Jimeng (Dreamina) Workflow Guide (Beta Feature)
Hand only the picture and native audio to Jimeng, while storyboards, dialogue, subtitles and stitching stay in this project — an alternative path when the Agnes API keeps returning 503.
During peak hours the Agnes free video queue keeps returning HTTP 503 (video queue is full), with waits of ten minutes or more, so automatic mode often gets stuck at the video generation step. Jimeng (Dreamina) offers another way: it grants daily credits, its Seedance video models support native audio, and the picture is stable. That is why v7.6 adds External Asset mode — the whole storyboard → voiceover → subtitles → stitching pipeline stays exactly as it is, and only the video generation step moves to Jimeng. Jimeng handles the picture and native audio; this project handles orchestration, normalization, subtitles and compositing. The feature is still in beta, so give it a try, and feel free to send us the work you make with it.
About this beta feature
External Asset mode has been available since v7.6 and is still being polished (checkpoint prompts, batch import and subtitle alignment may all keep changing). If you run into a problem, please report it with the task ID; if you have a free video generation tool that works better for you, tell us too and we will evaluate adding support.
The strengths and weaknesses of Jimeng
👍
- ✓Daily credits, so you can try a finished video for free first
- ✓Native audio: dialogue and ambient sound are generated together with the picture, saving you a TTS step
- ✓Stable video generation — your odds of getting a finished clip are far higher than gambling on the queue
⚠️
- –Free credits are limited; once they run out you need to pay or wait for the next day’s refresh
- –The free tier only allows lower-tier models, so picture quality and detail lag behind paid tiers
- –Subtitles Jimeng paints into the picture often go wrong (typos, broken glyphs), so do not count on it for on-screen captions
Compared with automatic and manual mode, there is just one difference
| Step | Automatic / manual mode | External Asset mode (Jimeng) |
|---|---|---|
| Storyboard and prompts | Generated by this project | Generated by this project, plus a per-scene prompt set you can paste straight into Jimeng |
| Picture and audio | Agnes video models (the API frequently returns 503; only the voiceover comes from this project’s edge_tts) | Generated by Jimeng (with native audio), exported as one mp4 per scene and imported in a batch |
| Subtitles | Aligned to word-level timestamps from the voiceover (more accurate) | Estimated from each storyboard’s real duration (may drift, check it at the subtitle checkpoint) |
| Stitching and final cut | Project stitches | This project normalizes + stitches while keeping the original audio track + burns in subtitles |
The end-to-end workflow
When creating a task, choose “🌐 External assets” and set “Video generation tool = Jimeng”, then leave the pipeline to automatic mode and the generation to Jimeng:
Create a task and choose “External assets”
There are three execution modes: ⚡ Automatic / ✋ Manual / 🌐 External assets. Choose the third one, and in the expanded options confirm that the video generation tool is “Jimeng (Dreamina)”.
Set the audio source and subtitle source
Audio source: use the video’s own audio (recommended — keeps Jimeng’s original track and this project stops generating TTS) or use this project’s voiceover (replaces the original track with an edge_tts narration). Subtitle source: narration (default) or character dialogue (the subtitle text comes from each scene’s dialogue).
Copy the prompts to Jimeng at the storyboard checkpoint
This project produces the storyboard first. The “🎬 Jimeng collaboration” card gives you a creative brief with picture, audio and duration constraints; at the “Video clips (videos)” checkpoint the card’s prompts are this project’s own storyboard script, numbered [Scene i/N · duration], ready to paste straight into Jimeng and generate clip by clip.
Export videos scene by scene
Export each clip as mp4 after generating it in Jimeng. Jimeng’s storyboard video tiers are 5s / 10s; when they differ from the project’s duration tiers (4 / 5 / 6 / 8 / 10 / 12 seconds), the import aligns on the actual duration and never truncates the final video.
Batch import to fill them back in
At the Video clips checkpoint, use “Batch import” to select several mp4 files at once. The system normalizes them to the task resolution, 30fps and H.264/AAC, then lists duration, resolution, frame rate and whether each has an audio track (a missing track gets a yellow warning but does not block you).
Check at the subtitle checkpoint, then render
With “Use the video’s own audio” selected, the voiceover checkpoint passes automatically (no pause) and subtitles are generated with a cumulative offset from each storyboard’s real duration. Check sentence breaks and the timeline at the subtitle checkpoint; once you confirm, this project stitches while keeping the original audio track and burns in subtitles to produce final_video.mp4.
What each of the five checkpoints does
- 1
Storyboard scenes
Confirm the storyboard script and narration; the “🎬 Jimeng collaboration” card provides a creative brief with picture + audio + duration constraints that you can copy straight into Jimeng or any LLM to refine.
- 2
Reference images
Optional. Import the storyboard images Jimeng produces (named like scene_1.png); they are scaled to the task resolution and padded to center. Creative long videos register imported images as that scene’s reference image.
- 3
Video clips
The core step. The card’s prompts are the storyboard script you can paste straight into Jimeng; after generating, batch import the mp4 files. Before you confirm, the system checks that every scene has its video and refuses to continue if any is missing.
- 4
Voiceover
With “Use the video’s own audio” it does not pause and passes automatically; with “Use this project’s voiceover” it works like manual mode, where you preview and regenerate the narration.
- 5
Subtitles
Check the SRT’s sentence breaks and timeline: the last cue must not run past the audio track, adjacent cues must not overlap, and Chinese lines must stay within 18 characters. This is the place in the Jimeng path that most needs a human look.
How to name exported files so nothing goes wrong
On import the system maps each file to the Nth scene using the first number N in its filename. For example, if you export scene_1.mp4 it lands in segment 1 (scene_0/video.mp4); if you export scene_3.mp4 it lands in segment 3 (scene_2/video.mp4).
- →scene_1.mp4 → segment 1 (scene_0/video.mp4)
- →scene_3.mp4 → segment 3 (scene_2/video.mp4)
Two things to keep in mind: the number is only used when it falls between 1 and the scene count, so a filename carrying a year or timestamp (such as 20261015.mp4) will not be mistaken for scene 2026; and files with no usable number are mapped in upload order. To assign manually, pass an explicit mapping at import time (for example 1 → scene_2) to put the first file into the third scene. Manuscript-type tasks use a para_N directory, and digital-human tasks have only a single clip/clip.mp4.
- •Always keep the audio track when exporting: turn on Jimeng’s “Generate audio” switch before exporting, or the finished video will have no sound.
- •This project requires every storyboard to share the same resolution; it does not matter that Jimeng exports come in all shapes and sizes, since import scales and pads them uniformly.
- •The easiest way to misalign is mixing in files from another task in the same batch, so glance at the echoed scene mapping after import before confirming.
The four most common pitfalls
1The clip has no sound: nine times out of ten the Jimeng export carried no audio, or you exported a silent version. Just re-export an mp4 with audio and import it again; the import echo tells you directly via has_audio.
2Subtitles are a second or two early or late overall: subtitles are estimated from a cumulative offset based on each storyboard’s “real duration”, so any mismatch in Jimeng’s speaking pace or line length accumulates. Editing the SRT by hand at the subtitle checkpoint is the least effort, or you can merge Jimeng’s exported subtitles and import them back.
3Garbled subtitles are overlaid on the picture: subtitles inside Jimeng’s picture often go wrong, so do not let it do subtitles. Remove the subtitle copy from the prompt so only the single layer the project burns in remains, and the picture will be clean.
4Not enough credits: first run a short 2–3 scene task to make sure the prompts, export naming and import flow all go smoothly, then move on to longer tasks.
Try it, then share what you make
This workflow is still in beta and what it needs most is real use cases. Run one short creative long video or manuscript video through it, then post the finished video or a screenshot to GitHub Discussions “Show and tell” and we will collect it into the site’s showcase section (feel free to attach your storyboard prompts and Jimeng settings so others can reproduce it).
Want the details
The full description of External Asset mode (the checkpoint artifact matrix, the batch import interface, the normalization rules) is in section five of the manual mode guide in the project repository; the cards inside each checkpoint also inline the prompts you can copy and the fill-back instructions.
Read the manual mode guide →How to handle a 503
Try them in this order — in the vast majority of cases you will have a clip within three steps:
Jimeng workflow FAQ (including how to handle Agnes 503 / queue is full) →Other free video generation tools? Recommendations welcome
This project’s tool list is an extensible registry: adding a tool only means adding a set of prompt templates and fill-back instructions, and neither the workflow nor the compositing layer has to change. If you know a better free video generation channel (Kling, Hailuo, open-source local models…), open an Issue on GitHub and tell us: the tool name, whether it has native audio, its per-clip duration tiers, and whether it can export subtitles. We will evaluate adding support.
Suggest it on GitHub →Run one right now
External Asset mode is a desktop feature: download the app and choose “🌐 External assets” when creating a task. The online Demo and the web Studio use automatic mode.
Related Pages
Jimeng workflow FAQ (including how to handle Agnes 503 / queue is full)
What do you do when Agnes returns HTTP 503 or video queue is full? Should you wait? Can you take another route? Plus the questions people ask most while using the Jimeng workflow.
Manual Mode Guide — Auto vs Manual: Vibe Generating or Precise Output | Agnes Video
How to use Manual Mode in Agnes Video Generator: pause at checkpoints (storyboard, references, video clips, narration, subtitles, final) to review or refine each artifact before continuing, optionally pairing with stronger external AI or tools for a precise, predictable result. Auto Mode fits vibe video generating from a rough idea; Manual Mode fits clearly specified outputs with fine-tuned artifacts.
Agnes Video Generation Errors: HTTP 503 Queue Full, 401 Token & Retry Fixes
How to fix Agnes Video Generator errors: HTTP 503 video queue is full, 401 invalid token, failed or interrupted generations with checkpoint resume, plus system requirements and API Key basics.
Narration & Subtitles: Add Voice-over and Captions
The digital human mode ships with built-in TTS and lip-sync; for regular text-to-video, add captions in an editor.
Sources
This page's description of Jimeng's capabilities, the external-asset checkpoints, and the media normalisation rules is based on the following sources.
- 即梦 AI (Dreamina)：AI 图片与视频创作平台ByteDance · 即梦官网 (2025)
Jimeng (Dreamina) is the tool that generates picture and native audio in this workflow; its site is the basis for the advantages/trade-offs section of this page.
- agnes-video-generator — 开源 AI 视频生成项目lcy362 · GitHub (2026)
The external-asset implementation (checkpoint prompts, batch import, keep-audio concatenation) follows the project source: core/providers.py, web/routes/video_routes.py and utils/video.py.
- FFmpeg Documentation — 音视频处理官方文档FFmpeg 项目 · ffmpeg.org (2025)
Both normalisation on import (scale-and-pad, uniform frame rate, H.264/AAC) and subtitle burn-in concatenation are done by ffmpeg; parameter semantics follow the official documentation.
Why you can trust this
Written by a practicing developer and cross-checked against authoritative primary sources such as arXiv papers, vendor technical reports, and the Stanford AI Index.
SandGrid@lcy362
Author of Agnes Video Generator · Full-stack Developer
Independent developer and author of Agnes Video Generator, an open-source (MIT) AI video generation tool built on Agnes AI's free video models. Has helped 1,000+ creators produce AI videos at zero cost. Focused on making video generation models accessible and production-ready; content is based on hands-on practice and primary research.
View on GitHubLast updated：2026-08-21
Ready to Start Creating?
"Making world-class AI belong to everyone." — Bruce Yang. It's completely free, no credit card, and you won't need a high-end GPU. Your first AI video starts at zero cost. Want to use Agnes AI's free video models? This is the easiest way in.
Clone the GitHub repo and launch in 2 minutes