Load HelloMemeVideo Pipeline

If Load HelloMemeImage Pipeline builds the still-image generator, this node builds the video one. Same idea, same inputs, same presets - but the pipeline it assembles carries the temporal machinery (the motion modules and chunked sampling) that HelloMeme Video Pipeline needs to animate a photo over many frames instead of one. It's the node you load when the goal is a talking-head clip, not a still.

Almost everything said about the image loader applies here, so this is the short version: the differences, not the full essay.

The inputs that matter

• checkpoint - SD 1.5 base, either a [preset] (auto-downloaded from the author's HuggingFace) or anything you have in ComfyUI/models/checkpoints. Realistic presets fight the SD look better.
• version - v1 through v5 (note: no v5b/v5c here, those are image-pipeline-only). Default is v2, which is where the pack got its lower-VRAM and third-party-checkpoint compatibility improvements.
• stylize - x1 or x2 super-resolution pass. On video, x2 multiplies the cost across every frame, so only reach for it when you have the VRAM.
• lora / lora_scale - optional style LoRA, preset or from ComfyUI/models/loras.
• vae, deployment, dtype - same as the image loader; fp16 when VRAM is tight.

Output: hm_video_pipeline (HMVIDEOPIPELINE), wired into HelloMeme Video Pipeline.

How it works

Same assembly as the image pipeline - a diffusers SD1.5 pipeline with the HelloMeme spatial-knitting modules injected - plus the video path: a motion module for temporal coherence and chunked generation so you can produce clips longer than a fixed frame budget. The first load downloads the HM modules and any preset checkpoint, so expect a slow first run.

Installing it

Part of the HelloMeme pack; install once, get all nine nodes. ComfyUI Manager keyword hellomeme-api, or:

cd ComfyUI/custom_nodes
git clone https://github.com/HelloVision/ComfyUI_HelloMeme
# restart ComfyUI

Dependencies are the pack's standard heavy stack - transformers, diffusers, accelerate, peft, onnxruntime, opencv-python, scikit-image. Models auto-download on first use.

Common issues

• VRAM is the real limit on video. Video multiplies the image pipeline's footprint by the number of frames in flight. If you're under ~12GB, stay on v2, avoid stylize x2, and prefer short clips - the README's own testing ran on an 11GB 2080 Ti with the lighter path.
• Version mismatch between loader and sampler. Load the video pipeline and feed HelloMeme Video Pipeline, not HelloMeme Image Pipeline. It's easy to cross-wire the two when you're copy-pasting a workflow.
• First-run download storm. New versions fetch new modules. If it hangs, switching deployment to modelscope has rescued people with flaky HuggingFace access.