Face Transfer
Face Transfer takes a reference photo of a person and carries their identity into a new generation. The output can be set anywhere — medieval armor, anime cel, cyberpunk neon, Renaissance portrait — and the face still reads as the same person. It's powered by InstantID and runs against the SDXL family of models inside Pocket.
This is the mode to reach for when "looks like this person" matters more than anything else in the image.
#What it does
Face Transfer extracts a facial embedding from the largest face it finds in your reference, then conditions the generation on that embedding. The result keeps the recognizable features — bone structure, eye shape, mouth, overall identity — while the rest of the image is invented from your prompt.
A few specific rules to know up front:
- Only the largest face in the reference is used. Crop or pick a photo where the subject's face dominates.
- The face's size, position, and angle in the reference carry over. If the face is small and off-center in the reference, expect the same in the output. Reposition the source if you need a different framing.
- Body pose is ignored. Face Transfer is about identity, not posture. If you also want pose control, run a separate generation with Face and Pose Capture, or stage the prompt to describe the body.
- Works best with SDXL, SDXL Turbo, and SDXL Lightning models. Pocket steers you toward those automatically when Face Transfer is on.
#Face Transfer vs Face and Pose Capture
The two sound similar, but they capture different things:
| Mode | Captures | Use when |
|---|---|---|
| Face Transfer | Who the face is (identity) | You want the same person in different scenes/styles |
| Face and Pose Capture | How the face is posed (landmarks, expression) | You want a different person to inherit a pose or expression |
You can do "this person, this pose" with two passes or with the desktop ControlNet stack — Pocket runs one ControlNet at a time.
#How to start
- Open Advanced Guidance and tap Face Transfer.
- Pick a reference photo. A clear, well-lit head shot with some 3D lighting (a shadow on one side of the face) works far better than a flat selfie.
- Pick an SDXL family model.
- Write a prompt describing the world, outfit, lighting, and style — anything except the face itself.
- Lower the Guidance Scale to 2–3. Face Transfer prefers a softer prompt influence than vanilla SDXL.
- Generate.
#Prompt tips
- Describe the scene, not the face. The face is locked. Prompt for armor, costume, environment, lighting, mood, genre.
- Style triggers work well. "Anime, Studio Ghibli inspired" or "oil painting, Caravaggio lighting" gives the model a strong target without fighting the face conditioning.
- Adjust ControlNet strength if the face is dominating too much (lower it) or if the model is drifting from the identity (raise it).
- Avoid extreme angles in the reference if you want flexibility in the output — a slight 3/4 angle with shadow gives the best results.
#Workflows
Genre swap. Run the same reference through Cyberpunk, Medieval Knight, Studio Ghibli Anime, and Renaissance Portrait prompts. You get a coherent "character across universes" set.
Avatar set. Generate 8–16 variations with the same prompt and a different seed each, then pick the strongest as a profile picture or character card.
Storybook insertion. Turn a friend or family member into the protagonist of a fairy-tale or storybook illustration. Combine with a stylized prompt for a gift-worthy result.
Cosplay preview. Test whether a costume idea suits a specific person before committing to a real-world build.
#Tips
- The keeper-rate on Face Transfer goes way up with a single clean reference. Five bad references won't beat one good one.
- Photos with even, soft lighting feel safer but produce less convincing identity transfer than photos with directional light.
- Re-roll seeds rather than rewriting the prompt — identity preservation varies more by seed than by phrasing.
- If results look "almost right but uncanny," raise the ControlNet strength a notch.