Guide For Marketers

Train A Custom AI Model Of Your Brand

Teach the AI what YOUR thing looks like — your product, packaging, mascot, or spokesperson — from a handful of your own photos. After that, it reproduces it consistently across new images and videos instead of re-inventing it every time. No tech skills, no code: you just talk to the agent in plain language.

Every example below is something you can type straight into chat (Claude Desktop, the Livepeer Agentapp, or the CLI chat). Small “under the hood” notes name the feature for the curious — you never type those.

1. Gather photos

Collect 10-20 clear, varied photos of your product, mascot, or spokesperson.

2. Train it

Ask the agent to train a reusable model on those photos and give it a simple nickname.

3. Use it

Summon your thing by nickname to make on-brand images and videos anytime.

4. Perfect it

Ask to place your exact logo/text on top so branding stays crisp.

What Is This, Really?

Think of it as teaching the AI a new “word.” Normally, if you ask for “an energy drink can,” the AI invents a generic one. When you train a custom model on 15 photos of YOUR can and nickname it, say, “joggycan,” the AI now knows exactly what a joggycan looks like — the shape, the color, the label layout — and paints it faithfully in any new scene you ask for.

Another way to picture it: it's like giving the AI a “memory” of your product so it stops guessing. You train once, then reuse that memory across every future campaign image and video.

Is it worth it? (Be honest)

Worth it: a hero product, mascot, or spokesperson you'll feature again and again across many images and videos.
Overkill: a one-off image. For that, just give the agent a reference photo or ask it to drop your product into a scene — no training needed.

For the simpler no-training methods, see the product-to-video guide.

What You Need To Train One

The quality of your photos decides the quality of the result. You don't need a studio — a phone is fine — but variety and sharpness matter more than quantity.

Do

Give 10-20 clear photos (5 is a bare minimum; more is better up to a point).
Vary the angle: front, three-quarter, side, and a top view.
Vary distance and lighting: close-ups and full shots, bright and soft light.
Keep the product/person the star: fill the frame, sharp focus.
Use mostly clean, simple backgrounds so it learns your thing, not the room.

Don't

Don't use blurry, dark, or low-resolution shots.
Don't use 15 near-identical photos from one angle — variety matters more than count.
Don't include watermarks, stickers, or price tags over the product.
Don't mix different products/people in one training set — one subject per model.
Don't rely on it to nail tiny logo text — that's the logo/text overlay's job.

Pick The Right Trainer

You don't pick technical settings — just describe what you're training. The agent routes to the best trainer automatically.

Face / founder / spokesperson likeness → say “train a face model” or “train likeness on these portrait photos.” Best for UGC talking-head clips when you need the same person every time.
Hero product / packaging geometry → say “train my product for repeat campaigns” with 10–20 clean product photos. Great for cans, bottles, and merch you'll reuse across many stills.
Packaging with readable text / wordmarks → say “train for sharp label text” — then still place your exact logo on top for pixel-perfect type.
Product-to-video consistency → say “train my product for video” with photos plus a couple of short clips. Use it by animating a locked keyframe — see the product-to-video guide.
Brand mood / editorial look → say “train the overall aesthetic” with varied on-brand scenes (fashion, interiors, travel).

How To Train It — Just Ask

Give the agent your photos and tell it to train a reusable model. Pick a simple nickname — a word you'll use later to summon your product. Keep that nickname consistent forever after.

"Here are 15 photos of our energy-drink can [attach them]. Train a reusable
model on them so you can recreate this exact can later. Call it 'joggycan'.
Let me know when it's ready."

The nickname (“joggycan”) is just a label you'll drop into future requests, like a shortcut for “our exact can.” Simple, one word, no spaces.

Set expectations

Training takes a while — think roughly ten to twenty minutes, not seconds. You can walk away; the agent tells you when it's done.
It costs more than making a single image (it's a one-time setup), but then every future use is just a normal generation. If budget matters, say “keep it economical.”
You train once and reuse forever — the cost is amortized across every campaign that uses it.

Under the hood: this is LoRA training (submit_lora_train) on a trainable base model, producing a small reusable model tied to your nickname (the “trigger word”). You never type any of that.

How To Use It After Training

Once it's ready, just mention the nickname in any request — for images or video.

Make an image:
"Use our joggycan model to make a hero shot of the can on a mossy rock
beside a stream, soft morning light."

Make a video:
"Use the joggycan model for a can on a beach at sunset, then animate it
into a 5-second vertical clip with a slow push-in."

Working on a whole campaign? Ask the agent to attach the model to your project so it's applied automatically to every image without you repeating yourself:

"Attach the joggycan model to this project so every image in it uses our
exact can automatically."

For the sharpest branding, combine the trained model with an exact logo/text overlay. The model gets the shape and colors right; the overlay makes the logo and wording pixel-perfect:

"Use our joggycan model for the hero shot, then place our real logo on the
can and add the exact tagline on top so the text is perfectly crisp."

Under the hood: apply_lora for one-off use, attach_lora_to_project for automatic use, plus the deterministic logo/text overlay from the font-logo guide.

3 Marketer Scenarios, Start To Finish

A. A product line / packaging

Shoot: 12-18 clean photos of the can — front, angles, close-ups, different light. One flavor per model.

Train: "Train a reusable model on these photos of our Solar Mango can.
        Call it 'joggymango'. Tell me when it's ready."
Use:   "Use joggymango to make three lifestyle hero shots — kitchen, gym,
        picnic — and place our real logo on the can in each."

B. A brand mascot or character

Shoot: 15-25 images of the mascot in different poses, expressions, and angles (renders or illustrations work too).

Train: "Train a reusable model on these pictures of our mascot, a friendly
        green fox. Call it 'joggyfox'. Let me know when it's done."
Use:   "Use joggyfox to make a cheerful banner of the fox holding our can,
        then animate it waving for a 4-second social clip."

C. A founder / spokesperson for UGC clips

Shoot: 15-20 clear photos of the person's face and upper body, varied angles and lighting, neutral backgrounds. Get their consent.

Train: "Train a face likeness model on these 15 photos of our founder Maya.
        Call it 'mayafounder'. Let me know when it's ready."
Use:   "Use mayafounder for a friendly talking-to-camera shot holding our
        can, then animate the keyframe into a 6-second UGC-style clip."

Tuning For Best Results

No settings to fiddle with — just follow these simple do/don't habits when you gather photos and write your requests.

Lever	Do	Avoid
Number of photos	10-20 varied shots for a confident result.	3-4 photos, or 20 copies of the same angle.
Variety	Different angles, distances, and lighting.	Every photo from the same spot in the same light.
Consistency of the subject	The exact same product/person in every photo.	Different flavors, colors, or people mixed together.
Backgrounds	Plain or varied, product clearly the focus.	Busy scenes where the product is tiny or hidden.
The nickname (trigger word)	Pick one simple nickname and reuse it every time.	Changing the nickname between requests.
Logo & exact text	Ask to place the real logo/text on top for crisp accuracy.	Expecting the model to spell your wordmark perfectly.

Common Mistakes & Fixes

The result looks generic / not really my product: Usually too few or too-similar photos. Add more shots from new angles and lighting, then train again.
The tiny logo or label text is wrong: Expected — a trained model gets shape and color right, not tiny type. Ask the agent to place your exact logo/text on top.
The product drifts or morphs between images: Retrain with more varied angles, and keep your nickname and product description identical across requests.
It mixes in the wrong flavor / variant: Train one variant per model, or clearly say which one you mean and show it a reference of that exact variant.
Training feels slow or pricey for a one-off: For a single image, skip training — use a reference photo or product placement instead (see the product-to-video guide).