DynamiCrafter Simple

This is the node that actually does the thing. DynamiCrafter-Simple is the sampler half of this pack's base workflow: you give it a starting image, a prompt describing the motion you want, and a few sampling settings, and it hands back a clip. The "Simple" in the name is doing real work - there's no noise-prediction gymnastics, no elaborate conditioning stack. Model in, image in, video out. That's the whole DynamiCrafter pitch, and it's why the pack exists.

If you've ever wished a still image would just move, this is the 2024 version of that wish. The load-bearing caveat, which I'll get to at the end: it's a 2024 wish, and video moved on.

How it works

DynamiCrafter comes from the Doubiiu lab's 2024 paper (arXiv December 2023, ECCV 2024). The architecture is a video diffusion prior - an SD-style UNet with temporal attention - conditioned by a frozen CLIP ViT-H image encoder. Your start frame gets encoded into visual tokens that steer the denoiser, and the text prompt rides along as extra guidance. That encoder is why the pack downloads laion/CLIP-ViT-H-14-laion2B-s32B-b79K on first run, and it's why the image input is the star here: this model animates an existing image, it doesn't hallucinate video from a prompt alone.

Inputs a beginner actually sets

• image - your start frame. Generate a still first with any checkpoint, wire it in, and let the model push it into motion. This is required, not optional.
• prompt - describe the motion: "sleeping", "hair blowing in the wind", "camera slowly panning up". The README's examples are exactly this kind of verb-led prompt.
• steps (50) and cfg_scale (7.5) - standard sampler settings. Higher CFG means tighter prompt adherence and more artifact risk; the 7.5 default is the sane middle.
• eta (1.0) - DDIM stochasticity. Fun fact: 1.0 isn't a typo, it's what the DynamiCrafter authors themselves used at inference.
• motion (3) - how energetically things are allowed to move. Low values give a subtle warble; cranking it invites drift and distortion.
• seed (123) - reproducibility. Change it when results feel stale.

The output and what you do with it

The single image output is the whole clip as a batched IMAGE tensor - frame 0 is your still, the rest are the motion. Wire it into something that encodes video, like VideoHelperSuite's VideoCombine or ComfyUI's core Save Animated WEBP, to get a file you can actually watch.

Install, and the gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/chaojie/ComfyUI-DynamiCrafter
pip install xformers   # the README insists; the pack expects it
# restart ComfyUI

Or grab it from ComfyUI Manager under ComfyUI-DynamiCrafter. Models auto-download on first run; manually, the base checkpoint lives at ComfyUI/models/checkpoints/dynamicrafter_1024_v1/model.ckpt - the folder name has to match that exact resolution-based convention or the loader won't find it.

Where people get burned: expecting prompt-only text-to-video (the image input is mandatory), and underestimating the cost. At 576_1024 and 16 frames, a 4090 takes about three minutes - at 32 frames, double that. On weaker hardware, drop to 320_512 or 256_256 in the loader and cut the frame count. And if anything fails to load, xformers is the first suspect.

The honest take

DynamiCrafter is a museum piece now. Community activity died off after mid-2024, and if you just want image-to-video, Wan 2.x or Hunyuan will beat it in quality and speed on the same GPU. Install this pack because you're following an old workflow, want that specific early-I2V look, or need its interpolation sibling. For that last one, keep reading.