WanVideo SEI Clip Encode

This is the node that makes the whole pack worth using. Everything else here is standard Wan plumbing forked to match it - this is the one doing something Kijai's mainline wrapper didn't originally do: take two images, a start frame and an end frame, and turn them into conditioning that steers an entire generated clip from one to the other.

Worth knowing before you build with it: this predates the feature landing in Kijai's own ComfyUI-WanVideoWrapper. When Kijai added native start/end frame support in March 2025, the community explicitly noted it had "earlier been available separately in raindrop313's WanVideoStartEndFrames" - this pack is where the idea got proven out first. If you're wondering whether to use this or Kijai's now-native version: they do the same job, this is just the original standalone implementation, and plenty of existing workflows and tutorials are still built directly around it.

How it works

You feed it a CLIP vision model, your start and end images, and the VAE from WanVideoSEVAELoader. It encodes both images, resizes them to your target generation width and height, and builds a stack of image conditioning across your target frame count - one continuous embedding the sampler reads for the whole clip, weighted toward the start image early and the end image late.

The inputs and outputs that matter

• start and end - the two images. Their tooltips are plain about the job: "Image to encode." This is the entire point of the node.
• generation_width / generation_height (default 832×480, Wan's standard 480p aspect) and num_frames (default 81) - matches the README's own recommendation to generate 25 or more frames, since fewer than that tends to let character identity drift across the clip.
• start_frame_weight and end_frame_weight (both optional, default 1, range 0–10) - the real creative knob here. Their tooltips spell it out clearly: push a weight above 1 and that frame's features dominate more of the video; push it toward 0 and the model gets more freedom to drift away from it, with 0 ignoring that frame entirely. Crank start_frame_weight high and end_frame_weight low for a clip that opens faithfully and wanders by the end, or balance them for a clean transition between two fixed points.

A few more optional dials: noise_aug_strength (adds noise for extra motion, useful in I2V per its own tooltip), latent_strength (lower values give the model more room to move away from your source frames), clip_embed_strength (a multiplier on the CLIP conditioning), and adjust_resolution (on by default, reproducing the original code's automatic resolution handling - generally fine left alone).

Output is a single image_embeds socket, typed WANVIDIMAGE_EMBEDS, plugging directly into WanVideoSESampler.

How to install it

ComfyUI Manager, search "ComfyUI-WanVideoStartEndFrames," install, restart. Or manually:

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

You'll also need a Wan-compatible CLIP vision model - download it per Kijai's ComfyUI-WanVideoWrapper instructions, since that's the pack this fork's README defers to for every model file it uses.

Common issues & troubleshooting

Wrong CLIP vision checkpoint. The most common mistake is loading a CLIP vision model meant for a different family - an SDXL or Flux one, say. It'll load without erroring, but the encoding will be nonsense, since Wan's CLIP vision expects its own matching weights.

End-frame results look mushy. This is a known, honestly-documented limitation, not something you're doing wrong. The README says adding end-frame guidance to I2V "seems to degrade video generation quality" compared to plain start-frame generation - no LoRA or model fine-tuning has been done yet to fix it. Treat end-frame steering as genuinely useful but imperfect, not broken.

Distorted output that seems to trace back here. It might not be this node at all - the README's blunt warning that generating without a real positive prompt "can result in severe video distortion" points at the text side of the graph, not the image side. If your clip looks wrong, check your prompt before you re-tune weights here.