AI Thumbnail Generator on Astorie

The thumbnail generator produces click art for organic discovery: YouTube thumbnails with the headline baked in, podcast covers, Shorts and Reels verticals, Open Graph share cards, and A/B variants — all anchored to the same host portrait and brand palette.

Channels live and die on thumbnail consistency. A YouTuber publishing four videos a week needs every thumbnail to share the same host portrait, color script, and a readable episode title — across 200 videos a year. The first thumbnail looks great; the fiftieth has drifted on the host's face shape and the title is illegible against a busy backdrop. Without anchors, the channel reads as a stack of unrelated videos rather than one show.

Platform variants compound the problem: the same content ships as a 16:9 YouTube thumbnail, a 1:1 Spotify cover, and a 1200x630 share image, each with its own layout constraints. Tab-based tools generate one image per session and force manual re-crops per platform.

And there is the legibility problem. Most general-purpose image models still render in-image text poorly — 'EASY 30-MINUTE PASTA' comes out as 'EASY 30-MINUTE PSTAA' or worse. Ideogram handles short legible text better than its peers, but the fix requires a chain — Ideogram for the text pass, a photographic model for the backdrop — which single-model tools cannot express.

Anchor the host portrait and the brand color script once. Drop one clean host headshot as a labeled image node and the channel palette as another — both stay locked across every episode, so the whole season fans out from the same face and the channel stays on-model.

Add the per-episode prompt as a text node — the dish, the headline, the energy — and run the title-text pass through an Ideogram node. For short copy (4-8 words) the headline bakes into the composition legibly without a separate Figma overlay step.

Fan out the photographic backdrop in parallel across Flux (high detail), Midjourney (editorial composition), Nano Banana 2 (host-face fidelity), and GPT Image 2 (edit-aware refinement). Compare takes against the Ideogram text variant and pick the strongest combination per episode.

Generate platform variants from the approved take: once the YouTube 1280x720 lands, fan into the 1:1 Spotify cover, the 1080x1920 Shorts vertical, and the 1200x630 Open Graph card on the same canvas. If the text needs a contrast tweak, pipe the thumbnail through GPT Image 2 — edit-aware refinement keeps the rest of the image stable.

Save the canvas as the channel template; the next drop swaps only the headline and the host expression. The thumbnail is also the upstream of the video campaign — reuse the same anchors in /workflows/ai-product-video or /workflows/multi-shot-short-film so the thumbnail and the video ship from one visual lockup.

Astorie anchors the host portrait once and the brand color script once on the canvas, then fans every thumbnail off both anchors. The cooking-show host is one labeled image node; the brand color script (the channel's two-color palette and overlay style) is another. Every weekly thumbnail wires both anchors into the model node, plus the per-episode prompt — pasta dish, Asian noodle, cocktail, sheet-pan dinner. The host face stays consistent, the color script stays consistent, the channel reads as one show across 200 episodes.

Multi-model fanout for the text and the photo. The wedge here is Ideogram for the title-text-baked-into-image pass — 'EASY 30-MINUTE PASTA' rendered legibly inside the thumbnail rather than overlaid in a separate Figma export. For the photographic backdrop and host expression, fan out across Flux, Midjourney, Nano Banana 2, and GPT Image 2. Each thumbnail picks the strongest take from the fan-out; the chain runs every week without re-uploading the host portrait.

Cross-platform format generation chains downstream. Once the YouTube 1280x720 thumbnail lands, chain into a 1:1 Spotify cover variant, a 1080x1920 Shorts thumbnail, and a 1200x630 Open Graph share image — all anchored to the same host portrait and brand color script. The platform variants share visual identity automatically. Save the canvas as a template after one episode lands; future drops swap the per-episode prompt only and the chain re-renders everything in one go.