WeeLLM Video Generate

Video generation is the one place the low-VRAM crowd usually just has to give up. A 34B video model quantized and offloaded still wants far more than a 4 GB card has, and the usual GGUF advice stops being practical around the point where you're juggling a few gigabytes of spare VRAM. WeeLLM Video Generate is the sibling node that says "not necessarily" - it produces frame batches from big video models like MiniMax-H3 while the same layer-streaming engine from the WeeLLM Loader keeps peak VRAM near the 3 GB mark.

The catch list is real, and you should hear it before the excitement: this is a niche pack from a single author with essentially no community track record yet, it's slow (layer streaming trades VRAM for wall-clock time, and video multiplies that by every frame), and it only hands you frames - you still need a video-combine node to turn them into a file.

How it works

Like the image Generate node, it's a thin wrapper. The real work happens in the pipeline you fed it, so the setup matters more than the knobs:

1. In WeeLLM Loader, set model_path to a video-capable model repo (MiniMax-H3 or LTX 2.5 are the ones the underlying weellm library lists) and - this is the easy one to miss - set task to video. That loads WeeVideoPipeline instead of a text-to-image class.
2. Wire the WEE_PIPELINE socket into this node.
3. Generate, and the node unpacks the engine's output into a single IMAGE batch - a stack of frames in [T, H, W, C] order - and hands it to whatever comes next.

That last handoff is the whole design. The output is called frames and it's frames, not a movie. Route it into a Video Combine node (the VideoHelperSuite family is the usual choice) and encode there.

What you'll actually set

• prompt - multiline, does what you expect.
• width / height - default 960×544, step 8, capped at 2048. Video is memory-hungry even streamed, so don't treat the cap as a target on a 4 GB card.
• num_frames - default 75, floor 22, ceiling 500. At 24 fps, 75 frames is roughly a 3-second clip. More frames = more streaming passes = a lot more waiting.
• num_steps - default 6. With video, every extra step multiplies across every frame, so respect it.
• seed - reproducibility.

The two optional sockets are where the interesting workflows live: image conditions on a first frame (image-to-video), and last_image pins the ending too - MiniMax-H3's FL2VA path is built for start-and-end bookending, so you can generate a clip that resolves into a specific final frame.

Installing it

Same pack as the rest of the WeeLLM family - ComfyUI Manager, search ComfyUI-WeeLLM, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Jit-Roy/ComfyUI-WeeLLM.git
cd ComfyUI-WeeLLM
pip install -r requirements.txt   # installs weellm + a pinned diffusers commit

Where people get burned

Three things will bite you, in order of likelihood.

Loader task mismatch. If the Loader isn't on video, this node gets a text pipeline and the whole thing falls over or silently does nothing. Check the Loader first.

System RAM, not just VRAM. Layer streaming parks model blocks in system RAM while it feeds the GPU. Image generation is already RAM-hungry here; video on top of it will chew through whatever you've got. The community guidance on running video models is "32 GB RAM minimum, 64 GB to be comfortable," and that applies double when the model is streaming through RAM rather than sitting in VRAM. Insufficient RAM pushes you to disk swap, which both thrashes performance and wears out SSDs.

The model license. MiniMax-H3's Community License excludes the US, EU, UK, and South Korea from its applicable territory - check whether the repo is even accessible to you before you point this node at it. LTX 2.5 is the friendlier default if licensing matters.

And the honest closing note: this pack is young and barely charted. Zero downloads chatter, one author's roadmap (ControlNet, LoRA support, more video models) still in the TODO. If it works, you've done something genuinely unusual - video generation on a card that has no business running it. If it fights you, you're not wrong about the tool being rough; that's what "0 impressions" means.