Hobfolk
Studio · case study

A recurring character,
on-brand in every take.

Building a short-form video engine around one repeatable, on-brand character is not a single prompt done well — it is three separate, specific failure modes, each found live, each fixed permanently rather than papered over.

Start a conversationSee the whole crew
You go home. It doesn't.Every promise gets a row and a due dateAnswered the same day, not the next working dayA drawing before the quoteUnverifiable is never a passNine desks, one address eachNothing is sent that a human cannot read backThe folder exists before you price itYou go home. It doesn't.Every promise gets a row and a due dateAnswered the same day, not the next working dayA drawing before the quoteUnverifiable is never a passNine desks, one address eachNothing is sent that a human cannot read backThe folder exists before you price it
The story

What it actually takes to keep one character consistent

A character described well in a prompt is a suggestion. A character locked as its own reusable reference is a fact the next generation has to agree with.

The brief was ordinary: a recurring on-model character for a run of short-form social video, consistent enough that a viewer would recognise her from one clip to the next. The first attempts didn't hold — not because the likeness was wrong, but because the wardrobe drifted between takes even when the likeness was fine. The fix was to stop describing the wardrobe in every prompt and lock it instead as its own reusable reference, generated once and reused, not re-imagined per shot.

The second failure showed up at the finishing stage, on two separate tools: taking a cheap draft render and upscaling it to a finished resolution produced plastic, artifacted output both times, on tools from different vendors. The standing rule since is to iterate cheap at draft resolution and always finish on a native render at the target resolution — never upscale a draft and call it done.

The third was the quietest: a single unset setting on one video edit silently defaulted to the wrong frame shape, and the character drifted off-model as a direct result — not because anything about her had changed, but because the frame itself had. That setting is never left to its default again.

The reference that has to survive every new take
The reference that has to survive every new take

Sourced from Hobfolk's own operational record, dated and on file — see the other four. Client identity is never named; the numbers and the root cause are not softened.

What changed

The permanent fix, not just the patch

01

A locked identity reference, not a prompt description

The character is generated once as a reusable reference asset and reused across jobs, not re-described — and re-interpreted — every time.
02

Wardrobe locked separately from the character

Learned the hard way: wardrobe drifted between takes until it was locked as its own asset, independent of the character reference itself.
03

A written world bible, not just a character sheet

The setting, lighting and mood are specified once, in writing, so every new shot inherits the same world rather than reinventing it.
04

Finishing resolution decided up front, never upscaled

Cheap drafts for iteration, a native render at the real target resolution for anything finished — a rule that exists because upscaling a draft produced visible artifacts on two different tools.
05

Every job logged, cost and lesson both

What was spent and what went wrong are recorded per job, so a mistake is paid for once and not repeated by the next person who didn't see it happen.
06

A human grades before anything ships

Every finished clip is reviewed and approved before it is published — the generation pipeline proposes, it does not get the final word.
How it unfolded

The sequence, in order

  1. The character is designed and locked

    Identity generated once as a reusable reference, not re-described per job.

  2. The first drift is caught

    Wardrobe changing between otherwise-consistent takes.

  3. Wardrobe is locked as its own asset

    Separated from the character reference so it stops being reinterpreted per shot.

  4. Upscale artifacts are caught on two tools

    A cheap draft, finished by upscaling, comes out visibly plastic on two separate vendors' tools.

  5. The finishing rule is set

    Draft cheap, always finish on a native render at the real target resolution.

  6. A silent default is caught before it ships

    One unset setting on a video edit had been quietly changing the frame shape and dragging the character off-model.

Said plainly

What this desk still won't do

A case study is not a claim of no limits. The same boundary applies here as on the department page this one came from.

Where this desk stops

  • It does not invent your brand assets. If the real logo cannot be found, it asks for it. A close-enough redraw is how a brand quietly becomes two brands.
  • It does not sign anything on your behalf. No desk here does.
  • It does not commit print spend. Quotes come to you; the order is yours to place.
Questions

Asked before, answered here

Is this disclosed as AI-generated to the audience?
Yes, plainly, and before anyone would need to ask — the same disclosure discipline this whole site runs on. Undisclosed synthetic media isn't a service this desk offers.
Does this use a client's real brand assets, or invented ones?
Real assets only — logo, palette and any existing photography are used as supplied, never redrawn or invented. Nothing here fabricates a brand identity that doesn't already exist.
How do you actually verify consistency, rather than just claim it?
By locking the identity and wardrobe as reusable reference assets rather than prompt text, and by grading every finished clip against the locked reference before it is approved — the same eyes checking the same fixed thing, every time.
What happens when a generation just doesn't work?
It's logged with the cost and the reason, and it doesn't ship. The log is what turned three separate one-off mistakes into three permanent rules instead of three recurring ones.

How many takes would your current process need to catch a wardrobe change?

Tell us the platform and the run length, and we'll tell you plainly what actually breaks consistency at that scale.

Start a conversation Read the method