People assume an AI companion app is one intelligence that can talk, draw and speak. It is three or four separate systems wired together by the app, and most of the frustrations in this part of the product come from the seams between them.
Three engines, one interface
The language model handles conversation. It produces text and nothing else.
The image model is a completely different system, usually a diffusion model. It takes a text description and produces a picture. It has never seen your conversation.
The speech system turns text into audio. Also separate, also blind to everything except the sentence handed to it.
When you ask your companion for a selfie, this is what happens: the language model writes a short description of the requested picture, the app adds the character's stored appearance tags, that combined prompt goes to the image model, and a picture comes back. The chat then reacts to a picture it cannot see, using the description it wrote.
That relay is the source of nearly every complaint about media in this category.
Why the face keeps changing
Because a diffusion model is not retrieving your character, it is generating someone matching a description. "Long dark hair, green eyes, mid-twenties" describes millions of faces, and you get a different one each time.
Apps solve this to varying degrees:
- Stored appearance tags — the same descriptive string every time. Cheap, and only roughly consistent.
- A reference image conditioning each generation. Much better, and the usual approach when faces stay recognisable.
- A tuned model per character — the most consistent and by far the most expensive, so it tends to appear only on higher tiers.
This is a real differentiator worth testing in a free tier before paying. Generate four pictures of the same character in different situations. If you get four different women, no subscription will fix it. Character consistency across images is a large part of why Candy AI leads our ranking, and why Secret Desires, which builds looks, voice and personality together and adds short video, scores where it does.
Why media is always metered
Text is cheap. Generating a chat reply costs the operator a fraction of a cent.
An image costs meaningfully more per generation. Voice is billed by the second of audio. Video is dramatically more expensive than either.
That cost difference is the entire reason for the pricing structure you see everywhere in this category: generous message allowances, tight image caps, voice minutes sold separately, video restricted to upper tiers. It is not artificial scarcity designed to upsell you — it is the actual shape of the bill the operator receives, passed through.
Which is why "unlimited" in this market nearly always means unlimited text. Read it that way and the pricing pages suddenly make sense. The arithmetic is in what image generation really costs you.
What voice adds, and what it costs
Voice changes the experience more than most people expect. Reading a message and hearing it are different, and the shift is larger than the technical description suggests.
Two things to check before paying for it:
Latency. A three-second pause before every reply breaks the illusion completely. Test it on a free tier or a trial, not from a demo video.
Whether it is a call or a message. Voice notes — the character reads its reply aloud — are common and cheap. Real-time calls, where you speak and it answers, are a different product and usually a different tier. Nomi lists voice calls among its features; many apps list "voice" and mean notes.
What images will not do
Two limits worth knowing before you are disappointed:
Hands, text and fine detail remain unreliable across every generator in this category. Nobody has solved this, and an app promising otherwise is describing its best output rather than its average.
The image does not know your scene. The chat model writes a one-line prompt, so everything not in that line is gone — the room you described, what she was wearing three messages ago, the time of day. Being explicit in the request fixes more of this than any setting.
How to judge the media side in one evening
Spend a free tier's entire allowance in a single session rather than one picture a day. You are testing four things:
- Consistency — four pictures, same character, different situations.
- Prompt adherence — ask for something specific and see how much survives.
- Latency — for both images and voice.
- What counts against the cap — whether a failed or refused generation still costs you.
That last one is the least documented and most annoying to discover after paying. Along with the rest of what the free tiers actually include, it is the sort of thing that only shows up when you use the thing properly.
