How Many Reference Photos Do You Need for a Consistent AI Character?

How many reference photos does OFMAI need?

Three. A character is built from 3 reference photos — a portrait, a front view and a back view — with no training step, and it is ready in minutes. That one character is then reused in every generation: photos, short videos, voice notes and image edits all carry the same look, body and morphology.

Three is enough because nothing is trained. A training run infers an identity statistically from many examples, which is why dataset size matters to it; OFMAI stores the character instead, with the three photos anchoring the likeness. Creating one costs 12 credits: 6 for the batch of candidate portraits and 6 to confirm your pick and generate the front and back views. It runs in the browser — no GPU, no install, no model setup.

How many photos does training your own character model need?

Between 10 and 100, depending on the base model. A ComfyUI LoRA tutorial of 28 August 2026 recommends 15 to 30 images for a character and notes that "about 10 excellent images can sometimes work" (eastondev.com). The diyai.io guide of 23 October 2026 gives the same 15 to 30 for SDXL-based models and 30 to 100 for DiT-based ones, warning that below 30 the adapter does not generalise and above 100 you mostly add training time. Scenario, a hosted trainer, puts the sweet spot at 5 to 15, blocks training below 5 and caps the set at 25 to 100 by plan (read in October 2026).

What none of these numbers say is how long the photos take to collect. A character who does not exist yet has none, so the first job becomes generating 15 to 30 consistent images of a person who keeps changing — the problem the training was meant to solve.

How many reference images do general image generators take?

Usually one, and it is not training. Midjourney's character reference accepts several references in a prompt but its documentation says you "probably don't need to" use more than one, that it works best on Midjourney-generated images of a single character, and that it is not designed for photographs of real people and will distort them (Midjourney documentation, read in October 2026). One image steers one generation — the reference has to be passed again next time, and the person shifts a little each time. A reference steers a single image; a character is stored once and reused.

What makes a reference photo usable?

Coverage and sharpness, not volume. Scenario advises the same person in different poses, expressions and outfits and prefers a small curated set to a large inconsistent one (read in October 2026); diyai.io, 23 October 2026, says to vary angles, expressions and lighting and keep one character per image. Resolution has a floor: match the base model, 512×512 for SD1.5 and 1024×1024 or higher for SDXL (eastondev.com, 28 August 2026). The same logic applies to OFMAI's three photos, plus one rule: they must be of you, or of someone who has explicitly authorised their use.

What if the character does not exist yet?

Then the photo count is zero: a character on OFMAI can be designed from a text description instead of photos, and the result is the same reusable character. The training route solves it the long way round — diyai.io, 23 October 2026, covers it under "If your character only exists in your head": generate the reference images first, then train on them. Cloning a real look rewards a dataset; inventing one rewards starting from a description.

What does each route cost in time and money?

Training costs hardware and a wait. The ComfyUI tutorial of 28 August 2026 puts the starting VRAM at roughly 8 GB for SD1.5 and 12 GB for SDXL, and suggests 1,000 to 2,500 steps for a character; diyai.io, 23 October 2026, reports 15 to 30 minutes for an SDXL adapter on a hosted trainer and over an hour for a DiT one — before you collect and caption the dataset.

OFMAI costs credits and no wait: 12 for the character, 2 for a photo, 2 for an image edit, from 7 for a short video. Plans start at $9.99 a month for 30 credits, about $0.67 a photo, and annual billing saves 25%. Credits bought as a top-up never expire, and a failed generation is refunded automatically.

When is training your own model the right route?

When you need to own the weights, or when what you are after is not a person: a file that runs offline on your own GPU with no platform in the loop, a style or an outfit family rather than one individual, or a large clean archive of the same person you already hold. The diyai.io guide of 23 October 2026 sets its own bar at about a dozen or more good images of the character already in hand.

OFMAI is not a local model or a downloadable checkpoint — it is a hosted web app. If full offline control is the requirement, running ComfyUI or Forge locally is the honest answer, and you manage the GPU, the node graphs and the model updates yourself.

Photos needed, route by route

• OFMAI from photos - 3 photos (portrait, front, back). No training, ready in minutes, 12 credits, runs in the browser.
• OFMAI from a description - 0 photos. No training. Same reusable character across photos, video and voice.
• Character reference in a general image generator - 1 image per prompt, passed again every time. No training, no stored identity (Midjourney documentation, read in October 2026).
• LoRA on an SDXL-based model - 15 to 30 images, 512 to 1024 pixels, a GPU with roughly 12 GB of VRAM (eastondev.com, 28 August 2026; diyai.io, 23 October 2026).
• LoRA on a DiT-based model - 30 to 100 images; below 30 it does not generalise, above 100 adds training time (diyai.io, 23 October 2026).
• Hosted trainer - 5 to 15 images, 5 minimum, 1024×1024 recommended, capped at 25 to 100 by plan (Scenario help centre, read in October 2026).