Private alpha

A photo of you pulling a face is not a thumbnail.

Postlume builds the whole frame — subject, composition, and the headline as an editable layer — from an identity kit you upload once. Choose one to three deliberately different shots per video, at 1280×720, because which one earns the click is something you find out by testing.

No password — Google, or a link to your email.

One topic, up to three shots

  • You, large, on one sideface-forward
  • The thing the video is aboutobject-forward
  • The words carry ittext-forward
The three

One video. Up to three genuinely different bets.

Not three variations on one idea — three different answers to what should carry the frame. Choose only the shots you want; the model and number of shots determine the credits shown before you start.

  • Face

    You, large, on one side

    The expression does the work, and the other half of the frame is left open for the headline.

  • Object

    The thing the video is about

    The game, the product, the chart — front and centre, with you supporting it at a smaller scale.

  • Text

    The words carry it

    Built around the claim, with the subject supporting from the edge. For videos that sell on the sentence.

  1. 01

    Build an identity kit, once

    Five to ten photos, your logo, your palette. Faces go to the model as reference images; logos never do — they are composited at exact pixels, because a model asked to draw your logo invents the letterforms.

  2. 02

    Choose one to three shots

    Curated presets give each shot its composition. Pick the ones that fit this video, then choose an available model when you want to steer the result.

  3. 03

    Judge them at feed size

    Each selected shot has its own displayed credit cost. Preview the results in a mock feed in both YouTube themes, because a thumbnail that reads at full size can vanish at the size people actually decide at.

What’s different

The parts a model gets wrong, we don’t ask it to do.

Image models are very good at scenes and lighting, and unreliable at anything that has to be exactly right. So the things that must be exact are handled as layers instead.

  • Text is a layer, not a guess

    The headline is typed, not generated, so it is crisp at any size and rewording it costs nothing. Baked-in AI text stays available as a style, never as the architecture.

  • Your logo is placed, not described

    Brand marks composite at exact pixels with transparency intact. The alternative is something logo-shaped in roughly the right colours, which is worse than no logo at all.

  • Choose the model for the job

    Available models show their credit cost before you generate. Leave the choice automatic, or pick one that suits the shot you are making.

Platforms

Correct dimensions, without looking them up.

Sizes, safe zones, and file limits live in data, not in a blog post you found in 2023. Every tool inherits them, so the twelfth feature is as platform-correct as the first — and when a platform changes a spec, it changes in one place.

Import a photo to reframe it, add a border or place a logo from your kit. Export the result without running a model.

Launch set — excerpt
YouTubeThumbnail1280×72016:9 · under 2 MB
YouTubeBanner2560×1440safe area 1546×423
XHeader1500×5003:1
InstagramPost1080×10801:1
PinterestPin1000×15002:3
LinkedInCover1584×3964:1

Build the kit once. Ship every upload on brand.

It is early — this is a private alpha, pricing is not set, and the studio is still growing. What works today is the part that matters most: identity in, test-ready thumbnail shots out.

Start free

Google, or an email link