Explore

Gen AI Video Production Workflow

Step 1: Character Canvas Creation

Once a new show is onboarded, the first step is to create the Character Canvas. This involves defining and generating unique visual prompts for every character appearing in the episodes.

Time Estimation: For every 5 hours of audio, character canvas creation (assuming there are 20 characters) takes 1 full day of work (for one ACD). Manage your team’s time accordingly.

Team Structure: One Creative Director (CD) handles one show and manages a team of 4 Assistant Creative Directors (ACDs).

Character Density: In early episodes (first 10–20 hours), many new characters are introduced. As the show progresses, fewer new characters appear, so the workload reduces.

⁠

Process

Upload the SRT file (subtitle transcript) of an episode to a

Google Colab research link⁠

A JSON file is generated, listing all characters appearing in the episode, along with auto-generated physical attributes.

Important: The JSON is not fully accurate. It only captures attributes based on that episode’s context.

NOTE:

Always listen to the audio and manually verify physical traits and personality cues.

If a character is described as thin in Ep 1 but later revealed to be plump in Ep 5, all previous prompts must be updated.

⁠

Character Prompt Format (Canvas)

A strong character canvas prompt includes:

Age: Use specific numbers or ranges like “mid-20s” or “early teens.” Never use vague terms like “adult.”

Facial Features (5–7): Include facial structure, eyes, eyebrows, jawline, nose, beard/stubble, hair type/color/length, etc.

Body Structure (2): Include height, body build, posture, etc.

This level of detail will ensure unique and consistent character renders across episodes.

Quinn - An 18 year old male, (oval-shaped face with pointed chin, pimples on his face, big brown eyes, black short hair, sharp nose), short height, lean muscular build

Jake - A male in late teens, (square shaped face, blue eyes, light stubble, brown short curly hair, round nose, wavy hair), thin built, tall athletic body

Emma - A 17-year-old female, (round face, medium-length straight blonde hair, big bright blue eyes, thin arched eyebrows, thin lips, pointed nose, clear fair skin), slender build with narrow shoulders

⁠

Want to dig deeper?

Here’s a list of character archetypes and physical traits you can use to create them:

⁠

Character Archetypes⁠

⁠

Review & Testing

ACDs generate preview images for all characters and present them to the CD for approval.

CD checks for uniqueness, accuracy, and if the character fits the story tone (e.g. protagonist = likable & vulnerable; antagonist = ).

Avoid subjective traits like “elegant,” “authoritative,” or “classy” in prompts, as these restrict emotional range in image generation for scenes.

⁠

Validation

Once approved, ACDs test each character across multiple expressions, costumes, and locations.

If the images remain consistent, the prompt is locked.

Finally, the approved Character Canvas is fed into Auto AI (CMS tool), and episodes (e.g. Ep 1–5) are queued with the locked canvas.

⁠

Step 2: AutoAI Video Image Generation via CMS

Once the Character Canvas is finalized and tested, it is formatted according to the AutoAI requirement and prepared for upload.

Upload Process:

Go to the Video Generation Tool on the CMS, where AutoAI is hosted.

Upload the completed Character Canvas Sheet along with the Show ID.

Batch Queue 5 episodes at a time for generation.

The episode is broken down into 4-second timestamps, giving a total of 15 images per minute.

For each row (timestamp), 1 prompt is generated and 4 image options are created.

⁠

How AutoAI Works (Creatively):

Transcript + Character Canvas + Seed Values for characters are fed into AutoAI.

AutoAI uses a backend system (Python + ChatGPT APIs) to understand the story in chunks of 7 rows (~28–30 seconds).

Based on context, it creates prompt variations for each scene dynamically.

Example: If a character like Emma is introduced:

Emma is standing on a street → Prompt: Emma standing on a street, office attire, character canvas prompt for Emma, morning.

Emma enters a cab → Prompt: Emma sitting inside a cab, character canvas prompt for Emma, same attire.

Emma walks into an office building → Prompt: Emma entering a corporate building, character canvas prompt for Emma, same attire.

Even though the Character Canvas only defines physical traits, AutoAI uses context to generate scene-specific prompts (location + costume + expression).

⁠

Prompt to Image Generation:

These dynamic prompts are then passed to Leonardo AI automatically.

Based on the selected model (AnimeXL, Realistic, etc.), Leonardo generates 4 image options per prompt.

Final renders (images + prompts) are pulled into a Google Sheet — 100 rows = 400 images.

⁠

Output for Review:

This output sheet is the final render sheet for the episode.

It is handed over to the Assistant Creative Directors (ACDs) for Quality Check (QC) by the CDs.

⁠

Step 3: Quality Check (QC) by ACDs

Once the character canvas is tested and images are generated via AutoAI, the Assistant Creative Director (ACD) receives a Google Sheet for a specific episode. This sheet contains 4 image options and 1 prompt for each row of the episode. The ACD listens to the full audio and reviews the entire sheet following strict QC guidelines.

CD’s job is to ensure that the following standards are maintained throughout the episode and there’s tonal consistency throughout the show apart from stylistic consistency and overall visual treatment (look & feel).

The QC checklist is divided into 5 major steps:

1. Hygiene

1a. Art style consistency across all frames

1b. Character must match the canvas prompt exactly

1c. Eliminate character distortions or anatomical errors

1d. Ensure character consistency across episodes

1e. Reuse existing approved character images wherever applicable

2. Pacing

2a. Maintain frame rate: For episodic shows,

Without LA: 15 IPM for 30 minutes

With LA: 15 IPM for 2 hours (or 8 IPM otherwise)

2b. No single image should last more than 10 seconds

3. Direction

3a. Include proper contextual or insert shots where needed

3b. Use multiple camera angles to avoid flat storytelling

3c. Avoid repetitive shots with same character pose

3d. Characters must have visible expressions, mouth open when talking

4. Scene Consistency

4a. Maintain costume consistency in a sequence

4b. Background/environment must match throughout the scene

4c. Ensure image aligns with audio narration for relevance

5. Operations

5a. Maintain the Excel sheet meticulously, including character canvas references

5b. Verify audio match and duration match in CMS (±5 seconds allowed)

5c. Ensure all video links are working

5d. Ensure the CMS episode audio matches the audio playing in the final video

Responsibility Order for QC:

First CD

Operations for Ops QC

Once QC is completed by the ACD, the approved image set is passed to the Creative Directors for a final review before moving to editing and stitching.

⁠

Want to print your doc?
This is not the way.

Try clicking the ··· in the right corner or using a keyboard shortcut (

CtrlP

) instead.