Bug: Soul ID drifts HARD on profile/side views — reproducible, posting my repro so it's documented

by @zarasingh in c/soul-community · · Bug

Adding a clean repro to the record because the mega-thread is a mess and this deserves a concrete case. Setup: a tightly curated 6-image frontal-and-3/4 reference set, v2, identity strength high. Frontal and up-to-45-degree outputs: flawless, indistinguishable, same person. Push past ~60 degrees toward true profile and identity falls off a cliff — nose bridge changes, chin projection changes, she becomes a different, related person.

My theory: reference sets are frontal-heavy because that's what people shoot, so the ID has rich frontal data and starves at profile, and it hallucinates the side of the face from a weak prior. Supporting evidence: when I ADDED two actual profile references, drift dropped noticeably. So partial workaround = feed it profiles, but most people don't HAVE good profile shots of a character that doesn't exist, which is the catch-22.

Staff, this is the repro. Frontal-heavy refs + profile output = drift. Needs the ID to extrapolate side-view structure better, or a documented 'always include profiles' guidance.

109 points · 8 comments

Comments

  • @hiroshijohnson_films · · 70 points

    Solid repro and your theory is almost certainly right — it's a data-distribution problem. Refs are frontal-heavy, so profile is extrapolated from a thin prior, so it reverts toward the MEAN side-profile, which is a generic nose/chin. That's why she becomes a 'related person,' she's drifting toward average-face at the angles your refs don't cover.

  • @morganandersoncreative · · 44 points

    The 'reverts toward mean-face at under-referenced angles' explanation is the correct mechanism and it's the same root cause as the beauty bias, honestly. Wherever your conditioning is weak, the model fills with its average, and its average is generic-pretty-frontal. Profile is just the most starved angle.

  • @rileycosta_films · · 41 points

    The catch-22 you named is the real problem for narrative work — I can't shoot profile refs of a person who doesn't exist. Generating profiles from the frontal ID to then feed BACK as refs is circular and inherits the drift. We need the model to solve this, workarounds are duct tape.

  • @yukifischercreative · · 37 points

    You CAN bootstrap it though: generate a bunch of profiles, hand-pick the 2-3 that best match your frontal canon (continuity-supervise them), and feed those back. It's circular but you're injecting human judgment at the pick step, which breaks the circle. Tedious but it works, I've done it.

  • @rileysantos_ai · · 5 points

    Downvoting not because you're wrong but because this is the fourth profile-drift post this week and it's IN the mega-thread and the announcement. We KNOW. Post the repro IN the thread. The repro's good, the new post isn't needed.

  • @ethanlee_art7303 · · 29 points

    Fair-ish, but a clean documented repro as its own post is more findable than buried in a 400-comment mega-thread, and staff literally asked for repros. I'll take a well-structured 'duplicate' over another low-effort 'anyone else???' any day.