Make Consistent AI Animation with Codex: A 5-Step Workflow So Your Characters Stop Drifting (with the Full Project File Structure)
Most people making AI video start by pasting the story into the model and hoping it generates a complete short film in one go. The result is almost always the same: characters look different from shot to shot, shots don't cut together, and the art style drifts further and further off.
This tutorial doesn't take that approach. We split "story → screenplay & storyboard → prompts → reference images → video" into a file-backed pipeline: Codex decomposes the task and produces the storyboard plus two sets of prompts, and Jimeng Seedance image-to-video makes it move. Follow the 5 steps and a ~30-second, 8-shot 2D hand-drawn anime short keeps its characters consistent.
Bottom line first: a 5-step workflow — ① Put the story in
input/article.md→ ② Codex produces the screenplay & storyboard → ③ Split out two prompt types, "image prompts" and "video prompts" → ④ Generate the reference images one by one → ⑤ Generate each shot's video using "reference image + its matching video prompt." The core idea in one sentence: the image prompt decides "what the frame looks like," the video prompt decides "how the frame moves" — feed them separately and your characters won't drift.
Contents:
- Prepare Your Tools
- Why "One-Shot Generation" Fails
- Get to Know the Project Structure
- Step 1: Put the Story into input/article.md
- Step 2: Generate the Screenplay & Storyboard First
- Step 3: Generate Image Prompts and Video Prompts Separately
- Step 4: Generate the 8 Reference Images First
- Step 5: Generate Video with Reference Image + Matching Prompt
- Final Checks & Export
- Pitfalls to Watch Out For
- FAQ
Prepare Your Tools
- Codex (OpenAI): reads the story, writes the screenplay & storyboard, and produces both sets of prompts. You need a ChatGPT account with Codex quota (how to pick a quota: see Codex quota comparison — top-up info at the bottom of this article).
- Jimeng (Seedance): image-to-video, making each reference frame move.
- The copy-ready prompts provided in each step of this article.
Why "One-Shot Generation" Fails
Most people making AI video start by pasting the story into the model and hoping it generates a full short film directly. This approach easily runs into three problems: inconsistent characters, discontinuous shots, and continuously drifting style.
The problem isn't the model — it's the division of labor. When a single model has to manage "story, storyboard, frame, motion" all at once, it inevitably drops the ball somewhere. So this tutorial takes a different route: build a complete project file structure first, then generate story, storyboard, reference images, and video separately — each step does only one thing.
Get to Know the Project Structure
This is the foundation of the whole method. The directory tree:
project/
│
├── input/
│ └── article.md
│
├── script/
│ ├── story-analysis.md
│ ├── screenplay.md
│ └── storyboard.md
│
├── prompts/
│ ├── image-prompts.md
│ └── video-prompts.md
│
├── images/
│ ├── scene01.png
│ ├── scene02.png
│ ├── scene03.png
│ ├── scene04.png
│ ├── scene05.png
│ ├── scene06.png
│ ├── scene07.png
│ └── scene08.png
│
├── videos/
│
└── output/
└── final.mp4
The point of this structure is to split "story, storyboard, frame, video" into distinct stages. The AI no longer does all the work at once — each step handles one thing, and every deliverable is a file you can inspect, edit, and reuse.
Step 1: Put the Story into input/article.md
Place the original story in input/article.md. This file is the input of the whole project. It isn't fed directly to video generation — it's handed to Codex for decomposition.
Then give Codex this key prompt:
Read ./input/article.md.
Adapt it into an original 2D hand-drawn anime short film of about 30 seconds and 8 shots.
Don't generate any images or videos yet. Only complete:
story analysis, character design, scene design, emotion curve, and storyboard planning.
Step 2: Generate the Screenplay & Storyboard First
This step isn't about writing a pretty screenplay — it's about giving the later images and videos a single unified basis. Codex will generate three files under script/:
script/
├── story-analysis.md
├── screenplay.md
└── storyboard.mdstory-analysis.md: the story, characters, emotion, and visual themesscreenplay.md: what happens in each shot within the 30 secondsstoryboard.md: shot size, composition, camera movement, and timing
Step 3: Generate Image Prompts and Video Prompts Separately
This is the conceptual core of the whole tutorial: image prompts and video prompts must not be mixed together.
image-prompts.md owns the frame itself:
- Character appearance
- Clothing
- Scene
- Composition
- Style
- Lighting
- Negative constraints
video-prompts.md only owns the parts that move:
- Character motion
- Wind chime swaying
- Hair and hem lines
- The cat's movements
- Camera movement
- Stop character drift and photo-realism
One sentence to tell them apart: the image prompt decides "what the frame is," the video prompt decides "how the frame moves."
Step 4: Generate the 8 Reference Images First
Generate images/scene01.png through scene08.png in order. Don't let Seedance improvise from the story directly — first lock each shot down as a reference image, then let the video model's only job be making that image move.
Step 5: Generate Video with Reference Image + Matching Prompt
This is the core of the whole method: each reference image is paired only with its own shot's video prompt.
scene01.png + Scene 01 video prompt
scene02.png + Scene 02 video prompt
scene03.png + Scene 03 video prompt
...
scene08.png + Scene 08 video prompt
In Jimeng Seedance, choose image-to-video, then drop in the reference image and its matching video prompt. You can generate the 8 shots one at a time and stitch them together, or let Codex string the whole flow into a single finished video.
The principle boils down to one sentence:
The Source Image decides the characters and art style; the Video Prompt only handles motion and camera. Feed them separately and you'll noticeably reduce character drift, outfit changes, and style loss.
Final Checks & Export
Put the generated clips into videos/, and move the confirmed final version to output/final.mp4.
Before exporting, check four things:
- Whether the characters are consistent
- Whether the shots cut together continuously
- Whether it stays 2D anime
- Whether anything has gone photo-real or 3D
The point of this workflow isn't to have AI generate an entire film at once — it's to break a complex task into a set of files that can be checked, edited, and reused.
Pitfalls to Watch Out For
- Codex quota: image and task generation burns through quota faster than chat — when it fails, check your remaining quota first and don't retry repeatedly.
- Reference images set the ceiling: if the 8 reference images are stylistically consistent, the final result stays consistent; once the reference frames drift, no amount of downstream tweaking can save it.
- Don't describe appearance in video prompts: only describe motion and camera. Repeating character appearance actually makes the model change the art style.
- Do one thing at a time: the shorter and more clearly scoped the instructions you give Codex, the more stable the output.
FAQ
How is this different from the previous 'Codex image generation + Seedance image-to-video' tutorial?
Do I have to use Codex? Can another AI replace it?
I followed the steps but my characters still drift / frames look inconsistent. What now?
My Codex quota isn't enough and it keeps saying quota insufficient. What do I do?
Closing
If you want to make short films with AI, don't start with "generate a video for me." Start by creating a project folder: the story goes into input, the storyboard into script, the prompts into prompts, the frames into images, the video into videos, and only the confirmed final version into output.
Once the file structure is clear, AI video stops being a random gacha pull and becomes a repeatable production pipeline.
Need Codex quota? Check the Codex top-up page to see supported plans, payment methods, and delivery.
About the Author
UPGPTs Team creates step-by-step, screenshot-heavy tutorials in Chinese for AI tools like ChatGPT, Claude, Grok, and Codex. Find more in the UPGPTs Tutorial Hub, check the FAQ, or reach us via the Contact page.
Related Reading
- Codex Image Generation + Seedance Image-to-Video: Celestial Palace Cinematic Tutorial — the entry-level version that turns a single image into a single video
- ChatGPT Codex App Guide — the full Codex onboarding flow for new users
- Add Vision to Codex — give Codex eyes
- Codex Quota Comparison: Plus / Business / Pro at a Glance — a full quota comparison across the four plans