Higgsfield Faceless Video

Production workflow

1. For factual topics, research reliable sources and retain a short source list. For personal stories, use supplied facts. Plan one illustrated scene and one narration line per 10-second block; use a shorter final block when the requested duration is not a multiple of 10. Avoid cutting a sentence. For a final visual-only tail of at most three seconds, create silent audio and use --last-visual-only.

2. Create a style key with a supported image model. For a selected preset, resolve it via the CLI and follow the returned media contract. Reuse the style and any selected character consistently. Preserve the user's chosen aspect ratio. Check a first rendered clip before submitting the rest; inspect each result for visual and identity drift.

3. Select one voice and keep the exact voice_id and voice_type pair. For a full 10-second block, start with approximately 26–30 spoken words in one flowing line and use a consistent delivery cue and a nine-second time window, for example [voice: warm, clear, lively conversational delivery, starts speaking immediately] [00:00-00:09] <spoken line>. Keep the cue out of the authored caption text. This is a starting point, not proof of timing or wording: inspect the generated take and adjust its text for the selected voice and language. For a shorter final block, plan to its actual window. Inspect the current seed_audio schema, then use higgsfield generate create seed_audio --prompt "<line>" --voice-id <id> --voice-type <type> --wait --json. Download each completed take and, before generating video, run bash "$SKILL_DIR/scripts/check_voice.sh" voice01.wav 10 (use the actual shorter window for a short final block). Each spoken take must end within [window−2.0, window−0.2] seconds; rewrite and regenerate a failing line, never speed up the audio. Allow at most three attempts per line, then report the timing failure without claiming a completed result. Keep each job ID, result URL, and block number as soon as returned.

4. Inspect the current minimax_h3 schema. Submit a 10-second background for each block with the chosen aspect ratio and style reference when supported; pass the completed style image's job ID directly when the media parameter supports it. Wait for terminal results. Use the returned job output URL, never construct one. Generate speech separately and exclude spoken words from background prompts. Respect the live model schema rather than passing unavailable flags.

5. Download completed clips and voices to the working directory using ordered names such as clip01.mp4 and voice01.wav. Confirm every file is nonempty and decodable. Assemble in order, for example:

bash "$SKILL_DIR/scripts/assemble_blocks.sh" --out final.mp4 \
  --last-seconds 5 clip01.mp4 voice01.wav clip02.mp4 voice02.wav

The example makes a 15-second video. Omit --last-seconds for full blocks. The script validates speech timing, duration, audio and decoding and writes an assembly receipt. For a requested music bed, inspect sonilo_music, generate one instrumental track with higgsfield generate create sonilo_music --duration 40 --prompt "<music description>" --wait --json, download it and pass it via --music; omit it when music is declined. If a gate fails, correct the input and rerun; do not bypass it or return loose clips as a completed video. Review the final video and audio for content, sync, framing and visual continuity. 6. If captions were requested, follow references/captions.md, inspect their timing and legibility, then use the captioned file as the deliverable. 7. To provide a hosted result when requested, run higgsfield upload create final.mp4 --json and use the returned confirmed URL; otherwise deliver the local file if the host supports it. Report actual duration, aspect, language, subtitle status and sources.

Resume from saved completed outputs. Poll running jobs with higgsfield generate wait <job_id> --json; do not submit duplicates while a job is running. Respect generation failures and refusals.