Add directors-cut: an LLM directs, PixVerse shoots with native audio

[SamyMe]

Summary

This adds a fourth demo, directors-cut/, which shows off PixVerse's video models.

You describe a film in a sentence. An LLM director writes it as a strict JSON shot list: a "style bible" that describes the characters identically in every shot, plus N shots, each with its action, camera move, sound and length. PixVerse shoots every shot with native audio, and ffmpeg cuts them behind a title card. Any shot can be reshot with a new note, which re-renders only that shot and re-cuts the film.

Shooting modes:

• Every shot at once: parallel, about 60–65 s.
• Continuity: each shot starts from the last frame of the one before (image-to-video).
• One multi-shot take: V6 or V5.5 cuts the whole sequence itself in one generation.

Directors: eight LLMs, grouped on the page. Each is pinned to an endpoint that supports JSON-schema output, with a fallback serving the same model elsewhere.

• Anthropic: Claude Sonnet (the default) and Claude Opus.
• OpenAI: GPT-6 Astra, Sol and Luna, served by Azure.
• Open weights: Kimi K3, GLM 5.3, and DeepSeek V4 Pro (served by Nebius).

Page: a light theme (off-white background, white panels, dark video frames), with a live Direct → Shoot → Edit pipeline, storyboard cards, the final cut and a cost table.

Testing (real key, 2026-09-29)

• 4-shot film, Claude Sonnet directing: 66 s, $0.78, a 16 s film. The same rider, kit and bike appear across all four shots, every shot has native audio, and each follows its camera move.
• Reshoot of shot 3: 46 s, $0.19, matching the new note.
• 3-shot film through the page, GPT-6 Sol directing: 60 s, $0.59. The film loads and plays in the page.
• All eight directors, same premise: each returned a valid 4-shot plan, in 4.5–23 s. GLM 5.3 took 44–90 s.
• Offline with fakes: all three modes, reshoots, silent and mismatched clips, 16:9 / 21:9 / 9:16, and the page flows.

Notes

These are problems on the providers' side or in the Eden AI platform, not in the demo:

• OpenAI's own endpoint returns 401 Incorrect API key for GPT-6 (and for OpenAI TTS), with is_byok: false. That's Eden AI's platform OpenAI key. The demo uses Azure, with OpenAI as the fallback.
• DeepSeek's own endpoint rejects the JSON-schema response format, so the demo uses Nebius.
• GLM 5.3 can't have its thinking turned off (Z.ai rejects the request), so it's labelled as slower.
• Eden AI's MCP video_generation tool takes no provider_params, so it can't switch on PixVerse audio or multi-shot. The demo calls the REST API instead.
• PixVerse options on Eden AI: camera_movement, style and last_frame_image aren't on the allow-list, so camera moves are written into the prompts.

🤖 Generated with Claude Code