Skip to main content

DualCLIPLoader

The DualCLIPLoader node is here to settle the age-old argument of which CLIP is better — by just using both. This node lets you load two CLIP models simultaneously, giving your workflow extra perspective, better text-vision understanding, and more power than a single CLIP could ever offer alone.

Whether you're running SDXL, SD3, Flux, or even Hunyuan Video, this node helps unlock the full potential of multi-CLIP workflows in ComfyUI.

DualCLIPLoader

🔧 Node Type​

DualCLIPLoader
Outputs: CLIP

📁 Function in ComfyUI Workflows​

The DualCLIPLoader node loads two distinct CLIP models into a single combined output object, which downstream nodes (like CLIPTextEncode, KSampler, and others) can use for better prompt conditioning. It’s especially useful in workflows where you're combining:

  • Vision + text encoders (e.g., SDXL-style)
  • Two stylistically different CLIPs (like realism and anime)
  • CLIPs trained on different languages or datasets
  • Experimental comparative generation setups

It’s a smart choice when you want more control, more nuance, and less guesswork in how your prompt is interpreted.

🧠 Technical Details​

This node:

  • Loads two .safetensors CLIP models
  • Outputs a combined CLIP object
  • Supports architecture-specific types (sdxl, sd3, flux, hunyuan_video)
  • Can optionally be assigned to specific devices (like CPU or cuda:0) for advanced load balancing

Internally, the node makes sure both models are loaded properly and tied to your workflow’s current recipe.

⚙️ Settings and Parameters​

🔲 Field💬 Description
clip_name1Filename of the first CLIP model. This is your primary encoder, typically used for prompt conditioning. Must be a valid .safetensors CLIP file in your models folder. Example: clip_l.safetensors.
clip_name2Filename of the second CLIP model. Used in tandem with the first — often a vision encoder or stylistic variant. Example: clip_vision_g.safetensors.
typeModel recipe this CLIP pair is being used with. Must match the type of your checkpoint and downstream pipeline. Options include: sdxl, sd3, flux, hunyuan_video.
device(Optional) The device you want to load the CLIP models onto. Use "default" for automatic assignment (usually GPU), or set to "cpu" or "cuda:0" if you’re managing memory manually.

✅ Benefits​

  • Multi-modal power – Combine language and vision understanding in one pass.
  • Style fusion – Blend the strengths of two different CLIPs (realism + anime, anyone?).
  • More accurate prompts – Dual interpretation gives better grounding to both positive and negative prompts.
  • Better SDXL/SD3 results – These architectures were made for dual-CLIP setups, and this node is the plug-in brain for them.

⚙️ Usage Tips​

  • Always match the type field with the checkpoint type you're using. Don’t mix sdxl with flux or your generation will go sideways.
  • If you’re running low on VRAM, offload one model to cpu by setting device to "cpu" — but expect slower performance.
  • You can mix a vision CLIP with a language-focused CLIP to create crazy accurate visual storytelling. Try it with SDXL for best results.
  • Want to experiment with prompting style? Use one CLIP trained for realism, and another trained for fantasy — balance prompt weights accordingly.

📍 ComfyUI Setup Instructions​

  1. Place your CLIP model files (.safetensors) in your ComfyUI models/clip folder.
  2. Add the DualCLIPLoader node to your workflow.
  3. Set clip_name1 and clip_name2 to the exact filenames.
  4. Set the type field to match your pipeline (sdxl, flux, etc.).
  5. Optionally assign a device (or just leave it as "default").
  6. Connect the output to any node that expects a CLIP input.

📎 Example Node Configuration​


clip_name1: clip_l.safetensors clip_name2: clip_vision_g.safetensors type: sdxl device: default

In this setup, we’re using two different CLIPs with an SDXL-based model. This is ideal for workflows where SDXL’s text+vision conditioning is leveraged for higher fidelity generations.

🔥 What-Not-To-Do-Unless-You-Want-a-Fire​

  • ❌ Don’t mix model types. Your type must match your checkpoint or you’ll get garbage output (or no output at all).
  • ❌ Don’t assign both CLIPs to GPU on low-VRAM systems. You will crash.
  • ❌ Don’t try to load non-CLIP .safetensors files here. It won’t work. You’ll sit there wondering why your workflow’s frozen.
  • ❌ Don’t assume CLIPs “merge” into a single model — this node runs them in parallel, not as a fusion.
  • ❌ Don’t use massive CLIPs on a 4GB GPU unless you really enjoy watching your system swap memory like it’s 2006.

⚠️ Known Issues​

  • VRAM hungry – Two models = more memory. No surprise here.
  • Slow on CPU – If you offload to CPU, expect a noticeable slowdown.
  • No CLIP validation – If your filename is wrong, the node won’t tell you nicely — it’ll just break silently or downstream.

📚 Additional Resources​

  • How SDXL Uses Dual CLIP Architecture
  • Example model files:
    • clip_l.safetensors
    • clip_vision_g.safetensors

📝 Notes​

  • This node is highly recommended for SDXL workflows and is borderline required if you want to use SDXL the way it was meant to be used.
  • Great for prompt tuning, multimodal alignment, and style fusion.
  • If you're experimenting with new CLIPs, try this node in a sandbox workflow before plugging it into a production chain.