What training an AI on your face means
Mostly, you do not need to train an AI on your face anymore. We sell AI face training on Doppelshot, and we think it is on its way out for faces. In our tests, a few selfies used directly as references gave the same likeness as weights trained for 35 minutes. This article shows the measurements, explains why modern AI photo generators made training optional, and covers the few cases where training still earns its place.
Until recently, getting an AI photo generator to draw you meant teaching it your face first. People searching for how many photos to train an AI would find the same recipe on every guide. You uploaded 10 to 30 photos, a trainer ran for 20 to 60 minutes, and you got a small file called a LoRA. A LoRA is an add-on that bends a base photo model toward one identity. If you wanted a LoRA of your face, you had to pick a base model, caption the photos, set the steps and learning rate, test the strength, and fix the waxy skin that appeared when you overtrained.
That recipe still works. The real question is whether you need it. Search trends give a clear hint: worldwide searches for "lora training" fell by about 90 percent between July and October 2026, while "nano banana" and "ai photo generator" stayed among the largest AI image searches, and "nano banana 2.1" became one of the fastest-rising queries. People stopped asking how to run AI face training. They started asking which generator to use. You can also read the short summary in our guide on whether you still need to train an AI on your face.
What we measured
We took the same designed person, the same scenes, and the same face recogniser as a judge on a scale from 0 to 1. Then we compared the three ways to create an AI twin:
- No training (Instant): your selfies guide the photo engine on every picture. Likeness scored 0.62.
- A profile learned in a few minutes (Studio): our Studio option, powered by Phota, learns a face profile in a few minutes. Likeness scored 0.70, the sharpest we measured.
- Weights trained for about 35 minutes (Trained): likeness scored 0.66 on the best trainer, and 0.62 to 0.63 on the others.
The 35-minute training landed right between the two methods that require no long training wait. It was not worse, and it was not better. You can inspect the photos from each method side by side on our AI twin options comparison page.
Why training is fading for faces
Photo models already know human faces very well. They only need to be told which face to draw. The newest engines accept reference photos inside the generation request itself. Google's Nano Banana (both Pro and 2.1), OpenAI's GPT Image 2.5, xAI's Grok Imagine (which edits from up to five source images since its Image 2.0 release in August 2026), Ideogram 4.5 (released on 30 September 2026), and Black Forest Labs' FLUX.3 (released in October 2026, taking up to ten reference images) all work this way. In our runs, Grok Imagine swapped the costume and the scene while keeping the face in 12 to 23 seconds per photo. A handful of selfies now does the job a trained LoRA used to do.
Training also carries a cost beyond the wait and the price. Trained weights have defects of their own. In our runs, trained models sometimes mirrored the face or doubled a beauty mark. Reference-based generation avoids that mistake because the engine looks at the real photo every single time.
The trend only moves in one direction. Each new model keeps a face better from fewer photos. When your AI twin is built from reference selfies, it automatically follows the best new engines as they come out. When your twin is a file of trained weights, it stays tied to the specific model it was trained on.
Where training still earns its place
Doppelshot keeps a Trained option, and we would not keep it if it were useless. Training still wins in three situations:
- You want to own the weights. A LoRA is a portable file. You can download it, store it on your drive, and load it into other tools. Reference selfies live inside a product.
- The subject is not a face. A specific product, a garment, or a custom drawing style that the base model has never seen tends to drift more when using references alone. Training teaches the model the exact object or style.
- You need a consistent character across hundreds of pictures or long videos. If you are building a fictional AI influencer, a brand mascot, or a comic hero that must look identical across hundreds of shots, dedicated training still helps.
Our Trained option runs on Ideogram's trainer, which scored the highest among the trainers we tested. If you just want clear photos of yourself for work or social media, pick the way that needs no training.
How to get an AI twin without training
If you want photos or videos of yourself without waiting half an hour for a trainer to finish, the process takes five steps:
- Take a few selfies. One clear selfie is enough to start on standard photo cards, where a result is ready in about 20 seconds. For a reusable AI twin, 10 to 20 selfies taken on different days and against different backgrounds give a stronger likeness, so the model learns your face rather than the lighting in your kitchen.
- Skip filters, sunglasses, and beauty modes. If you upload a filtered photo, the engine learns the filter too.
- Upload them as references. On Doppelshot, your first preview on a single-selfie card is free with no card required. For a full twin, you can start with an Instant twin with no wait, or let a Studio twin learn a face profile in a few minutes. Every option runs on credits, the price shows before you run, and you can check all packs and monthly plans on our pricing section.
- Describe a scene or pick a pack. You can type any scene you want, or generate a batch of six photos at once with the LinkedIn headshot pack or the Travel photo pack.
- Check the likeness, then keep going. Once you have a photo you like, the same result can step into the video tools. You can make the photo speak a script in your cloned voice, sing a 10-second jingle, move through a camera orbit, or star in a 60-second vertical short drama for TikTok, Reels, or Shorts. Every paid result includes a commercial license with no watermark.
Before you upload your face anywhere
Before you upload selfies to any AI photo generator, ours included, ask three questions:
- Where are my selfies stored, and is there a public link?
- Are my selfies used to train the company's base models?
- When are my photos and face models deleted?
Your face is not a password that you can reset if it leaks. On Doppelshot, uploaded selfies and twins live in a private bucket with no public link. External AI providers receive a photo only to produce the single generation you requested, through a short-lived signed link. Your selfies are never used to train Doppelshot's models. Selfies and twin models are automatically deleted 90 days after their last use, or immediately whenever you delete them from your account. Because Doppelshot is made by AitekLabs, MB in Lithuania, European Union GDPR rules apply in full. You can read every detail in our privacy policy.
FAQ
Do I need to train an AI on my face to get a good likeness?
No. In our tests, selfies used directly as references matched 35-minute trained weights for facial likeness, and a Studio profile learned in a few minutes scored higher than both.
How many photos does it take to get AI photos of myself?
One clear selfie is enough for single photo cards, which finish in about 20 seconds. For a reusable AI twin that you can place into any custom scene or photo pack, upload 10 to 20 selfies taken in varied lighting.
How long does AI face training take compared to reference models?
Training custom weights takes about 25 to 35 minutes. Using reference selfies on an Instant twin requires no wait at all, and learning a Studio profile takes only a few minutes.
Is it legal to use someone else's face in an AI photo generator?
You can use your own face, or the face of another adult who has agreed to appear in AI-generated content. Doppelshot is strictly for adults (18 and over), and photos of minors are never allowed. Every generated result also carries a machine-readable AI marker under the EU AI Act.
