WanVideo Model Loader(SE)

Every workflow built on raindrop313's ComfyUI-WanVideoStartEndFrames starts here. WanVideoSEModelLoader points at a Wan2.1 checkpoint sitting in ComfyUI/models/diffusion_models and hands back a loaded model for the rest of this pack's nodes to run. It's not a general Wan loader you'd reach for elsewhere - it's this fork's own copy of Kijai's model loader, kept separate so the sampler and image encoder in this pack can plug into a version of Wan wired for two anchor frames instead of one.

Worth knowing before you build with it: this whole pack is a modification of Kijai's ComfyUI-WanVideoWrapper, adding the ability to guide a generation with both a start image and an end image. It wasn't a side experiment either - it's the original implementation. When Kijai's own wrapper later shipped native start/end frame support, the community explicitly noted it had "earlier been available separately in raindrop313's WanVideoStartEndFrames." So if you've run into start/end frame support baked into the mainline wrapper and wondered why this standalone fork still exists - this is where it came from, and a lot of existing workflows still point straight at it.

How it works

Straightforward loader mechanics: pick a checkpoint file, pick a precision and a quantization scheme, decide whether it lives on GPU or gets shuttled from CPU RAM, and out comes a loaded model object the sampler consumes. Nothing generative happens in this node - it's pure setup.

The inputs and outputs that matter

Four required fields, and really two of them decide anything:

• model - the dropdown of files ComfyUI finds in models/diffusion_models. Your Wan2.1 checkpoint, 480p or 720p depending on your target resolution (the README supports both).
• quantization - disabled by default, or one of a long list of fp8 and torchao schemes. This is your real VRAM lever: fp8 roughly halves memory versus full precision for a usually-small quality cost.
• base_precision (bf16 default) and load_device (main vs. offload device) round out the required set - leave both on default unless you have a specific reason not to.

Optional inputs are for squeezing performance on smaller cards: attention_mode (sdpa default, or flash_attn_2/3, sageattn, spargeattn - sageattn is the popular speed pick but needs Triton installed, which is notoriously fiddly on Windows), lora for stacking a LoRA onto the model as you load it, and block_swap_args / vram_management_args for offloading transformer blocks (the latter's own tooltip calls it "more aggressive... but can be slower" than block swapping).

Output is a single model socket, typed WANVIDEOMODEL, feeding straight into WanVideoSESampler.

How to install it

Via ComfyUI Manager: search "ComfyUI-WanVideoStartEndFrames," install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/raindrop313/ComfyUI-WanVideoStartEndFrames

then restart ComfyUI. One thing the README is upfront about: it doesn't repeat model download or setup instructions, it just points you at Kijai's ComfyUI-WanVideoWrapper for those, since this fork shares his code and model layout. Install that pack alongside this one too - you'll need it for the text encoder node this pack doesn't ship, and its docs are the ones that actually tell you which Wan checkpoint files to grab.

Common issues & troubleshooting

Empty model dropdown. You haven't put a diffusion model file in ComfyUI/models/diffusion_models yet, or you added one without restarting - the folder gets scanned at startup, not live.

Quantization import errors. The torchao options need the torchao package installed, which is a heavier ask than the fp8 modes. If you don't specifically need torchao's schemes, fp8_e4m3fn is the simpler, more commonly used choice and works with far less setup friction.

sageattn fails to load. That's Triton/SageAttention not being installed correctly on your system, not a bug in this node. Confirm the default sdpa attention mode works first before troubleshooting the faster backends.