Consistency by reference
Bind one reference image to a character or a scene and every later shot reproduces the same face, the same background. Consistency comes from the image, not from re-rolling the prompt.
Capabilities
From one line to a finished cut one pipeline, end to end
Script, shot breakdown, reference images, voice-over, batch generation — all of it happens in a single workspace, with no shuttling between five separate tools.
Core mechanisms
Four things that separate it from a wrapper
Not a bigger model — engineering that makes consistency, concurrency, backend choice and cost actually hold up in production.
Task-pool scheduling
Bulk submissions enter a persistent queue, and the scheduler feeds them out in batches under a max_inflight=8 concurrency cap. Failed jobs are refunded automatically.
Three levels of backend choice
Global default, then per project, then per shot. One batch can partly run on your own GPU and partly in the cloud.
Every credit is on the ledger
Every generation is charged at 1 credit per second of video and 0.5 credit per image, then written to the ledger. Failures are refunded, line by line, on demand.
Persona matrix
Fifteen writers, not fifteen templates
Every persona brings its own subject library, fact-checking habits and storyboard instincts into the conversation — not a renamed general model.
Short Drama History & Heritage
A short-drama scriptwriter for history and heritage, turning historical anecdote and legendary figures into shot-ready scripts — rigorous like a scholar, charged like a storyteller.
Short Drama People’s Heroes
A screenwriter devoted to hero narratives, adapting and originating stories of martyrs, role models and exemplars across every field — finding the grand sentiment in a small, human frame.
Short Drama Xianxia & Myth
A senior xianxia and myth writer, skilled at Eastern fantasy worldbuilding and fated, tormenting romance — every script is ready to be shot the moment it lands.
Short Drama Urban Romance
A working writer of Shanghai-set urban romance. Offices, subway cars and corner stores become the sites of falling — every page reads like a shot list.
Short Drama Mystery & Thriller
A cold-blooded writer of mystery shorts. Twists woven into airtight logic — a hook in every episode, a reversal by the third.
Advertising Drinks & Snacks
A direct-response director for drinks and snacks, filming every bite as a 30-second film of desire you cannot refuse to buy.
Advertising Product Showcase
A narrative ad director for product showcases, turning every item into a short film you want to buy immediately.
Advertising Beauty & Fashion
A director built for beauty and fashion brands, fluent in the language of "the product is the protagonist" — every frame a place where desire takes root.
Advertising Tobacco, Spirits & Tea
A visual storyteller for premium tobacco, spirits and tea, using restrained light and framing to wake the aged character and Eastern grace of the product.
Short Drama Mythic 3D Showrunner
A showrunner for large-format 3D animation in myth, versed in Eastern fantasy worlds like Journey to the West and Investiture of the Gods — cinematic motion language and tight narrative structure for series- and theatrical-scale scripts.
Short Drama Tao Te Ching
A fixed-IP series built on the Keeper of the Archives and Wen Zi, producing 16:9 chapter-by-chapter drama shorts under 150 seconds — ancient micro-scenes paired with modern narration, turning abstract classic into concrete story.
Short Drama Old Dream Gramophone
A professional AI short-drama director, turning a one-line idea into a complete shot plan ready for AI video generation. Every rule is derived from 349 shots of five breakout videos and the field data of five top accounts.
Advertising Maternal & Baby Retail
A short-video strategist for the maternal and baby segment, shooting formula, nappies, baby food and toys as conversion clips any parent would save.
Advertising 10–15s Douyin Selling
A script strategist for 10–15 second Douyin selling videos, closing the full loop of hook, selling point and call to action in seconds.
Advertising Story-Driven Ads
A strategist for narrative-led conversion, using a complete story to carry the product pitch so the audience buys after they have been swept into the plot.
End-to-end flow
Eight steps, brief to download
This is the flow the product actually runs, not a concept diagram. You can take over at any step — the AI never locks you inside a black box.
-
Open Donjian Xujing and brief it
Pick a persona in AI Chat (e.g. the commercial director) and describe the film in one line — genre, length, target platform.
-
AI extracts characters and scenes
The AI drafts 5 characters and 3 scenes; once you confirm them it generates a reference image for each.
-
Choose the reference images
Each character and scene comes with 3 candidates. Pick one and apply it to the project — that image holds the subject consistent across every shot it appears in.
-
The AI breaks down shots and writes H3 prompts
The script is split into shots, each with an auto-generated 6-part H3 prompt (character / action / cinematography / composition / colour / sound). Every shot stays editable — prompt, duration, reference image.
-
Submit shots in bulk
Select shots in the project view and submit them in bulk. Everything enters the task pool, and the scheduler feeds the generation backends in batches under a max_inflight=8 concurrency cap.
-
Track the task pool
Each shot moves through pending → submitted → generating → done. Cloud video backends average 2–3 minutes a shot; local backends run at whatever your GPU manages.
-
Voice-over
Batch-select voices and styles in the voice manager; voice_status runs the same pending → generating → done cycle.
-
Preview and download
Preview a single shot inline, or select several and download them as a zip. The current build does not auto-assemble — the final cut is assembled by hand in Jianying or OBS.
Glossary
Twelve terms worth memorising
The definitions from the product manual, carried over word for word — shared vocabulary is what stops alignment going in circles.
- Persona Persona
- A persona — persona brief, greeting, quick commands and avatar — used to role-play the assistant in AI Chat.
- Concept Concept
- Project-level style references (4–9 images) that set the visual tone for every shot in the project.
- Character Character
- A recurring character. Once a reference image is bound, every shot they appear in reproduces the same face.
- Scene Scene
- A recurring location. Once a reference image is bound, every shot set there reproduces the same background.
- Shot Shot
- The smallest unit of output. Each shot carries one reference image, an H3 prompt, a duration and dialogue subtitles.
- H3 prompt H3 prompt
- MiniMax H3’s six-part prompt structure (character / action / cinematography / composition / colour / sound), generated automatically.
- T2V prompt T2V prompt
- A three-part text-to-video prompt, used when there is no reference image to condition on.
- Voice Voice
- An audio track for a shot, specified by a reference voice, the text to read, and style direction.
- TaskPool Task pool
- A persistent submission queue. The scheduler feeds jobs to local ComfyUI or cloud backends in batches under a concurrency cap.
- Dispatcher Dispatcher
- The task pool’s executor: it routes each job by task_kind to the right generation logic — shot, character, voice, and so on.
- backend_kind backend_kind
- Which multimodal backend to use: comfyui_local, hailuo, kling, seedance or cloud_image01.
- Credits Credits
- The single unit generation is billed in: 1 credit per second of video, 0.5 credit per image. Credits sit in one shared pool you can split freely between image and video; top-ups are ¥1 per credit.
Multimodal backends
Five backends, switch at will
The same batch of shots can move between your own GPU and four cloud models — chosen per project, and per shot if you need to.
backend_kind Use Speed Cost Quality
Local ComfyUI
comfyui_local Six local ComfyUI workflows Medium (GPU-dependent) Free (GPU electricity only) High MiniMax H3
hailuo Cloud MiniMax H3 video Medium (30–90s per clip) 1 credit/second (by final-cut duration) Very high Kling
kling Cloud Kling Slow (1–3 min) 1 credit/second (by final-cut duration) High Seedance
seedance Cloud Seedance Medium 1 credit/second (by final-cut duration) High image-01
cloud_image01 Cloud image-01 text-to-image Fast (5–10s) 0.5 credits per image High Recommended combinations
- Production short drama hailuo video + cloud_image01 references (cloud stability)
- Testing / debugging comfyui_local throughout (spends no credits; depends on your GPU)
- Bespoke enterprise comfyui_local video + custom workflows
- Bulk, lower fidelity kling / seedance (cheaper, more variance in quality)
Model matrix
What it runs on
The models in production, their roles, their specs and measured speed.
Model Role Spec Speed Quality
Qwen2.5 LLM Brain 72B / 14B <2s 3/5
GLM-4 LLM Brain 9B <2s 3/5
Seedream Cloud Image 4B 5–10s/img 3/5
ComfyUI Local Local Image SDXL / SD3 8–15s/img 4/5
MiniMax H3 Cloud Video I2V / T2V 30–90s/clip 5/5
Seedance Cloud Video Pro / Lite 45–120s/clip 4/5
Kling Cloud Video 1.5 / 1.0 60–180s/clip 4/5
Speech-2.8-HD TTS Voice Multi-lang <3s/sentence 4/5
Music-1 BGM Gen Mood-driven <10s 3/5
FAQ
The real questions that come up
Three entries lifted verbatim from the manual’s FAQ — every answer is something you can verify in the product yourself.
What if the AI shot breakdown is not what I wanted?
Edit the H3 prompt shot by shot in the project view, or go back to Donjian Xujing and ask the AI to rewrite a given passage. Neither route disturbs characters or reference images you have already confirmed.
A shot keeps failing to generate — why?
Read the error field on that row in the task pool. Three causes account for most failures: sensitive words in the prompt, a reference image below 512×512, and local ComfyUI running out of VRAM (confirm in the logs).
How do I retry a batch of failed shots?
Filter the shot list by status=failed in the project view, select the shots, and use “Retry Selected”.
Want to see it running?
Registration is invite-gated. Tell us the kind of film you are making and we will open a workspace configured for that scenario.