The thumbnail is a CTR device — not a pretty image but a reason to click. Always pull a variant set (4 by default) and hand it to A/B testing. The generation policy is the same as the image-gen skill: all generation goes through the Higgsfield CLI, default model gpt_image_2 — do not silently switch to another path (failure = report).
Hard rules
- No single image, always 4+ variants. At least 3 should be competing candidates for the same video — so they can be entered directly into YouTube "Test & compare" (3-thumbnail A/B).
- Copy = matches the video's first claim (hook). If it clashes, it comes back as drop-off. ≤12 chars (KR) / 3–5 words (EN), within 2 lines, left placement (the bottom-right on mobile is covered by the duration badge). Keep one number/stake visible ($20, 99%). No clickbait.
- Full generation including the copy is the default — the finished thumbnail (background + person + text) in one
gpt_image_2 shot. The prompt must be exactly line by line:
first line reads exactly "…" in <color>, second line reads exactly "…"+ "render every character EXACTLY". After generation, character-by-character READ verification is mandatory — a common failure is a missing symbol ("$20"→"0"). Typos = regenerate with the same model (you may simplify the copy to be shorter), no inpainting, no shipping typos. HTML/PIL overlay is the last-resort fallback used only when full generation repeatedly fails (after reporting to the user; no hard rectangles for scrims, only directional gradients that fade to transparent). - Real faces = full generation with cutout references. Use only user-provided photos, no inventing
faces. Do not insert the photo as-is: rembg(
birefnet-portrait) cutout → remove burned-in subtitles/caption bands before cutting out (set that area's alpha to 0 → crop to the alpha bbox) → collage on a white background into a 한 장을--imagereference and full-generate with "preserve the EXACT faces". For 2+ people, combine into one image with left-right placement and name LEFT/RIGHT in the prompt. If you need to change hair/clothes, edit the reference first (--image+ "keep exact face, change ONLY the hair to …") and then generate with that. Compositing a cutout onto a gradient background with PIL is forbidden — the look splits into cheap. - At least 1 of the variants must use the target channel's actual feed grammar. Before generating, look at 3–5 thumbnails from that channel (or top channels in the same genre if none) and extract the grammar — colors/highlight boxes/ speech bubbles/composition — and make a variant that copies it exactly. The rest are dramatic high-CTR family (neon, exaggerated expressions). If all 4 are the same look, it's not A/B but the same card 4 times.
- Verify by eye, at actual display size: shrink each variant to 168×94 (YouTube list size) and READ — if the copy is readable and the expression is visible, it passes. Compare faces against the reference photo, and regenerate if distorted. Do not pass it by looking only at desktop size.
- Setup/commands/JSON parsing/download snippets are in the image-gen skill —
higgsfield account statuscheck account and credits before generating.
Spec
| Target | Size | Notes |
|---|---|---|
| YouTube long-form | 1280×720 (16:9), <2MB | --aspect_ratio 16:9Generate as → resize to 1280×720, JPEG q90 if over 2MB. Upload to organic-social's mediaItems[].thumbnail |
| Shorts/Reels cover | 1080×1920 (9:16) | First-frame rule takes priority — for the cover, account for the IG grid 1:1 center-crop safe area (y 420–1500) |
| Community/banner | By use case | Route to image-gen |
Workflow
- Gathering materials: the video's hook/first line (use the transcript as-is if available), face photo (→ create a cutout reference per rule 4), brand tokens (DESIGN.md), target channel thumbnails observe 3–5 (grammar extraction per rule 5).
- Designing variation axes (4 by default): ① Copy the channel grammar ② Person reaction + large copy ③ Typography centered on a number hook ④ Result/before-after or curiosity gap. Four images with overlapping axes are not an A/B.
- Batch generation: per variation, include composition + per-line copy·color + "render every character EXACTLY"
- cutout reference in one prompt
higgsfield generate create gpt_image_2 --image ref.png --aspect_ratio 16:9 --quality high --resolution 2k --wait --json. For Korean copy, specify meaning-unit line breaks in the prompt.
- cutout reference in one prompt
- QA (in order): character-by-character READ → 168×94 downscale READ → compare against face reference → resize·file size. If any one fails, regenerate only that variation.
- Delivery:
thumb_A~D.png+ a one-line axis description for each variation + the top recommendation and why. If YouTube, instruct to run 3 in Test & compare; learn the next set from CTR results.