Back to Livepeer Agent

Best-Practice Guide

Product Shots To Consistent AI Video

You have real product photos — a can, bottle, sneaker, gadget — and you want them to show up accurately and consistently across many AI-generated video clips. This guide is written the way you actually work: you just talk to the agent in plain language. Every example below is something you can type straight into chat.

No commands, no code. Small “under the hood” notes name the underlying capability for the curious — you never type those.

1. Share the shot

Give the agent your product photo (drag it in, paste a link, or attach the file).

2. Lock the frame

Ask for an accurate hero frame with your product in the scene, and put the exact logo + text on it.

3. Animate it

Ask to animate that frame into a clip, keeping the product locked to the first frame.

4. Keep consistent

Reuse the same frame, brand colors, and product across clips, then ask to stitch them into one reel.

TL;DR — Just Say This

The single most important rule: a real product should enter the video by animating a frame you approved — not by asking for a video from scratch, which reinvents the product every time. Lock anything that must be exact (logo, wordmark, price) by asking the agent to place it on top rather than draw it.

Type this to the agent:

"Here's my product photo. Put this exact can into 4 lifestyle scenes and
keep the label identical in all of them. Place my real logo on the can —
don't redraw it. Then animate each scene into a 5-second vertical clip,
keeping the product locked to the first frame. Finally, stitch the clips
into one reel using my brand colors."

Under the hood: object placement + deterministic logo composite + image-to-video + concat/export. You don't type any of that — the agent picks the right capability from your words.

Which Approach, And When

Four ways to bring a product into generated media. Most real projects combine them: drop the product into scenes, lock the exact mark on top, and train a product model only if the product recurs across many jobs. Just describe what you want — the agent chooses.

Exact logo + text, locked on

Use when: The logo/wordmark and exact copy MUST be pixel-perfect (packaging, prices, legal, brand marks). The most reliable route — the mark is placed, never redrawn. Works on curved cans/bottles by wrapping the logo to follow the surface.

Under the hood: logo_composite + deterministic text overlay

Drop my product into many scenes

Use when: You have ONE product photo and want it in many scenarios ("put my can in 6 different lifestyle scenes"). The agent keeps the product identical across every scene. Great for a scene set you then animate.

Under the hood: place_subject

Keep it, change the setting

Use when: A single scene: keep the product identity but change environment, lighting, or style. Fast and cheap. Fine wordmarks can drift — ask to lock the text/logo afterward.

Under the hood: reference-image edit (nano-banana / kontext-edit / gpt-image edit)

Video LoRA — same product in motion

Use when: A hero product you'll animate repeatedly and need the can/bottle to stay identical shot-to-shot. Train on 10+ photos and a couple of short clips, then always animate from a locked keyframe — never describe the video from scratch.

Under the hood: Wan 2.2 video LoRA (train on photos + clips, animate keyframe)

Teach the model my product (stills)

Use when: A recurring HERO product for still campaigns where the model should know geometry and decals. Worth it for repeat use; overkill for a one-off. Needs ~10-20 photos, roughly $6–10 and 10-20 minutes.

Under the hood: product LoRA training for images, then reuse

The Product-Shot To Video Pipeline

The recommended end-to-end flow, each step written as something you say to the agent. The product identity is set once in the keyframe, then preserved through the animate step.

Step 1 — Share the product shot

Get your photo to the agent. In Claude Desktop you can attach or paste the image; if you have a link, just give the link.

"Here's my product photo — [attach the image, or paste its link].
Use this exact can as the source for everything that follows."

Under the hood: upload / hosted-URL ingest so the tools can read the image.

Step 2 — Ask for an accurate hero keyframe

Put the product into the scene while keeping its identity, then lock the exact logo and copy where accuracy matters.

Put it into scenes:
"Put this exact can into these three scenes and keep it identical:
a sunlit kitchen counter, a runner holding it mid-stride, and on ice
at a rooftop party at dusk."

Or restyle one scene:
"Keep this exact can but place it on a mossy rock beside a waterfall,
cinematic lighting."

Then lock the real logo:
"Take my Joggy logo and place it on the front of the can, wrapped to
follow the curve — don't let the model redraw it."

Under the hood: place_subject or a reference-image edit, then logo_composite for the pixel-exact mark.

Step 3 — Animate the keyframe

Animating your approved frame keeps the product faithful, because the clip starts from that exact frame. This is the biggest lever for accuracy.

"Animate this exact frame into a 5-second vertical clip. Keep the product
locked to the first frame — slow push-in, condensation glistening, gentle
bokeh. Make it fast."

Under the hood: image-to-video (fast = pixverse-i2v, balanced = seedance-mini-i2v, HQ = kling-o3-i2v).

Step 4 — Keep the look identical across clips

Describe the product the same way every time, and give the agent your brand facts once so they apply to everything.

"Remember my brand: the colors are green #00b934, pink #e42467, and near-
black #0a0a0a; the logo is the one I gave you; keep headline casing exact.
Use this on every scene and clip from now on. This is the same can in
every shot — keep it identical."

Under the hood: a brand kit / context kit plus (for a hero product) a project-attached product model, auto-applied per scene.

Step 5 — Finishing pass

Stitch the clips, size them for the platform, and add a brand mark.

"Stitch these clips into one reel, export it vertical for TikTok, and put
my logo as a small brand bug in the corner."

Under the hood: ffmpeg concat / export / overlay finishing steps.

Full Worked Example — A Canned Drink

One product photo becomes a 4-scene vertical reel with pixel-exact branding. You can say it all in one message, or step through it.

Say it all at once

"Here's my can photo [attached]. Grab the exact Joggy logo from
getjoggy.com. Put the can into 4 lifestyle scenes — kitchen counter,
desk beside a laptop, held mid-run on a city street, and on ice at a
rooftop party — and keep the can identical in all of them. Place the exact
logo on the can, wrapped to the curve; don't redraw it. Then animate each
scene into a 5-second vertical clip, keeping the can locked to the first
frame. Stitch them into one reel with my brand colors and check that the
logo and text are exactly right."

Or step through it

1. "Here's my can photo [attached]. Use it as the source."
2. "Capture the exact logo and colors from getjoggy.com."
3. "Put the can into these 4 scenes, identical each time: ..."
4. "Place the real logo on each can, wrapped to the curve — no redraw."
5. "Animate each scene into a 5s vertical clip, product locked to frame one."
6. "Stitch them into one reel in my brand colors, then check the logo/text."

Works identically in Claude Desktop, the Livepeer Agentwebapp chat, and the CLI chat — it's the same natural-language request.

Training A Product Model (For Hero Products)

If the same product shows up across many future videos, you can have the agent train a small model on it so it “knows” the product's shape and label. Restyle edits invent a fake brand name most of the time; a trained model fixes that. Worth it for a recurring hero product — overkill for a single clip.

Gather good photos

  • 5-10 sharp photos minimum; 15-25 for a tricky product.
  • Vary the angle (front, 3/4, side, top), distance, and lighting.
  • Mostly plain backgrounds so it learns the product, not the room.
  • Keep the label readable and the product large in frame.

Ask for it in plain language

"Train a product model on these 15 photos of my can so you can reuse it
in future shots. Call it 'joggycan'. When it's done, use it to generate
the can on a beach at sunset, and still place my exact logo on top for
crisp type."

Under the hood: submit_lora_train → apply_lora / attach_lora_to_project. Rough cost/time: about $10 and 10-20 minutes.

Highest fidelity: combine both — let the trained model carry the product's shape and materials, and still ask the agent to place the exact logo/wordmark on top so the type is pixel-perfect rather than model-drawn.

Video LoRA — when motion must match every time

For a hero product you'll animate in many reels, train on photos plus one or two short clips, then always animate from a locked keyframe. Say you want the product “trained for video consistency” — the agent picks the video trainer.

"Train my can for video consistency on these 12 photos and 2 short
clips. Call it 'joggycanvideo'. When it's ready, animate THIS keyframe
into a 5-second vertical clip and keep the can identical to frame one."

Tuning For Best Performance

What to say to push a result toward accuracy or toward creativity.

LeverWhat it doesDefaultWhat to say
How the product entersWhether you drop it into scenes, restyle one scene, or teach the model."put my product into these scenes" (place)More accuracy: ask to lock the exact logo on top and, for a hero product, train a product model. More freedom: just describe a new setting.
Animate from a frame, not from scratchAnimating your locked keyframe conditions every frame on the real product; describing a video from text reinvents it.always animate the keyframeAccuracy: "animate this exact frame, keep the product locked to the first frame." Ambient B-roll only: "generate a video of..."
Speed vs fidelityHow polished the motion is.fastSay "make it fast" for quick drafts, "balanced" for steadier multi-shot consistency, or "in HQ" for premium 4K hero clips.
Lock the lookKeeps re-runs matching instead of drifting.let it varySay "keep the same look/seed across all clips" so a batch stays visually stable; ask to vary only when exploring options.
How strongly to hold the productHow tightly the original identity is enforced.balanced"Hold the product exactly, don't restyle it" for maximum fidelity; "you can reinterpret it a bit" to let the scene take over.
Same product every shotBias toward the product looking identical across shots.balancedSay "this is the same can in every shot — keep it identical" when consistency matters most.
Curved surface fitHow the logo wraps onto a can/bottle.flat"Wrap the logo to follow the curve of the can" for cylinders; "keep it flat" for boxes and flat packs.
How much to train (hero product)More training = sharper identity, longer wait.standard pass"Do a quick training pass" for speed, or "train it thoroughly" for a stubborn product. Diminishing returns past a heavy pass.
Prompt structureWhat the model weights most.product facts firstLead with the fixed facts ("a slim matte-pink 250ml can, black Joggy wordmark"), THEN the scene and motion. Keep those facts identical across clips.
Clip lengthLonger clips drift more.3-5s per clipKeep product-critical clips short and cut them together; save long single takes for ambient shots.

Accuracy & Consistency Checklist

  • Product enters by animating an approved frame, not a from-scratch video.
  • Exact logo/wordmark placed on top, never redrawn.
  • Product described the same way in every clip.
  • Same look/seed held across the batch.
  • Brand colors given to the agent once and reused.
  • Clips kept short (3-5s) to limit drift, then stitched.
  • Final reel checked for exact logo and text.

Common Problems & Fixes

Logo drifts / wrong shape
Don't let the model draw it. Ask the agent to place your exact logo on top, or train a product model so it stays right.
Wordmark / text garbled
Never trust AI-rendered text. Say "generate it clean, then put the exact text on top" and ask the agent to check the spelling.
Product morphs between shots
Say "animate from this exact frame, same product in every shot, keep the same look." For many shots, train a product model.
Color shift across clips
Give the agent your brand colors once ("use these brand colors everywhere") so every clip and the final export match.
Warped / smeared label on a curve
Say "wrap the logo to follow the can's curve" and nudge its size/position. For full photoreal print, ask to train a product model.
It invents a fake brand name
Expected from restyle edits. Ask to place the real logo on top instead, or train a product model on your photos.