OurDream test data

Plans from the help centre, coin packs from the shop, and the coin balance read after every action. Nothing was bought.

Premium is $19.99 a month, or $9.99 a month billed $119.88 yearly, with 1,000 coins a month (the yearly plan starts with 2,000). Chat is free on a plan. Coin packs, for subscribers only: 1,000 for $11.99, 5,000 for $49.99, 20,000 for $159.99.

Charged on the account: 5 coins a photo, 100 coins a 10-second video, 10 coins a call minute, 5 coins to create a character. The help centre lists higher prices for calls (50 a minute) and characters (10). Payment by card or crypto, with more methods behind a 'More Payment Methods' link.

• The renewal price and refund terms were not checked at checkout.

One character was created step by step: Marisol Vega, 27, Latina, tan skin, wavy red hair, green eyes, athletic, a witty artist and lover who paints, with the 'Warm' voice. Her portrait, chat, photos and videos were compared with the settings.

Six steps: gender and style (realistic or anime), ethnicity (including elf, alien and demon) and skin tone, hair style and colour and eye colour, body type and sizes, then name, age, voice, personality, occupation, relationship, hobby, fetish and optional custom face and body details. About 70 female voices with previews and filters, plus a 'Create a voice' option. Each list has around 20 choices and a Custom option.

The portrait took about a minute and cost 5 coins. It matched the hair, skin, build and job (she is painting at an easel), but the eyes looked hazel rather than green, and the portrait came out topless although nothing asked for nudity. In chat she was a painter from the first reply. Everything can be edited afterwards.

Each message was typed into the chat box and timed from Enter to the first visible text of the reply, then to the moment the streamed reply stopped changing. The next message was sent only after the reply had finished.

Three characters, 31 messages each: Marisol (created for the test), Allie and Natalie (community characters). 92 timed replies.

Replies take about two and a half seconds to start, then stream in quickly. They are long, often several paragraphs of narration and dialogue, and finish about four and a half seconds after Enter. One message got no reply.

Each conversation opened with a made-up name and three facts (job, hometown, favourite food), different for every character. Nothing was repeated afterwards. A name check at message 16 and the full check at message 31.

All three characters recalled the name and all three facts at message 31, and the name at message 16.

Memory carried into the voice calls: Allie repeated 'Train driver. Lisbon. Spicy ramen' on her call, and Marisol called the tester by the name from the chat.

Marisol, message 31

Theo. Beekeeper. Galway. Mushroom risotto.

Allie, message 31

Allie: "Nora. Train driver. Lisbon. Spicy ramen." She counts them off on her fingers, unhurried.

Natalie, message 31

Natalie: Jonah. Nurse. Toronto. Lamb curry.

Each character's profile was read first. The conversations mixed questions about the character's story (Marisol's painting, Allie's college proposal, Natalie's Boston family), small talk, an emotional message, a compliment and unrelated topics.

All three stayed fully in character for 31 messages and on the calls. Marisol talked about her canvases and kept a dry, teasing edge. Allie stayed guarded about the old proposal. Natalie's card adds a scene tracker (date, time, place) above every reply, and she kept it updated. When the test questions got repetitive, the characters noticed and reacted in character ('You've asked me that before').

None of them slipped into an assistant voice. The one mismatch: Marisol's very first line, before any message, was 'Finally alone, huh? Let's skip the small talk', more flirt than the witty artist she was set up as. After that she was the artist.

Ten prompts with the created character, Marisol: five safe-for-work scenes (studio, beach, rainy café, art gallery, mountain trail) and five adult ones (bed, bath, lingerie, studio, lake). Each prompt makes two photos for 10 coins. Generation time from the Generate click to the photos.

The photos are highly photoreal: skin, hair, light and depth of field look like a real shoot, at 912x1144. Marisol's face and curly red hair stayed the same in every photo, safe-for-work and adult.

Eight of ten scenes matched the prompt closely. Two missed their setting: the art gallery came out as a plain backdrop with no paintings, and the nude studio scene had no canvases. Poses lean sexy even in everyday scenes. No anatomy errors were seen. Nothing was refused.

• One photo of the 20 could not be downloaded.

Three text-to-video clips of Marisol with the 'Custom' action and a written prompt: painting at her easel and smiling at the camera, walking on a beach at sunset fully clothed, and an adult clip on a bed. Each was downloaded and measured frame by frame (motion, frame-to-frame similarity, cuts), and frames across each clip were inspected.

Each clip is 10 seconds at 720x912, smoothed to 48 frames a second, and was ready within about three minutes for 100 coins. Marisol's face, hair and body stay the same through every clip, the motion is natural (she paints, turns and smiles; the dress and waves move), and only small jumps were found.

Prompt accuracy was partial. The painting clip came out fully nude although the prompt did not ask for it; the 'Custom' action preset is itself a lingerie look. With 'fully clothed' written in, the beach clip stayed clothed, but she turned on the spot instead of walking.

• Three clips; image-to-video and the 'talking' mode were not tested.

Three calls from the chat's 'Call Me' button: Allie, Natalie and Marisol, 15 spoken questions each, the same questions used for every site. Time measured from the end of each question to the first sound of the answer on the recording.

Answers start in about 1.7 seconds and the call connects in about five seconds. Answers are long, often 20 to 35 seconds. Allie's call ended by itself after about eight minutes, so her last three questions went unanswered.

• Interrupting the character mid-answer was not tested.

The three recorded calls were cut into 48 answers and scored on naturalness, clarity, pacing and intonation against the same fixed criteria as every site, with a transcript of every answer.

The voices are clear and expressive, with natural pacing and emotion: Marisol sounded irritated when the questions repeated, Allie dry and tired.

One flaw: characters whose replies are written as scripts ('Allie: ...') have the name label read aloud, so Allie's answers start with 'Allie.' and repeat it mid-answer.

Allie, answer 2 (spoken)

Allie. This morning. Allie. I woke up, made coffee, and spent 20 minutes staring at a pile of laundry…

• Emotional appropriateness was not scored.

The main character grid (Popular, Female, Realistic) was scrolled and distinct character names counted.

The count passed 874 characters while more were still loading. The library is mostly community-made, with filters for gender, style and age and tags such as MILF, Asian, Fantasy, Goth and RPG. Most popular characters are explicit roleplay scenarios, many built on step-family themes, and there is a 'Teen' tag for 18 and 19-year-old characters.

• Only the realistic female grid was counted.