The split of work after ~500 generations: Doubao Work plans the shots, ChatGPT draws first frames, Meta Muse renders each 10-second shot, and they are joined into a 3-minute film
Jarvis ran 488 Meta Muse video generations and concluded that unstable long videos usually come from unsettled prompts and first frames. His final split: Doubao Work researches viral videos on the theme plus Seedance's prompting tips and writes a 10-seconds-per-shot plan, ChatGPT draws a first frame per shot, and Muse renders each clip from its frame and stitches them.
Visit site Meta Muse + 豆包 Work + ChatGPT ↗- Pricing
- Freemium
- Region
- Global
- Where it's from
- Chinese community
- Added
- Oct 18, 2026
- Version we used
- Meta Muse video (~10 s clips), Doubao Work, ChatGPT GPT Image 2.5, as described 2026-10-16
- Original author
- Jarvis|AI 商业实验室 @jarvis11x ↗
- Our test
- Not tested by us
- Cost / time
- Muse has a free weekly allowance (Meta also sells $20/mo and $100/mo plans); ChatGPT image generation needs a suitable plan; check Doubao's own pricing for Work.
The task
Make a consistent ~3-minute video from a model that only produces 10-second clips.
Steps
- Pick the theme, length and aspect ratio, and split into 10-second shots (about 18 for 3 minutes).
- Hand the idea to Doubao Work (instruction below): find viral Douyin/TikTok videos on the theme, summarize how they're shot, check Seedance's official prompting tips, then write the full plan, each shot giving the frame at 0 s, the motion over 10 s, the ending composition and the hand-off to the next shot.
- Read the plan end to end. Once character, set and prop descriptions are settled, reuse them verbatim.
- Send the whole plan to ChatGPT, have it absorb the style, then generate one first frame per shot (he used GPT Image 2.5). Fix the image before moving on if character, framing, set or light is off.
- Give Muse each first frame plus its 10-second prompt, pick takes shot by shot, and check each against its neighbors for character, motion and camera continuity.
- If a shot still fails after 10–15 tries, simplify the action or change the shot instead of generating more.
- Have Muse write an ffmpeg script that normalizes aspect ratio first (his clips mixed 3:2 and 16:9, unified to 1280×720) and then concatenates; watch the whole thing, focusing on every cut and the music.
Prompt
Write prompts for a three-minute "heavenly palace" video at 10 seconds per shot. First search the web for viral videos of this kind and summarize their content and how they're shot, then output the prompts.
What the author says
The author (2026-10-16) says he generated 488 Muse videos in about 20 days across three stages: all-Muse (a 3-minute episode, lots of rework), Doubao Work plan + Muse (an 18-shot palace film, much easier review), and Doubao Work + ChatGPT first frames + Muse (a 21-shot, 3.5-minute Ghibli-style reel with clearly better detail and light, less "AI look"). Muse ran on the free allowance.
Our test
We have no Muse account and have not reproduced this workflow. Steps and results are from the author's post.
How to get in
The original is a Chinese X article. Doubao Work is ByteDance's Chinese-language agent (doubao.com); any web-researching agent can stand in for the planning step, though we haven't compared results. Meta Muse availability varies by region (third-party guides say US and Canada, 18+).
Who it suits
Creators making AI story episodes, scenic reels or music-video-style long pieces with 10-second video models.
Pitfalls
- When long videos drift, check the prompts and first frames before blaming the video model.
- Rewording a character or set description invites drift; reuse the locked text verbatim.
- If a shot fails 10–15 times, redesign it; the same applies when native dialogue won't come out right.
- Clips can come back in mixed aspect ratios (3:2 and 16:9); normalize before stitching.
- Expect rework on music and watermarks: he went through three music versions and moved the watermark.
More in this category
- A demo video per docs page that flags itself when the UI changesVideoScreenplayClaude CodeTool free; uses your agent's quota◐ Partly testedOverseas
- Code-drawn stick-figure explainers for about $0.54 eachVideoClaude CodeFFmpegChromeNeeds Claude plan or API; renders locally○ Not tested by usChina
- One topic in, a narrated tech explainer video out, locallyVideoClaude CodeChatterboxffmpegSkill free; needs a Claude Code plan◐ Partly testedOverseas
- Render restrained motion graphics from code, frame by frameVideoClaude CodePlaywrightffmpegNeeds a Claude plan; other tools free◐ Partly testedOverseas
- One reference image → a take-apart machine simVideoteardownClaude Codethree.jsSkill free; needs Claude Code/Opus◐ Partly testedOverseas