CLIP Vision Encode

"When your image needs to speak fluent CLIP, this is the translator."

The CLIP Vision Encode node is a powerful utility that encodes an image into CLIP Vision's latent embedding space. This allows you to leverage CLIP’s understanding of visual content for all kinds of fun and/or chaotic AI generation tasks — from style transfer and similarity search to image-to-image guidance and multi-modal workflows.

This node takes your input image, compresses it via a VAE (if needed), optionally augments it, and runs it through a CLIP Vision model to output image embeddings — both positive and negative — plus latent samples if you need to push things further.

Essential for AI artists who don’t just want pretty pixels — they want semantically meaningful ones.

🧩 Node Inputs

Input	Type	Description
clip_vision	CLIP Vision	The loaded CLIP Vision model. Must be initialized with a CLIP Vision Loader. This is the core encoder that makes the magic happen.
init_image	IMAGE	The image you want to encode. Garbage in, garbage embeddings out — high-quality images are highly recommended.
vae	VAE	Variational Autoencoder used to encode the image into latent space. Required for downstream workflows that operate in latent format.
width	INT	Width (in pixels) to resize the image before encoding. Must match model expectations. Default is typically 512.
height	INT	Height (in pixels) to resize the image before encoding. Use the same guidance as width.
video_frames	INT	(Optional) Number of frames to generate for video workflows. Only needed if you’re encoding frame sequences.
motion_bucket_id	INT	(Optional) Identifier used for motion sequence conditioning in video tasks. Groups sequences together.
fps	INT	(Optional) Frames per second metadata for video playback.
augmentation_level	FLOAT (0.0–1.0)	Controls how much noise/augmentation is applied during encoding. Helps improve generalization. Too high and you’ll just confuse the model.

🎯 Node Outputs

Output	Type	Description
positive	LIST	Embeddings and metadata representing positive conditioning. These are used to guide generation toward your input image’s features.
negative	LIST	Embeddings and metadata representing negative conditioning. Great for telling the model what not to do.
samples	LATENT	Latent space tensor derived from the input image. Feed this into other nodes for diffusion, transformation, or image generation.

🛠️ Recommended Use Cases

• Image-to-text alignment: Use embeddings to match images with prompts.
• Image-guided generation: Feed the positive output into workflows where you want image features to guide the result.
• Reference style matching: Encode the "vibe" of an image and apply it elsewhere.
• Training augmentation: Use augmentation_level to simulate variation in the same input for robust downstream training.
• Video frame encoding: Turn sequences of frames into CLIP embeddings for multi-frame workflows.

⚙️ Usage Tips

• Keep your init_image clean and high-quality. CLIP is good, not psychic.
• Use matching dimensions for width and height. Usually divisible by 8 or 64 — 512x512 and 768x768 are safe bets.
• For multi-modal workflows, try moderate augmentation_level values (0.2–0.4). Higher values inject useful noise but can distort the image’s intent.
• When using video-related inputs, ensure video_frames > 1 and your CLIP model can handle batch inputs.
• Not using video? Leave video_frames, motion_bucket_id, and fps at default or 0. They won’t hurt anything.

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

• ❌ Feed in mismatched resolutions: If the width and height don’t play nicely with your CLIP model, expect crashes or distorted embeddings.
• ❌ Skip the VAE: You need it for anything latent-related. No VAE, no samples.
• ❌ Abuse augmentation: Setting augmentation_level to 1.0 will mangle your image into unrecognizable mush. Not ideal unless you’re trying to invent glitchcore.
• ❌ Use negative embeddings without understanding them: Negative outputs are not the opposite of positive — they’re meant for contrastive use in workflows that support them.

⚠️ Known Issues

• VRAM usage can spike if you feed in large image resolutions or lots of video frames. Resize before encoding if needed.
• Some CLIP Vision models are picky — make sure the one you loaded supports the resolution you’re using.
• Augmentation noise isn’t standardized. What’s “moderate” for one model might be “absolute chaos” for another.

🧪 Example Node Setup

{
  "clip_vision": "ViT-H-14-CLIP-Vision",
  "init_image": "reference_image.png",
  "vae": "vae_model.safetensors",
  "width": 512,
  "height": 512,
  "video_frames": 0,
  "motion_bucket_id": 0,
  "fps": 0,
  "augmentation_level": 0.3
}

This setup encodes a 512x512 image into CLIP Vision space using a standard VAE with light augmentation.

📝 Notes

• This node plays beautifully with DualCLIP, Prompt Conditioners, and KSampler workflows that take positive/negative embeddings.
• If you're doing anything with image + text matching, this node is practically a requirement.
• This is not a text encoder. Only visual embeddings here, folks.