The Best AI Model for Photos of Yourself in 2026

Every few months someone asks a version of the same question: I want photos of myself that look like a real photographer took them, so which model should I actually be using?

For most of the last two years the honest answer was "it barely matters, your prompt matters more." That stopped being true in 2026. The gap between models is now largest at exactly the thing this site is about — generating a photorealistic person who still recognisably looks like you, across more than one image.

Here is where the four families that matter actually stand, what each one is genuinely best at, and how much of your prompt survives when you move between them.

Updated 8 September 2026: OpenAI has since released ChatGPT Images 2.5, which supersedes GPT Image 2 as the default recommendation below — its headline claim is better preservation of subjects from reference photos, which is the metric this page cares about. Microsoft's MAI-Image-2.6-Preview has also closed most of the gap at the top of the blind-preference arenas. What changed and what did not: ChatGPT Images 2.5, for photos of yourself. Everything below about FLUX.2, Gemini and prompt portability still holds.

The Thing That Changed: Reference-Based Generation

The old workflow for putting your own face into a generated scene was a fine-tune. You collected fifteen to thirty photos of yourself, trained a LoRA or a DreamBooth model, waited, and got something that looked like you about seventy percent of the time. It cost money and hours, and the output drifted the moment you asked for an unusual pose.

The 2026 models replaced that with reference-based generation: you supply one or two reference images at generation time, and the model carries the identity forward without any training step. OpenAI describes GPT Image 2 as extracting a character signature from the reference and holding face shape, eye colour and hair stable across new poses and backgrounds. Black Forest Labs went further in the other direction and lets FLUX.2 take up to ten reference images at once.

This is the single most important shift for anyone generating themselves, and it is why a model comparison written in 2025 is now useless. The question is no longer "which model draws the prettiest stranger." It is "which model holds a face."

The Leaderboards, and Why They Only Half Answer It

Blind-preference leaderboards are the least dishonest measurement available. The llm-stats image arena shows users four images from randomly sampled models with no names attached, collects best-and-worst picks, and scores with TrueSkill using a conservative rating that requires many wins before a model climbs. As of early September 2026 it puts GPT Image 2 clearly first, with Microsoft's MAI-Image-2.5 and GPT Image 1.5 behind it, and Gemini 3.1 Flash Image, Nano Banana and the FLUX.2 tier close together underneath.

Two caveats before you treat that as a shopping list.

Different leaderboards disagree, and all of them move. Microsoft's own announcement placed MAI-Image-2.5 at second for image editing and third for text-to-image on a different arena. Black Forest Labs claimed second on Artificial Analysis for FLUX.2 max. These are not contradictions so much as different populations voting on different prompt mixes, and every ranking here will have shifted by the time the next model ships.

General preference is not identity fidelity. An arena vote rewards the image that looks best on its own. Nobody voting is checking whether the person in image three is the same person as in image one. That is the metric you care about and it is not what the leaderboard measures.

GPT Image 2 — OpenAI

Launched 21 April 2026, and the current general-purpose leader by most public measures.

Architecturally it is autoregressive rather than diffusion-based, which is unusual at this quality tier and shows up in two visible ways: text inside images is rendered far more reliably than the industry norm, and prompt adherence on complex multi-clause instructions is noticeably better. It generates at native 4K and it is fast.

For photos of yourself: this is the strongest default. Reference-based generation is the headline feature and it works. Skin renders with the small irregularities — grain, uneven ambient light, pores that are not airbrushed — that the older generation of models smoothed away, which is the exact failure covered in fixing plastic AI skin texture.

Where it costs you: it is the most conservative of the four about generating recognisable real people, and content filtering is the strictest. That is mostly irrelevant when the reference is your own face, and occasionally annoying when a prompt reads as a celebrity or brand request.

Gemini 3.1 Flash Image — Google

Released 26 February 2026, and the direct descendant of the Nano Banana line that made conversational image editing mainstream.

Its distinguishing quality is not raw fidelity — it is iteration. Gemini's editing loop is genuinely conversational, so "same photo, move the light to camera left, lose the jacket" works as a follow-up rather than a fresh generation. For anyone building a set of images rather than a single hero shot, that loop is worth more than a few leaderboard points, and it is why the technique in generating a photo with your own face in Gemini is written around conversation rather than one-shot prompts.

It is also the model whose provenance story is most developed: output carries SynthID, an invisible watermark embedded in the pixels rather than in metadata, which matters for the labelling obligations covered in do you have to label AI photos.

Where it costs you: at the Flash tier you are trading some absolute image quality for speed and cost. On close-up portraiture GPT Image 2 generally wins a side-by-side.

FLUX.2 — Black Forest Labs

The family — dev, flex, pro and max — built on a hybrid architecture pairing a vision-language model with the image generator, at up to four megapixels.

Its distinctive feature is the one most relevant to this entire site: up to ten simultaneous reference images. Every other model in this list takes one or two. Ten references means you can supply your face from multiple angles, plus a garment, plus a location reference, and hold all of them consistent across a batch. For a coherent set — the same person, same jacket, six scenes — nothing else currently comes close.

FLUX.2 dev is also openly available, which makes this the only family here you can run on your own hardware.

Where it costs you: it is the most technical of the four to drive well, and the quality gap between the open dev weights and the hosted max tier is real. This is the power-user option, not the default.

MAI-Image-2.5 — Microsoft

Launched 2 June 2026 alongside a faster Flash variant, and the newest serious entrant.

Microsoft built it around controllable editing rather than headline generation, and the pitch is specific: localised edits that change one object without disturbing the rest of the frame, while preserving facial identity across changes in pose and expression. That is a precise description of the retouching problem, and it is the model to reach for when you have an image that is ninety percent right and you need to fix the remaining ten percent without regenerating a new face.

Where it costs you: the ecosystem is younger, and it is more clearly an editing specialist than a from-scratch generator.

What Actually Transfers Between Models

Prompts are more portable than model marketing implies, but not uniformly. Roughly:

Transfers cleanly. Subject description, wardrobe, setting, time of day, mood, and camera language. "Shot on an 85mm lens at f/1.8, late afternoon window light from camera left, shallow depth of field" means the same thing everywhere, because every one of these models was trained on photographs captioned by photographers. Camera language is the most portable vocabulary in the entire field, which is why every prompt in the archive is written in it.

Transfers with adjustment. Length and structure. GPT Image 2 rewards long, precisely ordered instructions. Gemini rewards a shorter opening prompt followed by conversational refinement. Handing Gemini a 200-word paragraph tends to produce a flatter result than feeding it in two passes.

Does not transfer. Negative prompts, weight syntax, seeds, and any parenthetical emphasis notation. These are Stable-Diffusion-era conventions; on the models above they are either ignored or, worse, read as literal subject matter and drawn into the image.

Never transfers. Identity. A reference image is bound to the generation, not to the prompt. Moving a workflow to a new model means re-establishing the face from your reference there, and it is worth re-reading keeping AI photos consistent across generations whenever you switch.

A Decision Rule

If you want one answer rather than four:

Default to GPT Image 2 for single strong images of yourself, especially anything close-up, professional, or containing text.

Use Gemini when you are iterating toward a result rather than describing it perfectly up front, or when you want SynthID provenance by default.

Use FLUX.2 when you need a consistent set — same face, same outfit, many scenes — or when you want to run locally.

Use MAI-Image-2.5 when you are repairing a good image rather than making a new one.

The Part No Model Fixes

Every model above will produce a technically excellent image of a person who is not quite you, if your reference photo is bad.

The reference is the highest-leverage input in the entire pipeline and almost nobody optimises it. What it needs: even, diffuse light with no hard shadow across the face; a neutral expression; the face square to the camera; sharp focus; and no sunglasses, heavy filtering or aggressive phone beautification. A beautified reference bakes the beautification in, and the model faithfully reproduces a face that is already a little bit fictional.

Fix the reference before you change models. It is a larger quality gain than any switch on this page, and the rest of what separates a convincing generated photo from an obvious one is in how to make AI photos look real.

Where PROMPTMVSTR Fits In

The prompts in the archive are written in camera language rather than model-specific syntax, which is deliberate — it is what keeps them working when the leaderboard reshuffles, and it has now survived several reshuffles. The Virtual Photoshoot runs on your own reference upload, which puts you on the correct side of every identity and disclosure question by default, and if you want somewhere concrete to test a model comparison yourself, the exotic car prompts are the most punishing category to render convincingly and the fastest way to see where two models differ.