DynamiCrafterLoader

This node does exactly two things, and one of them is where every DynamiCrafter render-time decision gets made. It loads the base image-to-video checkpoint, and it hands that model - plus your chosen resolution and frame count - to the DynamiCrafter-Simple sampler. That's the whole job. No secrets, no settings menu worth fearing.

Why you'd reach for it: it's the boring half of a two-node video workflow. The interesting part is the sampler that actually moves your image, but you can't get there without this loader, and the frame_length value you set here is the single biggest lever on how long your generation takes. The README's own numbers on a 4090: 16 frames is about 3 minutes, 32 frames about 6. Read that twice before you crank the slider.

What it's loading

The "1024" checkpoint is the main DynamiCrafter model from the Doubiiu lab's 2024 paper (it hit arXiv in December 2023 and landed at ECCV 2024). Mechanically, it's a video diffusion prior - an SD-style UNet with temporal attention - that's conditioned through a frozen CLIP ViT-H image encoder (laion/CLIP-ViT-H-14-laion2B-s32B-b79K). That encoder is why the pack quietly downloads that specific OpenCLIP checkpoint the first time you run it, and it's why DynamiCrafter is fundamentally about animating an image you already have rather than conjuring video from nothing.

The two inputs that matter

• resolution - 576_1024 (default), 320_512, or 256_256. The default is portrait and looks the best, but it's also the most expensive. If your GPU is modest, the pack makes it easy to drop to 320_512 or square 256_256.
• frame_length - default 16. This is the one you'll actually change. More frames equals a longer clip and a much longer wait; the 32-frame setting is the README's headline endurance test.

There are no optional inputs to discover later. It outputs a single model object of type DynamiCrafter, which you wire straight into DynamiCrafter-Simple.

Installing it

The whole pack installs together - you don't get this node without its siblings:

cd ComfyUI/custom_nodes
git clone https://github.com/chaojie/ComfyUI-DynamiCrafter
# then restart ComfyUI

Or use ComfyUI Manager and search for ComfyUI-DynamiCrafter. The README is insistent on one extra dependency: pip install xformers - it's not optional here, the pack expects it. Models auto-download on first run; if you'd rather place them yourself, the base checkpoint goes to ComfyUI/models/checkpoints/dynamicrafter_1024_v1/model.ckpt (the folder name matters, see below).

Honest verdict

This model is from the exact moment before Wan and Hunyuan made U-Net image-to-video feel ancient - community mentions cratered after mid-2024 and never recovered. If you googled "comfyui image to video," you'd be better served by a Wan workflow. But if you're following an older workflow or want the specific DynamiCrafter look, the loader still works exactly as written. Its interpolation sibling, DynamiCrafterInterpLoader, is the one with a genuinely current use case.