurator half.

What it actually does

At execution time it checks which GPU you're on - auto-detected, or you can pin a specific device - reads its total and free VRAM plus compute capability, and runs a recommendation engine that picks the model precision, batch size, chunk size, attention backend, tiling, and model-cache residency. It then emits four outputs:

• dit (SEEDVR2_DIT) and vae (SEEDVR2_VAE) - complete model configurations, using the same socket types as the main package's loaders
• auto_settings (SEEDVR2_AUTO_SETTINGS) - the runtime knobs (batch, chunk, color correction, offload…) that override the upscaler's controls when connected
• report - a plain string explaining what it decided: GPU, usable VRAM, chosen model, batch and chunk sizes, cache residency, attention backend

Wire all three config outputs into the Video Path Upscaler and you're done. The report is genuinely useful rather than decorative: it's the node showing its work, so hook it into a Show Text node and you can see exactly why it picked what it picked.

The inputs that matter

Defaults are mostly sane. The four you'll actually touch:

• profile - balanced (7B quality while preserving headroom), maximum_quality (the highest-fidelity model that safely fits), maximum_throughput (smaller native quantized models, larger batches). Start balanced.
• target_resolution - the short-edge output resolution; batch and tiling recommendations scale with its pixel cost.
• reserve_vram_gb - VRAM deliberately kept free for ComfyUI, previews, and other nodes. The 4 GB default is sensible; only cut it if you're genuinely tight.
• device - auto picks the obvious GPU; choose explicitly when you have several.

Optional extras worth knowing: batch_size_override (a known-good 4n+1 value - 1, 5, 9, 13… - replacing automatic sizing; 0 keeps auto), enable_torch_compile (off by default, because the first run's compile cost is steep and it only pays off on repeated workloads), and an optional video input that lets it count the real frames and avoid oversized batches or chunking on short clips.

How to install it

This is an add-on, not a replacement - install the main package first, then:

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

Install its requirements into your ComfyUI venv and restart:

.venv/bin/python -m pip install -r custom_nodes/ComfyUI-SeedVR2_VideoPathUpscaler/requirements.txt

ComfyUI Manager also has it - search "ComfyUI-SeedVR2_VideoPathUpscaler". Models auto-download on first use into ComfyUI/models/SEEDVR2. Expect heavy dependencies (diffusers, peft, gguf, torch) and Python 3.12+ as the recommendation.

Gotchas

• Don't fight auto_settings. While it's connected it overrides the upscaler's batch, chunk, and color controls; twiddling those and wondering why nothing changes is a rite of passage.
• The config is a snapshot, not a promise. Free VRAM at execution time shifts the recommendation, so a crowded workflow may legitimately downgrade the model mid-run.
• Reserved VRAM is yours to lose. Cut reserve_vram_gb only if you know what's eating memory - ComfyUI previews and other nodes take real space.
• It's brand new (v2.5.24, December 2025) and honest about it: the README frames the output as starting points, since resolution, clip length, and kernel availability all move the result. Treat its config as a great first guess, then let enable_debug on the upscaler confirm it actually fits your machine.