OyJI Thumbnail
Adapted from higgsfield-youtube-thumbnail (MIT, (c) 2026 Higgsfield AI): the concept gate, the reference manifest, the prompt order, the post-render gate and surgical tweaks. Tools, models and the text overlay are ours. See ../../NOTICE-higgsfield.md.
Needs the OyJI MCP (generate_image, upload_from_url), model nano-banana-pro, and local Python with Pillow for the headline (scripts/thumb_text.py). Images are free.
Rules
- Respond in the user's language. Keep the model mechanics out of the chat.
- Truth. The thumbnail may only promise what the video really delivers. Never invent claims, results, products, people, screenshots or numbers.
- A style-reference thumbnail is for looking at (energy, framing, palette, emotion) and is never passed to the model as a reference image: copying its identity or exact layout is not allowed.
- Ask only what the brief does not answer, in one compact question: the video's topic and promise, who appears (0 to 3 people), a logo, a headline (2 to 4 words), the ratio.
- If a concept needs a person and no face photo was given, ask whether to use the user (their face profile, see
oyji-face-id), another provided person (with their permission), or an invented character. Never choose silently.
Ratios (OyJI: 1:1, 3:4, 4:3, 9:16, 16:9)
16:9 for YouTube; 9:16 for Shorts, Reels and TikTok (faces in the upper two thirds, the bottom is covered by the app's own buttons); 3:4 for an Instagram feed cover (OyJI has no 4:5). The picture comes at roughly 1 to 1.5 thousand pixels on the long side: enough for these uses; say so if 4K is asked.
Concept gate
Think of at least five truthful concepts internally and pick the strongest. Prefer one focal subject, minimal clutter, and an information gap: the viewer sees something that makes them need the answer, and the video gives it. Useful frames: a reaction face next to the one surprising object; a huge scale contrast; a number or result that begs "how"; a before/after (only when asked); a mystery object partly hidden; a versus (a split layout only if asked); a person pointing at nothing is not a concept. It must read in under a second at about 120 px wide: check by shrinking the result to 120 px and looking at it.
Emotion ladder when the user gives only a count: shock, hype, anger, awe, laughter, fear, smugness, charisma, confusion, determination, disgust.
References and the prompt
Reference order: faces first (character order), then the logo, at most 6 in all (nano-banana-pro). When two or more are attached, the first line of the prompt names them: IMAGE REFERENCES: image 1 = CHARACTER 1 face; image 2 = brand logo.
Prompt order (our wording):
- Frame: a bold, punchy, poster-grade photoreal thumbnail, the chosen ratio, one unified frame (no split-screen unless asked).
- Scene: exactly the content the user described.
- Text: default
No text, no readable UI labels, no watermark.The headline is added afterwards (below). - People: large, foreground, chest-up or medium close, about 40 to 60% of the frame;
All faces crisply sharp.For each referenced person:CHARACTER N: the person from image K, exactly the same face: bone structure, eye shape, nose, lips, jaw, skin tone, hairline. Do not beautify, average or restyle. Expression: <emotion>. - Key elements: only the props that explain the gap.
- Logo: exact shapes, colours and letters, away from faces. (A glossy 3D logo is a separate image first; use it as the last reference.)
- Place, composition: one hero at a power third, clear scale hierarchy, strong subject-background separation.
- Background: a vivid high-contrast colour field or environment with a soft vignette.
- Light on people: a strong key light sculpting the face, a soft fill, and a bright rim light tracing the hair and shoulders.
- Grade: vivid, glossy, poster-punchy, deep blacks; restrain only for an explicit calm or premium brief.
Make one call per distinct concept or expression (count 1 to 2 is fine); vary only the concept, expression or camera line. Cap the whole job at about 16 generations.
Check every result (look at it)
Identity matches the references; no stray text or watermark; the face and the hero element still read at 120 px; the concept is true to the video. A hard failure: rerun the same prompt at most twice. If you cannot view the image, say so and let the user judge; do not claim it passed.
Surgical tweaks
Use the picked result as the only reference and strict: true: Change ONLY <the expression / the background colour / the rim light>. Everything else stays exactly as it is. Allowed scopes: expression, background replacement, background recolour, rim-light recolour. Never regenerate the whole picture for a small request.
The headline (local, never from the model)
python3 -I scripts/thumb_text.py spec.json, one headline of 2 to 4 words, in Kazakh or Russian. Presets: beast (yellow, thick black stroke), fire (orange to red), lime (neon lime), clean (white with soft shadow).
{"image":"pick.png","out":"thumb.png","text":"БҰЛ НЕ БОЛДЫ?","preset":"beast","position":"left","size":0.16}
position: left, right, top, bottom, center. Keep text off the face; if the base picture has no free area, regenerate the base with a clear zone instead of covering the subject. The script uses a font that has the Kazakh letters and stops if none does.
Deliver
The passing pictures with a short label each (for example shock / close-up), the ratio, and whether each is clean or has the headline. No prompts or IDs unless asked.