Load FlashVSR DiT (Stock Wan)

This is the "brain" of the pack, and the "(Stock Wan)" in its display name is the whole pitch. Where the official FlashVSR repo runs a standalone reference pipeline, this pack reimplements FlashVSR v1.1 on top of ComfyUI's stock Wan - so this node loads FlashVSR's diffusion transformer as an ordinary ComfyUI MODEL, and everything downstream (guider, sampler shell, VAE decode, attention patches) is standard ComfyUI. If you already run Wan workflows, this is the least surprising node in the pack.

FlashVSR, for anyone who arrived cold: a one-step 4x video super-resolution model from OpenImagingLab that became the open video upscaler people actually run - the community adopted it for speed, where SeedVR2 is the quality pick. This loader is where that model enters your graph.

How it works

It scans ComfyUI/models/flashvsr/ for matching .safetensors files and loads the chosen one with ComfyUI's normal diffusion-model loader. Then it does one non-obvious thing: it patches the model's model_sampling to a flow-discrete, constant (CFG=1) sampling with a shift of 5.0 and multiplier of 1000 - the sampling math FlashVSR was trained with. That's why you can't just load the checkpoint with a stock UNETLoader and expect it to work; the loader is setting up the one-step sampling regime for you.

The output is a single MODEL, which wires into Configure FlashVSR Upscaling, then into BasicGuider.

The one input: model_name

A dropdown of whatever it finds in models/flashvsr/. The matcher is component-based, not filename-exact: it picks files containing "flashvsr1_1" that aren't the LQ projector or decoder. So FlashVSR1_1.safetensors and dtype conversions like FlashVSR1_1-int8_convrot.safetensors both appear - the INT8 ConvRot variant is the one the README highlights as compatible with the pack's shared-activation Q/K/V path and low-VRAM setups. If you see only the "No dit safetensors found in models/flashvsr" placeholder, the file isn't there, isn't lowercase in the right folder, or ComfyUI hasn't been restarted since you dropped it in.

Two model-selection notes. First, use the v1.1 weights (FlashVSR1_1.safetensors); the older v1 checkpoint is not the intended main model for this integration and the README says so in so many words. Second, the Wan 2.1 VAE is not loaded here - it's stock, load it from models/vae with the regular VAE loader, and use whichever decode path you prefer (stock Wan VAE or the pack's Tiny Decode).

Installing it

Part of the ComfyUI-FlashVSR-Stock pack - one install, ten nodes. ComfyUI Manager: search ComfyUI FlashVSR Stock Wan, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Pizzawookiee/ComfyUI-FlashVSR-Stock.git
cd ComfyUI-FlashVSR-Stock
python -m pip install -r requirements.txt

Restart after installing. The Python deps are light (einops, safetensors, a Triton runtime); the real heft is the checkpoint. Download FlashVSR1_1.safetensors (plus the other three component files) from pizzawookiee/FlashVSR-1.1 on Hugging Face into lowercase ComfyUI/models/flashvsr/.

Common issues

• Empty dropdown / "No dit safetensors found." File missing, wrong case on the folder, or no restart. It's always one of those three.
• The model loads but output looks like plain Wan, not upscaled. You've wired the MODEL from this node somewhere that skips the FlashVSR conditioning - it must pass through Configure FlashVSR Upscaling before it becomes a FlashVSR model.
• OOM during sampling. The loader isn't the bottleneck - that's the sampler modes and decoder batching. See the sampler node's notes on streaming_faithful_lowvram.

This node is deliberately boring, and that's the compliment: it makes FlashVSR's DiT behave like any other Wan model in your graph, which is the entire design goal of the pack.