dhee-bundle-ugc-ad-product

A themed product / UGC video ad pipeline for Dhee. Give it a real product image, a short description, and a theme (e.g. Rainy season, Festive tea-time) and it produces a short, alive, multi-beat video ad with a designed, multilingual voiceover. Product-agnostic — all specifics come from the inputs.

Bundle: bundles/ugc_ad_product_v2.

The keystone rule

A product's printed label/text must never be re-synthesised by a diffusion model (it re-letters fine print). This bundle honours that two ways:

1. Keep it far / illegible — the product is model-composited (boogu) into the scene small and set back, so the label is sub-readable and any drift is invisible by design.
2. Real pixels for what must read — the brand logo end-card is the actual logo, overlaid (never regenerated). Brand recognition rides the logo + the voiceover, not the tiny on-pack text.

Pipeline

product image → comfy.matte (SAM 3 concept extract → product on white)
             → comfy.boogu  (composite far/blended into the themed scene)
             → comfy.ltx_director (alive motion)            ── ×3 beats
designed voice (omnivoice_design) → emotive voice acting (omnivoice_clone)
             → ffmpeg.concat → ffmpeg.overlay (logo end-card)

Inputs

product_description, product_image, product_concept (what SAM 3 segments), theme, language (English / Kannada / Telugu / Hindi / …), voice_style (e.g. calm and warm or high-energy voice acting), brand_logo, aspect.

Voice

Voice is designed then acted: tags (incl. indian accent) → a designed voice → each line acted in it with emotion (temperature/guidance_scale). voice_style drives the script energy.

Runners

llm.generate, comfy.matte, comfy.boogu, comfy.tts, comfy.ltx_director, ffmpeg.*.

Sample inputs in inputs/ are GRL Foods (Mangaluru) marketing assets, used as the demo product. Replace them with your own product image, description, and logo.