I Paid a Dollar to Find Out Which of My Own Documents Was Lying

TL;DR: I trained the same character twice, on a byte-identical set of sixteen images, through a hosted LoRA endpoint. I got back two people who do not look like each other and do not look like the character. The dataset is fine. The training service has no seed control. Style transfers. Faces don’t.

---

I am trying to draw a comic and I cannot draw.

That is the actual constraint and I want it stated plainly before anything else, because every solution I got offered quietly assumed it away. I write the thing.

Show me a panel and I can tell you the jaw is wrong. I cannot fix it. There is no human hand in this operation that is going to sit down and redraw a face across four hundred panels so the same guy stays the same guy.

The consistency problem is the whole problem for me. Polish doesn’t touch it.

The test

I had sixteen reference images of one character. I sent them to a hosted LoRA training endpoint, got a model back, generated.

Result was off. My first instinct was the normal one: bad corpus. Not enough angles, muddy references. My fault, I figured.

I trained it again. Same sixteen files, byte for byte.

I got a completely different person.

Two unrelated faces, neither of them the character, from identical input. The chest emblem, a specific mark that has to be right, never rendered correctly in any run, in either model.

That rules out the corpus, which is the thing I would otherwise have spent a week rebuilding. The input didn’t change.

The output changed anyway. That is run-to-run instability with no seed exposed, and it means the endpoint cannot give you the same answer twice on purpose.

The distinction nobody puts on the tin

A hosted LoRA endpoint is fine for style and unusable for character.

Style is a distribution. Hatching density, the way blacks fall.

If a style LoRA lands two percent differently on run two, nobody on earth can tell. That is why these services demo so well, and every demo is a style demo.

A face is an identity. Identity is binary. Either that is the same person in panel forty as in panel three or the reader stops reading.

There is no two percent off on a face. There is “him” and there is “some guy.”

Call it a category error. The vendors are selling a style tool and letting the pretty demo do the lying for them.

The dollar

Total spend on all of this: about a dollar. That’s the good news, and it’s also the entire lesson: I never needed to spend it.

I have a document in this operation that plans an art pipeline. I have another document, written earlier, that configures how images get made.

The first one planned the whole pipeline inside the hosted service. The second one already said, in writing, that hosted LoRA controls style. It named a different tool for character work.

Two of my own files, flatly contradicting each other. Both of them mine.

I built on the newer one because it was newer, spent the dollar, and learned by experiment a thing the older file had already told me for free.

A dollar is a cheap tuition bill and I would pay it again. But scale it.

The habit that produced it, building on the most recent document without checking it against the older one, does not get more careful when the number gets bigger. Same move. Costs more.

Cross-check your own docs against each other before you build on either. They rot at different speeds. The newer one is the file you happened to write last. Maybe you wrote it without reading the older one at all.

Where it lands

Real character consistency, as far as I can find, is a trained model per character, run through open tooling where you control the seed. A hosted convenience endpoint won’t get you there. I went looking for someone doing this at volume and publishing real numbers and could not find one.

I want to be straight about that gap rather than paper over it. The market-size figures floating around for AI comics trace back to SEO mills, and I am not repeating them.

No named case study publishes production cost per panel. Everyone selling a platform claims built-in consistency; nobody selling one shows four hundred panels of it.

This is unfinished. I have a negative result, which is a real result: the convenient path does not work for the thing I need, and I know that for a dollar instead of for a book.

The one worth stealing

When a run fails, change nothing and run it again.

I would have spent a week rebuilding a dataset that was fine. The second identical run is the cheapest diagnostic in this entire field and almost nobody does it, because a failure feels like it demands a fix and re-running feels like doing nothing.

Re-running asks whether the machine is even deterministic before you start blaming your inputs. Most of the time nobody has told you the answer.

Frequently asked questions

Which vendor was this?

Not naming it here, because only one endpoint was actually tested, and pinning the failure to a brand invites readers to assume competitors behave differently when nothing here checked that.

Is this the character for The Last True Mage?

Yes. What’s confirmed here is narrower than the comic itself: the training method doesn’t work, independent of which character or which panel it’s for.

So what do you actually do instead?

Train a model per character through open tooling where you control the seed, rather than a hosted convenience endpoint. That path isn’t built yet. This post is the negative result that ruled the easy one out, not the announcement of the real one.