PhotoMaker and InstantID
PhotoMaker and InstantID both solve the same problem — keeping the same face across many AI portraits — but they take different routes. PhotoMaker fuses multiple reference photos into one identity embedding that conditions generation. InstantID uses a face encoder plus an adapter to lock identity from a single image. This guide picks sensible reference counts and strength values based on how faithful you want the result to be.
How each tool works
PhotoMaker
PhotoMaker injects identity by stacking reference images into a single augmented embedding. Each reference image is processed by a CLIP-like image encoder and merged into the text embedding space using a special trigger word (commonly img). The embedding carries identity information across multiple samples of the same person, so varied references — different angles, lighting conditions, expressions — produce a more robust and generalisable identity representation than a single photo.
The main parameter is the style strength percentage, which controls how much the identity embedding influences the generation relative to the text prompt. High style strength preserves likeness; lower values allow the model to deviate from the reference and apply stylistic transformations more freely.
Trigger word placement matters: a photo of a img man places img directly before the class noun, which is the recommended pattern for binding the identity embedding to the subject.
InstantID
InstantID takes a different approach: it uses a dedicated face encoder (similar to IP-Adapter Face) combined with a ControlNet-style adapter that applies face landmark conditioning. This allows tight identity locking from a single image without needing multiple references.
The key parameters are:
- Identity scale — how strongly the face encoder’s output conditions generation (0–1.5, often best around 0.8–1.0)
- ControlNet conditioning scale — how closely the face structure from the reference is enforced (0–1.5)
Higher values on both produce strong likeness but can make the face look rigidly transplanted. Lower values allow the generation model more creative latitude while still preserving recognisable features.
Choosing the right tool for your use case
| Use case | Better choice |
|---|---|
| You have 3–4 reference photos | PhotoMaker |
| You have only 1 clear reference | InstantID |
| Strong style transfer (cartoon, painting) | PhotoMaker with lower style strength |
| Hard identity lock, realistic portrait | InstantID with high scales |
| Combine both for maximum quality | InstantID for identity + PhotoMaker or IP-Adapter for style |
How the guide sets recommended parameters
The two main dials are identity preservation and style strength. High identity preservation forces the output to match the reference face closely, which is great for “the same person” consistency but can look pasted-on if pushed too far. Lower preservation lets the model restyle the face into a new art style, age, or mood. The guide maps your chosen balance to the specific parameters each tool exposes — PhotoMaker’s style-strength percentage and trigger word placement, or InstantID’s identity and ControlNet conditioning scales.
Practical tips for faithful portraits
- Feed PhotoMaker variety. Three to four photos from different angles beat one perfect headshot — the embedding generalises instead of overfitting to one expression.
- Keep InstantID references sharp and frontal. The face encoder needs a clear, well-lit, mostly frontal face to extract a strong identity vector. Blurry, three-quarter, or heavily shadowed references reduce quality significantly.
- Don’t max out strength. Identity at 100% or 1.5 scale often looks rigid and pasted-on. Backing off 10–20% lets the model blend the face into the scene’s lighting and pose naturally.
- Put the PhotoMaker trigger word right before the class noun, e.g.
a photo of a img man— trigger word placement affects how strongly the identity embedding binds to the subject in the generation. - When combining tools in ComfyUI, keep total identity influence moderate — if PhotoMaker and InstantID both push at full strength simultaneously, they fight each other and the output degrades. Treat them as complementary layers, not additive amplifiers.