IP-Adapter (Image Prompt Adapter) lets a reference image act as part of your prompt. It extracts visual features — a face, an art style, a character's look — from that image and feeds them into generation alongside, or instead of, your text prompt.
A text prompt is limited to what you can describe in words. IP-Adapter sidesteps that by encoding a real image with a CLIP Vision model, then injecting those visual features as extra conditioning — which is why it's a go-to tool for keeping the same character's face or a consistent art style across multiple generations.
Where You'll See It
IPAdapter Unified Loader loads the IP-Adapter model, a Load CLIP Vision node reads the reference image, and Apply IPAdapter combines them with your MODEL before it reaches KSampler.
Quick Example
IP-Adapter FaceID uses InsightFace to extract precise facial features from a reference photo, then reproduces that same face across new poses, outfits, and settings — a common approach for character-consistency workflows.
IP-Adapter models are tied to a specific base architecture (SD1.5, SDXL, and so on) and a matching CLIP Vision model — mixing an SDXL IP-Adapter with an SD1.5 CLIP Vision model, or the wrong checkpoint, is a common source of broken or ignored results.
Common Mix-Up
IP-Adapter vs. ControlNet: ControlNet enforces structural layout — a pose skeleton, a depth map, an edge outline. IP-Adapter carries over visual style and identity — a face, a color palette, an art style — without dictating exact composition. Many workflows use both together for different jobs.
Yes — IP-Adapter needs a CLIP Vision model (a separate download from your regular text CLIP) to read the reference image. The Load CLIP Vision node handles this, and it must match the IP-Adapter model’s expected CLIP Vision variant.
FaceID is a specific IP-Adapter variant built around InsightFace, a face-recognition model, so it extracts and reproduces facial identity much more precisely than the general-purpose IP-Adapter models, which capture overall style and content more loosely.
Yes, and it’s a common combination — ControlNet holds the pose or composition steady while IP-Adapter carries over a face, character, or style from the reference image, letting you control structure and appearance independently.
See It In Action
Ready to keep a character consistent?
Our character-consistency guide covers IP-Adapter setup with real reference images.
