Full workflow: ultra-realistic AI vlog video with Higgsfield API and Seedance
EXM7777 · x · 2026-09-17
The author tested Higgsfield's newly released API in real conditions and produced a Paris vlog that looks like raw phone footage, sharing the complete 7-step workflow:
- Steal structure, not pixels: capture 3-5 scene structures from references, pull character inspiration from Pinterest and real vlogs
- Use Gemini to analyze collected videos/audio/keyframes into a shot-by-shot breakdown
- Build reference assets first with Nano Banana Pro: character sheet, per-location mood images, exact start frames — images are cheap, video is expensive
- Send the full timestamped script to Seedance 2.5 as ONE 30-second generation; one action per beat, identity locked at the end
- Revise like a director: regenerate only failed scenes from their start frames and splice with ffmpeg, changing one variable at a time
Key rules: iterate on images, commit on video; the start frame decides 80% of the shot; lock identity explicitly or it drifts between cuts.
More from Multimodal
- AI-generated 'Sweet Companions' made with Grok Imagine and Google Veo — creatoroff · 2026-09-18
- Cartesia's Multilingual Voices tops Speech Arena; Mamba author Albert Gu on board — alex_verem · 2026-09-18
- AssemblyAI demos Dictation API: custom key terms fix vocabulary, sub-second transcription in 19 languages — AssemblyAI · 2026-09-18
- Minimax H3 ecosystem roundup: audio gating, LoRA trainers, comic-to-video workflows — optimisticalish · 2026-09-18
- Runway CEO teases 'incredible things' ahead, says good research takes time — c_valenzuelab · 2026-09-18
- Magnific's first AI Original "The Sorrowful Figure" premieres in Madrid, streams Sept 24 — cuenca · 2026-09-18