📥 CLIPVision Download

CLIPVision Download is a one-job fetch node: give it a URL and it downloads a CLIP vision model, then outputs it as a CLIP_VISION object ready to wire into whatever needs one. Its reason to exist is that CLIP vision models are a quiet dependency of a lot of image-conditioning workflows - IP-Adapter, Flux Redux, and friends all need a vision encoder to turn a reference image into embeddings - and this lets a shared workflow carry that dependency as a link instead of a "download this and drop it in models/clip_vision first" footnote.

The inputs are minimal: Download_url, the direct link to the model, and Url_name, the filename to save it under (default model.safetensors). The one output, CLIP_VISION, is the loaded encoder, which you feed into an IP-Adapter apply node, a Redux node, or anything else that asks for a CLIP vision input. That's the entire surface - it's the "load" step with an auto-download attached.

The context worth having is what a CLIP vision model actually is, because matching the right one is the thing that trips people up. A CLIP vision encoder takes an image and turns it into an embedding the diffusion model can be conditioned on - it's how "use this picture as a style/content reference" works at all. Different adapters expect different encoders: the SD 1.5 IP-Adapters were built on OpenCLIP ViT-H-14, the SDXL ones on ViT-bigG-14, and newer Flux-era tools use their own (SigLIP-based, in some cases). The encoder is not interchangeable - an IP-Adapter built for ViT-H will misbehave or refuse if you hand it a bigG encoder. So the value here is convenience of downloading, but the responsibility of picking the correct vision model for your adapter stays with you.

Where it fits in a build: you're setting up image-prompt conditioning, your IP-Adapter or Redux node reports it wants a CLIP vision model you don't have, and rather than leaving the graph to go hunt HuggingFace, you paste the link into this node and keep working. It's plumbing for the reference-image side of a workflow.

Install comes with the pack. ComfyUI Manager, search SDVN_Comfy_node; or cd ComfyUI/custom_nodes && git clone https://github.com/StableDiffusionVN/SDVN_Comfy_node, then pip install -r custom_nodes/SDVN_Comfy_node/requirements.txt from your ComfyUI root and restart. The auto-download uses aria2c, so install it by hand on Windows/macOS; on Linux it's usually already present.

Troubleshooting is short. A failed download is nearly always a bad URL (it wants a direct model file link) or missing aria2c on Windows. The subtler failure is downloading the wrong encoder: if your IP-Adapter or Redux node errors on the CLIP_VISION input, or the reference conditioning does nothing, the odds are you fetched a vision model that doesn't match the adapter's architecture - line up the encoder with what your adapter was trained against (ViT-H for SD 1.5 IP-Adapters, ViT-bigG for SDXL) and it'll click. The node downloads whatever you point it at; it can't tell you the file is the wrong one for your graph, so that check is on you.