(Down)Load CLIPVisionModel

If you landed here you're almost certainly wiring up a DynamiCrafter workflow, hit a node called (Down)Load CLIPVisionModel, and want to know what to pick. Short answer: leave it on the CLIP-ViT-H-fp16.safetensors default, run the graph once, and let it fetch the file. That's genuinely it. But since you're here, let's cover what it actually does and where people trip.

What it is

DynamiCrafter is an image-to-video model - you feed it a still image and it animates it into a short clip, and it can also do frame interpolation (start frame → end frame) and looping video. It came out of CUHK and Tencent AI Lab back in late 2023, one of the early open image-to-video models alongside SVD and AnimateDiff. kijai - the guy behind basically every "Wrapper" node pack in ComfyUI, from WanVideoWrapper to the Hunyuan and CogVideoX ones - wrapped it so you can run it as nodes instead of a research repo.

Here's the thing the model needs to work: it doesn't just look at your raw pixels. It reads your input image through a CLIP vision encoder, which turns the picture into an embedding - a compact numerical description of what's in it - and conditions the video generation on that. Same family of image encoder that IP-Adapter uses to do image prompting; DynamiCrafter uses it to understand the frame it's about to animate. This node's whole job is to load that encoder and hand it to the sampler.

How it works

The "Download" in the name is the useful part. kijai's DownloadAndLoad* nodes auto-fetch their weights from Hugging Face the first time you run them, drop the file into ComfyUI's clip_vision models folder, and load it into memory. No manual downloading, no hunting for the right filename. First run is slow while it pulls a couple of gigabytes; every run after is instant because the file's already on disk.

The inputs and outputs that matter

There's exactly one input to set, and one output:

model - a dropdown with two choices: CLIP-ViT-H-14-laion2B-s32B-b79K.safetensors (the full-precision OpenCLIP ViT-H, ~3.9GB) and CLIP-ViT-H-fp16.safetensors (the same thing at half precision, ~2.5GB, the default). For this job the fp16 version is indistinguishable in output and lighter on disk and VRAM - just use the default. The only reason to grab the fp32 one is if you already have it cached from another workflow.
clip_vision - the output, a CLIP_VISION object. This wires straight into the DynamiCrafter loader/sampler node's clip_vision slot. That's its only destination.

That's the entire node. No hidden knobs.

Installing the pack

Two ways, both from the README:

ComfyUI Manager - open it, search ComfyUI-DynamiCrafterWrapper, install, restart.
Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-DynamiCrafterWrapper, then pip install -r requirements.txt inside the folder (portable users: python_embeded\python.exe -m pip install -r ...), and restart.

kijai flat-out recommends xformers - without it the memory usage for decoding is brutal (pip install xformers --no-deps). And this CLIP vision file is not the only download: the actual DynamiCrafter checkpoint goes in models/checkpoints (grab the bf16 safetensors conversions from kijai's Kijai/DynamiCrafter_pruned HF repo), and the refactored nodes also want a regular CLIP text model. This node only handles the vision-encoder piece.

Where people get burned

The auto-download fails behind a firewall or on an air-gapped box. If the first run can't reach Hugging Face, drop the matching .safetensors into your clip_vision folder by hand and it'll load it locally.

Confusing this with the text CLIP. kijai's refactor deliberately splits things - DynamiCrafter now uses a clip_vision model and a separate clip (text) model, and they're not interchangeable. The CLIP_VISION output here feeds only the vision slot; wiring it into a text-CLIP input just errors. If your graph complains about a type mismatch, check you haven't crossed the two.

Expecting this node to fix a broken generation. It's an encoder loader, nothing more. If your video looks bad, that's the sampler, the checkpoint, or your resolution - not this. Honestly, for a fresh image-to-video project in 2026 you'd probably reach for Wan or LTX first; DynamiCrafter is worth it mainly for its interpolation and looping modes, which still hold up.