Frequently asked.
Training a subject
How many photos do I need?
Five is the minimum, 15–30 is the sweet spot. Variety beats volume: different angles, expressions, lighting and backgrounds teach the model what actually stays the same about your subject. Thirty near-identical photos from one afternoon teach it the afternoon.
Which photos should I avoid?
Group shots (the model can't know which face is yours), sunglasses and heavy filters, extreme close-ups only, and long bursts of the same pose. One or two full-body shots help a lot.
How long does training take, and what does it cost?
Standard (SDXL) trains in roughly half an hour for 200⚡. Premium (FLUX) is our best quality at 1000⚡, or 1400⚡ for Rush, which jumps the queue. Nothing is charged until you confirm the job.
What if my trained portrait looks bad?
First re-train is free. Better photos in — more variety, fewer filters — is almost always the fix. After the second re-train we offer a partial refund.
Writing prompts that work
Why did the image ignore half of my prompt?
Every image model has a budget of roughly four or five distinct visual requirements per prompt. Past that it stops following your details and starts blending them into the most plausible overall picture — a requested detail can even slide onto the wrong part (“spiral tip” ends up as a spiral handle). It's a model limitation, not your wording. The fix: prompt the overall shape first, then fix the details one at a time with Repaint.
What does a good prompt look like?
Describe the picture in one visual sweep, the way your eye would scan it — for an object, top to bottom. Say what each part IS (“the center prong is a drill-like twisted spike”), not what it's like (“like a M48 cyclone” — the model doesn't know product names). Skip counts, negations and functional claims (“can be used as both”) — they spend your budget without adding anything visible. And describe the plain parts too: “a simple smooth steel shaft” is what stops decoration from migrating there.
Does image size matter?
Yes — the aspect ratio is itself an instruction. A tall subject (a full-length person, a staff, a bottle) wants a portrait canvas; a landscape wants a wide one. Forcing a tall subject into a square makes the model crop or shrink it.
What is seed locking for?
The seed is the randomness. Lock it and change one phrase at a time, and you can watch exactly what each wording change does to the same image — the fastest way to learn what the model responds to. Unlock it to explore variations of a design you like.
What is Draft mode?
Half the Sparks for a quicker, rougher render. Iterate on composition in Draft, then run the final at full quality.
Why was my prompt rejected?
A safety layer reviews prompts and photos before we spend your Sparks — it declines sexual content involving real people's likenesses, anything involving minors, and a few similar categories. Rejections cost nothing, and the reason is shown with the result. If you think it got yours wrong, contact us with the prompt.
Reference images
How do I control the exact pose or shape?
Words control style; a reference image controls geometry. In Compose, attach a reference and pick a mode: Pose copies a person's posture from your reference, Sketch follows a drawing — even rough marker art — for exact shapes and layouts, and Depth borrows the spatial composition of a photo. A reference adds 1⚡ to the render.
Which reference mode should I pick?
Pose for people (it follows the skeleton, not the clothes). Sketch when the shape itself must be right — logos, weapons, products, layouts. Depth when you want the same scene structure with everything restyled.
Fixing an image
What's the difference between Upscale and Enhance?
Upscale (1⚡) makes the image bigger and cleaner without changing it. Enhance re-imagines fine detail while enlarging — sharper results, priced by output size, and it may subtly reinterpret textures. Use Upscale for faithfulness, Enhance for punch.
How does Colorize know the right colors?
It doesn't — nobody can recover the true colors from a grey photo. It makes a plausible guess, and the hint box is where you beat it: “the car was British racing green, the door was red” steers the result. You were there; it wasn't. 4⚡.
Can I change just one part of an image?
Yes — Repaint (4⚡). Brush over the area, describe what should be there (or leave it blank to remove something), and only the brushed area changes; every other pixel is untouched. This is the intended workflow for detailed work: get the composition right first, then repaint the parts that missed.
Do these tools overwrite my original?
Never. Every tool saves its result as a new image next to the original.
3D Studio
Can I make 3D models?
Yes. Open any picture in your gallery and choose Make a 3D model (40⚡, about three minutes), or go straight to the 3D Studio. You get a real textured .glb you can spin around, edit, download, and share.
What source images give the best 3D models?
One object, centered, photographed straight-on, on a plain background — think product photo. We cut the background automatically, but a busy scene, harsh shadows, or multiple objects confuse the reconstruction. Objects and props come out great; realistic human figures aren't there yet.
Can I edit a model after it's built?
That's what the Studio is: paint directly on the surface, patch texture holes, sculpt the shape (raise, lower, even out), cut pieces off, simplify for game engines, and Regenerate an area (4⚡) to have the AI repaint part of the skin with real detail. Every edit is 2⚡ and saves a new version — nothing is ever overwritten.
Can I use Blender or Photoshop on my model?
Yes — the Files panel in the Studio downloads the model (.glb) and its texture (.png). Edit them in any tool and import them back; the import becomes a new version here.
Can other people see my models?
Only if you choose. Share to Explore puts a model on the public Explore page; toggling it off removes it immediately.
Sparks & billing
What are Sparks?
The credit that pays for everything: a standard render is 2⚡, premium 4⚡, image tools 1–4⚡, a 3D model 40⚡, training 200–1400⚡. Prices are always shown on the button before you click.
What happens if a job fails?
Your Sparks come back automatically — including when a worker dies mid-job; a background process finds stranded work and either restarts it or refunds it. If you ever see a charge without a result, that's a bug: tell us.
What payment methods do you accept?
Credit and debit cards, processed by Stripe — buy Sparks packs or subscribe right from your wallet. Crypto is on the roadmap.
API & webhooks
Do you have an API?
Yes — everything the site does, your code can do: train subjects, render images, run the image tools, build and edit 3D models. Create a key at Account → API keys, send it as a Bearer token, and read the full reference at loralume.com/developers — the quickstart examples there are tested against production, so they work as written.
My API key gets a 401 — why?
Three usual causes: the key was revoked (check your keys page), your plan doesn't include API access, or the header is malformed — it should be exactly Authorization: Bearer lm_live_…. Keys are shown once at creation; if you lost it, revoke and make a new one.
Why does my key get 403 on some endpoints?
By design: an API key can do everything your account can do except manage keys or change credentials. Those need a web login, so a leaked key can never lock you out of your own account.
Do you support webhooks?
Yes — register an endpoint and we POST signed JSON when renders, image tools, and 3D builds finish, so you never have to poll. See the API docs under “Receiving webhooks” for events, signatures, and retry behavior.
My webhook isn't receiving anything — why?
The checklist, in order: the URL must be public — localhost and private/LAN addresses are rejected for security. Your receiver must answer with a 2xx within 10 seconds. Failures are retried for about 2.5 hours, and an endpoint that fails 50 times in a row is switched off (the reason appears on the webhook). The fastest diagnosis: call the /test endpoint to send a ping, then read /deliveries — it lists every attempt with its error.
Privacy & safety
Do you share my photos?
No. Photos are used only for your specific training job. They are not used to train any other customer's portrait. Deleted from active systems 90 days after delivery.
Do you track what I generate?
No. We don't monitor your activity or browse your gallery. If you want our eyes on a specific image — for support or feedback — share it with us yourself.
Can I train on someone else?
Only with their explicit written permission. Our verification step requires their selfie matching every photo in your upload. If they haven't verified, you can't train on them.
What about minors?
All photos are scanned by a multimodal AI guard model before training. Any photo where the person appears under 18 is automatically rejected. We use this for every upload, no exceptions.