Consistent characters across generated images

Why seeds are not the answer

The usual first attempt is to fix the seed. That reproduces an identical image for an identical prompt, but change one word — the setting, the outfit, the pose — and the face drifts, because the seed controls the noise, not the identity.

What you actually need is to condition generation on a specific face. Two mechanisms do that: an adapter that reads a face embedding from a reference photo, or a small set of trained weights that teach the model the person.

Option 1: Face ID, from a single photo

Fastest path. Post one clear photo alongside your prompt and get back the same person in that scene. Nothing to train, nothing to store, and it works the first time a user uploads a selfie.

The reference drives the face only — describe the rest of the person in the prompt, because hair colour, build and clothing come from your text, not the photo.

curl https://api.fastgencloud.com/v1/images/face-id \
  -H "Authorization: Bearer $FASTGEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
      "image": "https://example.com/selfie.jpg",
      "prompt": "a woman with dark curly hair in a green raincoat, standing on a harbour wall, overcast light",
      "style": "realistic",
      "identity_strength": 0.9,
      "n": 4
    }'

Option 2: train a character

When the same person recurs — a brand mascot, a user avatar, a series — training gives a much closer likeness that holds up across wildly different scenes. Send 4–30 photos, wait about ten minutes, get an id.

# 1. start training
curl https://api.fastgencloud.com/v1/characters \
  -H "Authorization: Bearer $FASTGEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
      "name": "Ada",
      "class": "woman",
      "base_model": "general-image",
      "images": ["https://…/1.jpg", "https://…/2.jpg", "https://…/3.jpg", "https://…/4.jpg"]
    }'
# → { "id": "chr_0123456789ab", "status": "training", "progress": { … } }

# 2. poll until status is "ready"
curl https://api.fastgencloud.com/v1/characters/chr_0123456789ab \
  -H "Authorization: Bearer $FASTGEN_API_KEY"

Generating with a trained character

Reference the character by id on an ordinary generation call. The server resolves the weights and the trigger token — you never see either.

curl https://api.fastgencloud.com/v1/images/generations \
  -H "Authorization: Bearer $FASTGEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
      "model": "general-image",
      "prompt": "a woman in a linen shirt reading at a cafe table, soft afternoon light",
      "character": { "id": "chr_0123456789ab", "strength": 1.0 },
      "size": "832x1216",
      "n": 2
    }'

Which one to build on

A pattern worth stealing: offer Face ID immediately on upload so the user sees a result within seconds, and kick off training in the background. Swap to the trained character once it is ready.

User uploads a selfie, wants 4 images now

Face ID

Recurring character in an ongoing product

Training

Likeness must be convincing to someone who knows the person

Training

You cannot store biometric-derived assets long term

Face ID

Anime or cartoon output

Either — train against that family, or set style

Getting a better likeness

• Vary the training set — angle, distance, lighting, expression. Ten varied photos beat thirty near-identical ones
• One face per photo, no sunglasses, no heavy filters, no motion blur
• On Face ID, raise identity_strength toward 1.1 for closer resemblance, lower toward 0.7 if faces look stiff or waxy
• Generate several and pick — identity pipelines vary more between seeds than plain text-to-image does
• Likeness is always strongest on the realistic style; anime and cartoon carry a person more loosely by design

Consent, before you ship this

Both features operate on real people’s faces. Generating sexual or otherwise harmful depictions of a real, identifiable person without their consent is prohibited by the Acceptable Use Policy and will end an account.

If you are building a consumer product, get explicit consent at upload, and be clear with users about how long you keep their photos. Outputs here are deleted after 7 days; your own storage is your responsibility.