Here's the setup. I had one jpeg. Like, literally one. A single picture of myself, and I wanted a character LoRA that would lock my face, hold my body shape across wardrobe changes, and let me spit out references for every downstream workflow I'd ever care about.
The problem with training a LoRA on one image is that it either overfits catastrophically or learns nothing useful at all. You need a set — different angles, different expressions, different clothing, different lighting — all recognisably the same person.
So the interesting question was never "how do I train a LoRA." Character LoRAs are a solved problem if you have a dataset. The interesting question was how do I synthesise a clean dataset from one reference frame, for as little money as possible.
Final number: $4.20 for the good run. About $8 including every dev retry I threw away. Amortised cost per reference image after the LoRA is trained: basically electricity.
Here's how.
Why references are the choke point
Ask anyone who's built a consistent character pipeline and they'll tell you the expensive part isn't the sampler, the trainer, or the GPU hours. It's getting a reference set that actually teaches the model this is one person from different angles, and not these are five people who kind of look the same.
Stock datasets solve this with hundreds of photos per subject. I had one.
The move: expensive once, cheap forever
The trick is to treat reference generation as a two-stage pipeline where you pay for quality exactly one time.
Stage one uses a high-prompt-adherence, high-consistency image model. I used nano-banana-pro 2. It's expensive per image — roughly seven cents a pop at current pricing — but its ability to hold identity across pose, wardrobe, and lighting variations is unreasonably good. Stage two is training a small LoRA on top of whatever base checkpoint you like.
After stage two, the marginal cost of a new reference is the electricity your GPU burns running the sampler. On my box that's measured in fractions of a cent.
So the whole economics problem reduces to one question: how much does stage one cost to run well?
The reference recipe
You don't just ask the big model for "more pictures of me." You architect a reference set that isolates the things a LoRA is bad at learning, and strips out everything it's good at memorising by accident.
I ran two passes.
Pass 1 — face isolation
- Black background
- Black clothing, high collar, no pattern
- Soft key light, no hard shadows
- Many angles, many expressions
The reasoning: LoRAs love to overfit on the thing next to the face. If half your training set has a red hoodie in the frame, congratulations, you just trained "red hoodie" as part of the identity. Black-on-black collapses the entire visual signal down to skin, hair, and features. There's nothing else for the trainer to latch onto.
The background colour isn't cosmetic. It's an information filter.
Pass 2 — body lock
- Short gym clothes (tank top, shorts)
- Neutral studio backdrop
- Head-to-toe framing
- Varied poses, consistent proportions
This pass is about silhouette. Baggy clothing hides body shape from the trainer. If every reference you feed it is in an oversized hoodie, the LoRA will happily learn "this character wears an oversized hoodie" and give you a suspiciously different body shape the moment you prompt for anything tighter.
Gym clothes let the trainer see the actual proportions — shoulder width, waist, limb length, hip ratio — without cloth drape getting in the way. After this pass, the LoRA holds body shape across wardrobe changes, which is the single hardest thing for a character LoRA to do well.
The captioning trick
Caption the things you want the LoRA to ignore, not the things you want it to learn. If every image has a black background, put "black background" in the caption. The trainer files that attribute as a prompt condition instead of part of the subject. Same for clothing, lighting, pose.
Never caption anything about the subject you actually want the LoRA to learn. The trainer learns "the thing in the image that isn't called out in the caption" as the identity.
That's the whole principle in one line: caption what you want ignored, stay silent about what you want learned.
The numbers
| line item | cost |
|---|---|
| dev, retries, waste | ~$8.00 |
| final clean run | $4.20 |
| per reference after LoRA | ~$0 |
$4.20 buys enough high-consistency images to train a usable character LoRA if you're not being wasteful with stage-one prompts. The other $8 was what it cost me to figure out which prompts weren't wasteful.
Once the LoRA exists, every reference image I generate for the rest of the project is effectively free. That is the entire economic argument for this workflow, and it's a big one.
Proof
Same base checkpoint. Same positive prompt, same negative prompt. Same seed. Same 42 sampler steps. Same CFG. The only thing that differs between the two images below is a single scalar: the LoRA contribution strength, 0.0 on the left and 0.87 on the right.
On the left, the base model's idea of "green-haired girl in a black hoodie doing a peace sign." Generic anime, flat silhouette, rounder face, no ahoge, no black lipstick.
On the right, the LoRA-conditioned render. Fluffy layered hair with the signature ahoge, sharper jaw, proper hoodie drape, the black lipstick detail, the right body proportions under the fabric. Recognisably me.
Nothing about the prompt describes any of the things the LoRA brought in. That entire delta — hair structure, face shape, wardrobe drape, body proportions — is coming from weights that started life as one image.
What this unlocks
The reason this matters isn't the LoRA itself. It's the cost curve.
Once you can produce a high-consistency character for under five bucks, reference generation stops being a budget item and starts being background noise. That changes what workflows you can reasonably consider downstream. Multi-view generation. ControlNet pose rigs. Turnaround sheets. Pose libraries.
3D.
A lot of the AI-to-3D pipelines people are experimenting with right now break down not at the geometry stage but at the input stage — they need dozens of consistent views of the same character from specific angles, and nobody wants to pay seven cents a view when the pipeline might need two hundred of them to converge.
Free references change the arithmetic. More on where that leads another time.
For now, the recipe is the recipe. Pay once, isolate what you want learned, caption what you want ignored, and the rest follows.
One jpeg in. $4.20 spent. Infinite references out.