Load Qwen3.5 ViT (Anima)

This node loads the Qwen 3.5 vision transformer that powers image-reference conditioning on Anima. It exists so you can feed a character or style image into the graph alongside your text prompt, and let the visual content leak into the conditioning. It's the load half of a two-node feature - the encode half is Qwen35VisionEncode.

Straight talk up front: as of writing, the weights this node loads aren't published anywhere you can just download. The pack's HuggingFace repos host qwen35_4b.safetensors, the calibration files and the tokenizer - no qwen35_vit.safetensors. And the extraction script the code points you to, extract_vit.py, isn't in the GitHub repo either. The node's file picker is effectively a stub until the weights land, or until you extract them yourself from a Qwen 3.5 checkpoint if you're comfortable doing that. Don't feel dumb when you can't find the file; this pack shipped all three versions in a single day with the vision half still on the bench.

What it does when it has a file

It's a deliberately thin loader. One required input, vit_name, which scans the custom node folder and ComfyUI/models/text_encoders for a file with "vit" in the name (the source checks the node's own directory first, then the text_encoders folder). One output, vit, of type QWEN35_VIT, which wires straight into the vit input of Qwen35VisionEncode - the two always travel together.

The mechanism underneath is simple: the ViT turns a reference image into patch tokens, and those get injected into the text encoder's token sequence. Why bother when Anima already takes images for img2img? Because this is reference conditioning - the visual content becomes part of the prompt the diffusion model attends to, rather than the starting noise. It's the closest thing Anima has to a ControlNet-style reference feature, and Anima has no ControlNet at all.

Install

Same as the pack's CLIP loader:

cd ComfyUI/custom_nodes
git clone https://github.com/GumGum10/comfyui-qwen35-anima

Or search comfyui-qwen35-anima in ComfyUI Manager. Restart, and the node appears under loaders/Anima. There's nothing extra to pip-install - the pack adds no Python dependencies. The missing piece isn't installation, it's the .safetensors file itself, which means the practical next step is: add this node, watch it list no real choice, and keep an eye on the repo for the weights.

When it's worth your time

This is a "bookmark it" node right now rather than a "use it today" one. If the author publishes the ViT weights, it becomes the input half of a genuinely useful trick: style and character transfer on a model with no pose/depth ControlNet. Until then, treat the Qwen35VisionEncode page as the feature's design doc and this node as the thing that's waiting on its payload.