Qwen2.5-VL CLIP Loader

Before any of the Qwen-Image-Edit encoder nodes in this pack can do anything, they need a text encoder - and for the Qwen line that encoder is Qwen2.5-VL, a vision-language model. QwenVLCLIPLoader loads it and hands it out on ComfyUI's standard CLIP wire, so the rest of your graph can treat this multimodal LLM like any other text encoder. It's the loader you start the Qwen branch of a workflow with.

The naming is a bit of a fib, and it's worth clearing up: this is not CLIP. Qwen2.5-VL is a full vision-language transformer. It's presented as a CLIP purely so it plugs into ComfyUI's existing conditioning sockets. That "-VL" matters - the vision half is why Qwen-Image-Edit can look at your input image and reason about it, not just read your words.

How it works

Qwen-Image-Edit dual-encodes: your input image goes into Qwen2.5-VL for semantic understanding and into the VAE for appearance, at the same time. This loader supplies the Qwen2.5-VL half. Under the hood it's reading a language model's weights and wrapping them so downstream encoder nodes (QwenVLTextEncoder, QwenVLTextEncoderAdvanced) can push your prompt through the chat template the model expects. Everything the model knows about your instruction, it knows through this component - which is exactly why swapping to an LLM encoder changed prompting behavior so much across the whole 2026 generation of models.

There's not much to configure, which is the right amount for a loader.

The inputs and outputs that matter

• model_name (required) - a dropdown of the Qwen2.5-VL weights it finds on disk (the pack expects the 7B, e.g. qwen_2.5_vl_7b.safetensors). Drop the file in your text-encoder models folder and pick it here.

Output: clip - a single CLIP wire. Feed it into any of the pack's Qwen encoder nodes. That's the whole node.

How to install it

ComfyUI Manager: search ComfyUI-QwenImageWanBridge, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/fblissjr/ComfyUI-QwenImageWanBridge

then restart. The real work is getting the model file: you need the Qwen2.5-VL-7B weights in your ComfyUI text-encoder directory. The full-precision encoder is heavy - it's a 7B VLM - so on a tight card, grab a quantized (GGUF or fp8) build. Encoder quantization is loaded separately from the diffusion model, and how much quality it costs is genuinely disputed; test it rather than assuming it's free.

Common issues & troubleshooting

The dropdown is empty or your file isn't listed. ComfyUI only shows what's in the expected models folder. Confirm the Qwen2.5-VL safetensors is in the text-encoder/clip models path and restart so it re-scans.

You loaded Qwen3 by mistake. Different job. Qwen3-4B is the Z-Image encoder; Qwen2.5-VL is the Qwen-Image-Edit (and HunyuanVideo) encoder. They're not interchangeable - feeding the wrong one downstream will fail or produce nonsense. This loader is specifically the VL model.

It eats your VRAM. A 7B vision-language encoder is a big chunk on top of the diffusion model - on 2026 models the encoder is often what makes a workflow not fit, not the model. If you're OOMing, quantize the encoder first before touching anything else.