Nodes/ComfyUI-WanVideoWrapper/WanVideo FlashVSR Decoder Loader
ComfyUI Node Runs on cloud

WanVideo FlashVSR Decoder Loader

The fast upscale VAE for Wan

By kijai·Created about a year ago·Updated 3 months ago· 6,654
WanVideo FlashVSR Decoder Loader
    • vae
    model_name
    precisionbf16

    FlashVSR is the open video upscaler that actually broke through - it arrived in October 2025 and the community's first reaction was straightforward: "so much faster than SeedVR2." Speed, not top-end quality, is its whole selling point. This node loads the Wan VAE that FlashVSR's decode path in WanVideoWrapper uses - the piece that turns upscaled latents back into pixels quickly.

    Where FlashVSR actually fits

    Worth setting expectations honestly, because the picture has shifted since FlashVSR launched: it's fading relative to SeedVR2, and the two have settled into a fairly clean division of labor. FlashVSR wins when your source is already reasonable quality and you want speed. SeedVR2 wins when the source is genuinely low-resolution and you can afford the extra compute - a direct comparison in the community found SeedVR2 beats FlashVSR specifically on heavy upscales, the 256px-to-1024px kind of lift. If you're grabbing this node expecting FlashVSR to be the default answer for every video-upscale job, it isn't; it's the fast option for a source that doesn't need much rescuing.

    What this specific node does

    This is a VAE loader, not the upscaler itself - it loads a Wan VAE model from your standard ComfyUI/models/vae folder, at whichever precision you choose, and hands back a WANVAE object that FlashVSR's decode step in this pack consumes. Practically: you'll pair this loader's output with FlashVSR's sampling/upscale nodes elsewhere in WanVideoWrapper, not use it standalone.

    The inputs and output that matter

    model_name - a dropdown of Wan VAE checkpoints found in ComfyUI/models/vae. precision (optional, default bf16, also fp16/fp32) - the compute precision for decoding; bf16 is the sane default for most modern GPUs, fp32 costs more memory and time for marginal quality gain, fp16 is the lighter-weight option if you're VRAM-constrained and your hardware handles it well.

    Output is a single vae (WANVAE), which feeds into whatever FlashVSR decode node consumes it downstream.

    How to install it

    Via ComfyUI Manager, search WanVideoWrapper and install, or manually:

    cd ComfyUI/custom_nodes
    git clone https://github.com/kijai/ComfyUI-WanVideoWrapper
    

    then restart. You need a Wan VAE checkpoint already sitting in ComfyUI/models/vae for the dropdown to show anything - the same VAE files you'd use for regular Wan generation work here, no FlashVSR-specific VAE download required beyond that.

    Common issues & troubleshooting

    Empty model_name dropdown. This loader scans ComfyUI/models/vae directly - if you haven't downloaded a Wan VAE checkpoint into that folder, there's nothing to select. It's a loader, not a fetcher.

    Flickering at batch or chunk boundaries when upscaling long clips. This is a known limitation across video upscalers in this space, not unique to FlashVSR - SeedVR2 has the identical issue when it can't upscale a whole clip in one batch. If you see visible seams every couple of seconds, that's the boundary-joining problem, and the practical fix is usually processing in larger continuous batches where your VRAM allows, rather than tuning this loader.

    Wondering whether to reach for this over SeedVR2. Use the honest split above: source already decent and you want it fast, FlashVSR. Source genuinely low-res and quality matters more than time, SeedVR2. If you're not sure which bucket your footage is in, it's cheap enough to try FlashVSR first and fall back to SeedVR2 if the lift looks weak.

    CategoryWanVideoWrapper

    Inputs (2)

    NameTypeDefaultDescription
    model_nameCOMBOThese models are loaded from 'ComfyUI/models/vae'
    precisionoptCOMBObf163 options: fp16, fp32, bf16

    Outputs (1)

    NameTypeDescription
    vaeWANVAE