← Blog

Two kinds of character break platform training. Most PFPs are one of them.

2026-08-14 · avatar-lab

A profile-picture NFT is a token pointing at one image. The image is fine — it is the thing you chose, and for a lot of people it has been their face online for years. But the moment you want that character to do something it was not drawn doing — a different angle, a scene, eight seconds of video — something has to generate it again, and that is where it usually stops being the same character.

How far it drifts is measurable. A prompt that merely describes a character in words scores 0.18–0.27 on a face-identity scale where 0.363 is the line between “same person” and “different person”. Not a worse version of your character. A different one.

The obvious fix is the one that fails

Most platforms now offer a character system: upload a handful of images, wait for training, get a reusable character you can drop into any prompt. It looks like exactly the right tool. We measured what it does to the two kinds of character that profile-picture collections actually consist of — non-human creatures, and stylised or illustrated humans.

Non-human: refused at the door

Apes, frogs, cats, robots, blobs. We submitted 20 images of a non-human character to Higgsfield’s Soul training. It was accepted, queued, and then:

status: failed   fail_reason: "face_not_found"

The training step looks for a face. No face, no character. The cinematic variant behaved identically, so this is a platform-level check rather than a quirk of one model. To the platform’s credit, the failure costs nothing — but it also arrives after the submission looked like it had worked, which is why we never treat “queued” as success.

Stylised human: trains perfectly, comes back as someone else

Punks, anime, illustrated characters. This time training completed normally — 25 credits, about half an hour, status: completed, no error. Then we generated a close portrait with it. Hair colour wrong. Eye colour wrong. The star-shaped birthmark gone. Collar colours wrong. A competent picture of a different character.

None of the identity markers survived training. All of them survived simply attaching one reference image. The reason holds up: these systems lock the face. For a photorealistic human the face is most of the identity, so locking it is a reasonable trade. For an illustrated character the identity is mostly not on the face at all — it is hair colour, a birthmark, a palette, a collar, an accessory. Lock the face, let everything else drift, and you get a stranger with the right bone structure.

We saw the same pattern on a second, unrelated platform: attaching its character object measured no better than simply supplying the start frame, with the markers holding at 100% either way. Two independent platforms agreeing turned this from a platform quirk into a rule we now apply by default to any platform of that type.

What does work

Reference mechanisms — a written specification, a small set of reference images, and wording adapted per platform. For non-human characters those score 107% (Gemini), 105.5% (Midjourney) and 110.8% (Veo) against the reference set’s own ceiling: the generated frames agree with the character more closely than the reference images, shot across five extreme angles, agree with each other.

This is why the app will not sell you the expensive option. For a non-human character the training route is not offered at all; for a stylised one it warns that training measured worse than not training. The two warnings are worded differently on purpose — “this will fail” and “this will succeed and be worse” are not the same message, and blurring them would leave people thinking illustrated characters cannot be done here.

Limits, as always: five images per group, one character per type, single base model per platform. Read them as magnitudes, not as a ranking with decimal places. The full run is on the measurements page.

A word on rights, since people ask

The token and the licence are two different things, and the chain records only the first. Bored Ape holders were granted broad commercial rights to their ape. Nouns, Cryptoadz and mfers are CC0 — which means the artwork is free for anyone to use, and that includes anyone who is not you. Many other collections grant limited rights or none at all. Holding the token does not tell you which of those you are in; the collection’s terms do. Read them before you build anything commercial on the output. That is a description of how these licences generally work, not advice about your particular collection.

Registering a character here does not change any of it. An ID records that this specification existed here, unaltered, from a given date — evidence, not title. We are specific about that distinction, because a registry that overstates it is writing a cheque its records cannot cash.

Bring your character in →  ·  Argue with the numbers in Discord