Reference images: IP-Adapter, Redux, editing models
Using a picture as an example of a style, a subject or a face, rather than as a starting point.
In plain words
In image to image the picture is the start that gets redrawn. With a reference image the picture is a hint: "a character who looks like this", "in this style", "this product on a new background". The model encodes the reference and lets it steer the result, while the composition can be entirely new.
The ways to do it
| Way | Families | How it enters the graph |
|---|---|---|
| IP-Adapter | SD 1.5, SDXL (and others through packs) | A pack (for example ComfyUI_IPAdapter_plus) with an IP-Adapter model and a CLIP Vision encoder; applied to the MODEL with a weight |
| Flux Redux | Flux.1 dev | Built-in: Load CLIP Vision, CLIP Vision Encode, Load Style Model, Apply Style Model on the conditioning |
| Editing models | Qwen-Image-Edit, FLUX.1 Kontext, Flux.2 | The reference goes into the conditioning (for example ReferenceLatent or TextEncodeQwenImageEdit) and an instruction says what to change |
| CLIP Vision for video | Wan 2.1 image to video, Wan Animate, talking portraits | Load CLIP Vision (clip_vision_h) and an encode node give the model the look of the start picture |
| VACE reference | Wan with VACE | ref_images input keeps a look through the clip |
Values and what they change
- The weight or strength of the reference: too high copies the reference; too low ignores it. Start near the template's value.
- A clean reference (one subject, plain background) works best.
- The prompt still matters: say what should change.
Typical mistakes
- A CLIP Vision model or IP-Adapter for another family.
- Using a real person's face without consent: see Content rules.
- Expecting an editing model to work in an ordinary text to image graph: use its template.
How the app helps
The built-in knowledge describes Load CLIP Vision and CLIP Vision Encode, and its entry "Identity: adapter, LoRA or swap: which to choose" compares the ways to keep a face. The lessons: A reference picture with IP-Adapter, A reference picture for style (Flux Redux), Editing a picture by words, The same character in a new picture: FLUX.1 Kontext, One character in many pictures, The same face from one photo: IP-Adapter FaceID, Put a face onto a picture: ReActor, and for video A reference for the whole clip: Wan VACE. The lessons that use a real person's face ask for that person's consent.