render-ig-live-gallery

Render the ig-live-gallery format from a config. Each of the brand's REAL hero product stills is presented as the "live video" inside an authentic Instagram-Live card — dark gradient scrims top & bottom over the bright product, an IG gradient-ring avatar + username + verified check + red LIVE badge + viewer count + close X across the top, a short live-comment feed + an EMPTY "Add a comment…" bar + heart/share icons across the bottom, and pink/red reaction hearts floating up the right edge. A benefit sentence builds one short phrase per slide as the top overlay, and it closes on a clean brand logo card with the payoff line + CTA.

Deterministic and FREE — no generative image/video, no AI-rendered text. The paid music bed is a separate capability (create-music-elevenlabs) muxed on afterward.

Run

python3 scripts/build.py --config config.json --assets <stills-dir> --out master-silent.mp4
# quick check — one frame per slide:
python3 scripts/build.py --config config.json --assets <stills-dir> --stills <dir>

config.example.json is a full worked config (Glossier) — its phrases, comments, payoff and music prompt are that demo's creative answers, not defaults. Config shape:

• username, verified, palette, slide_dur, endcard_dur, crossfade
• slides[] — { image, phrase, viewers, comments: [[handle, text], …] } per product
• endcard — { logo, logo_is_white, payoff, handle, cta }
• music — { prompt, length_ms, trim_intro_sec } (consumed by create-music-elevenlabs)

Image paths in the config resolve against --assets.

Choices

The recipe asks the user these before any paid step; the renderer just draws what the config says.

• slide_story — what the slides show and how the sentence builds → slides[].image, slides[].phrase, endcard.payoff. Asked of the user; the demo used a makeup routine, one product per step.
• copy_tone — voice of the phrases and live comments → slides[].phrase, slides[].comments. Asked of the user; the demo used a soft, effortless beauty voice.
• music — the bed's genre/mood → music.prompt. Asked of the user; the demo used soft dreamy pads + bells.

Craft rules (baked into the renderer)

• Center each product on its TOP-HALF silhouette, never the full non-white bbox — the base drop-shadow otherwise inflates the box and the product reads shifted sideways. The renderer measures the top 50% of the silhouette with an L<246 threshold (catches white-product edge shading too) and centers that midpoint.
• Reaction hearts float up the right edge as a full column and may overlap the centered product — that's fine; don't shrink them to a thin edge ribbon.
• Comments must be claim-free brand-descriptor vibe (2-3 words, generic handles, NO emoji — PIL can't render color emoji); the comment INPUT bar stays empty. LIVE badge + viewer count + hearts are ambient live-stream chrome, not proof claims.
• Endcard is the clean logo card (bg + wordmark + @handle + CTA + payoff), never a model face.
• Real product stills + real wordmark only. Music bed only, no VO.

Output

A silent WxH (default 1080x1920) H.264 master. Generate + mux the create-music-elevenlabs bed with a short fade in/out + loudnorm (-map 0:v:0 -map 1:a:0), then QC with watch.