WanVideo Add FlashVSR Input

FlashVSR is a video super-resolution model - it takes a low-resolution clip and upscales it while keeping it temporally stable, which is the hard part of video upscaling. The KB's read: it "arrived October 2025 and broke through immediately," and it's the open video upscaler that actually got adopted, best when "the source is already reasonable and you want speed." This node is how you hand your low-res frames to a FlashVSR run inside Kijai's wrapper - it packages those input images into the Wan embeds the FlashVSR path expects.

The reason you upscale at all is baked into how people use Wan: generate at 480p, then climb, because generating natively at high resolution is slow and memory-hungry. FlashVSR is one of the two upscalers worth reaching for (SeedVR2 is the other; the KB says SeedVR2 wins "when the source is genuinely low-resolution and you can afford the compute," FlashVSR wins on speed).

How it works

You give it your Wan image embeds and the low-resolution images you want upscaled, plus a strength. It attaches those images as the FlashVSR input conditioning and returns modified embeds. The generation then produces the higher-resolution result guided by your low-res frames.

The inputs and output

Three inputs, all required, and only one you'll really fiddle with:

• embeds (WANVIDIMAGE_EMBEDS) - the embeds being modified.
• images (IMAGE) - the low-resolution frames to upscale. This is the source material.
• strength (default 1.0, range 0–2) - how strongly the input frames guide the result. At 1.0 it tracks the source faithfully; nudge up toward 2 if the upscale is drifting from your original, down if it's clinging so tightly it can't clean anything up.

The output is image_embeds (WANVIDIMAGE_EMBEDS) for the sampler.

How to install it

Comes with Kijai's WanVideoWrapper. ComfyUI Manager: search ComfyUI-WanVideoWrapper, install, restart. Manual:

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

then restart. The node adds no weights by itself, but the FlashVSR path needs the FlashVSR model support loaded. Heads up: FlashVSR can be VRAM-hungry on video - the KB quotes a user upscaling 640x880, 81 frames, 2x on a 24GB 3090 with tiling disabled, and noting tiling is "far slower" but "used less than a third" of the VRAM. So if you OOM, tiling (configured on the loader/sampler side, not this node) is the lever.

Common issues & troubleshooting

Ran out of memory. Video super-resolution is heavy. Enable tiling on the FlashVSR DIT and VAE side to trade speed for a big VRAM reduction - the KB's numbers show it dropping a 24GB job to under a third of that.

The upscale drifts from the source. Raise strength toward 2 so the output stays anchored to your input frames. If it's over-smoothing or losing detail instead, the model, resolution ratio, or tiling settings matter more than this one knob.

Wrong tool for a genuinely tiny source. FlashVSR shines when the input is "already reasonable." The KB is explicit that for very low-resolution restoration (256px→1024px territory), SeedVR2 beats it. If your source is truly rough, this may be the wrong upscaler regardless of settings.