Explore

US GenAI Updates

30th April, 2025

Exec Summary (1-minute read)

AutoAI Quality Fixes in progress: Character LORAs live on Leonardo (need AutoAI hookup); background slugline testing starts Friday; scoping “flat, emotionless image” fix once we get more inputs from writers.

May-June pipeline: Prompt misreads, static shots, and model limits sap speed and consistency.

Hiring Pipeline: Hiring 5 CDs + 20 ACDs in progress. Two good CDs (Shantanu, Manthan) identified already. Plan to complete by end of next week.

Guidance Needed: Leadership to rank next bets—Kling 2, MVS Spanish, or GPT-4o API—to finalize May-June experimentation roadmap. Same ask on key shows for CPI & in-app teams.

I. Updates on current 3 priorities

Character consistency solution - Character LORAs

LORAs are unlocked on Leonardo this week

LORA vs Manual Images -

Open link

⁠

We believe this solves the character consistency issue in promos and in-app

Background consistency solution - Sluglines

Waiting for the tech team to confirm the solution is ready for testing

Have informed the script cracking GMs to start preparing to add slug lines to their promos

We’ll add a check on AI VO creation so that these slug lines don’t flow into the promo

Improving flat, emotionless images & making them sub-genre specific - TBD

II. Top AutoAI concerns we are hearing from the team [May-June pipeline]

Summary of issues across CPI Fantasy, CPI D/R, In-app, and WW

Theme

How it Shows Up in D/R, Fantasy, In-app & WW

Why it Hurts

Example

Character consistency

Hair, skin-tone, age, gender, costume or scars change mid-promo; character gets swapped for random faces E.g. Medieval garb -> modern shirt, wolf tattoos change design/placement, swords disappear

Breaks story immersion, forces manual re-gen

⁠

link

⁠

Background / scene continuity

Locations jump seasons or architecture; indoor ↔ outdoor mid-dialogue; supernatural props vanish or mutate

Viewers notice “teleporting” sets, kills believability

⁠

link

⁠

Prompt-to-image mismatch (tool doesn’t understand metaphors)

Images show the wrong action, wrong object, or literal misread (“dragon of rage” → actual dragon)

Adds redo cycles; script ≠ visuals

⁠

link

⁠

Static, expressionless portraits

Model defaults to head-on studio shots; scarce wide, action or dynamic angles

Feels dull; loses dramatic beats every 4 s

⁠

link

⁠

Shot-style & lighting inconsistency

Art style, palette or camera angle swings between frames; model ignores “low-angle, OTS” cues

Disjoint visual language; harder to edit smoothly

⁠

link

⁠

Facial / limb distortions in multi-character frames

Hands, eyes or multi-character frames generate glitches, especially in action scenes

Requires multiple retries, delays delivery

hands distorted -

link

⁠

Missing or partial image sets

Some frames fail entirely or are flagged by safety filters

Cuts choice, forces extra generations

⁠

link

⁠

Model / tool limitations

Character canvas ignored; strict content filters flag kids/violence; tiny face pool yields look-alikes; female characters added in the background or get gender swapped

Teams juggle seeds, prompts, manual workarounds

⁠

link

⁠

There are no rows in this table

⁠

Deep dives by team

CPI D/R

CD1 - Keval

Problems:

1. Character consistency issues: costumes and hairstyles change between shots.

2. Generated images often appear as expressionless portraits (like studio photography) rather than dynamic/cinematic or action-driven shots.

3. Lack of contextually accurate elements (e.g., props, liquids, facial expressions like speaking or shouting).

4. Nuanced details are frequently missing, reducing the storytelling impact.

Examples:

[AutoAI Sheet Link 1] – In this promo, the main character’s hairstyle and outfit change mid-sequence, and emotional intensity is lost due to passive expressions.

https://docs.google.com/spreadsheets/d/1y2Zj_8KmvSGR_NwP1SIShrOVLPpcuflarXKCqDo-z58/edit?gid=1154536646#gid=1154536646⁠

⁠

[AutoAI Sheet Link 2] – Several shots use photographic-style poses with no alignment to action or dialogue;

https://docs.google.com/spreadsheets/d/1Cjr5B2wWHuzeY2npseUfESwU-9pD5UrG4aavGHhowkA/edit?gid=240524385#gid=240524385⁠

⁠

Frequency:

These issues are very frequent and occur randomly across all the Auto-AI-generated assets. Almost every asset contains at least one or more of the mentioned problems.

The occurrence is unpredictable but consistently present across most promos.

CD2 - Aniket

Problem:

The characters are inconsistent. The Auto-AI tool does not pick the character canvas prompt 5/10 times and creates the characters on its own, which look very different from the canvas.Sharing for a few characters:

Examples:

Charles

Auto AI sheet image:

https://drive.google.com/file/d/1ipoyQD9RDTkPf0i_cTT3tMYOAS51Z7dI/view?usp=sharing⁠

⁠

Charles original look:

https://drive.google.com/file/d/1pUv1nvNjIf0NtcACHTgOIth6E4RJzEuv/view?usp=sharing⁠

⁠

Cortez

Auto AI sheet image:

https://drive.google.com/file/d/1fElzMD8z-W4BoUhbNh0nAh8gHjVY81-n/view?usp=sharing⁠

⁠

Cortez original look:

https://drive.google.com/file/d/1hDn4-gvdnTNixN5uPdclC2-wa4frFG8M/view?usp=sharing⁠

⁠

Problems:

Most of the images we get are character images from the front angle, which reduces dynamism. Also, in most images, the faces of the characters are expressionless, which makes the visuals less appealing.

As we focus more on showing rather than telling, we have to create images manually to give more dynamic visuals.

Sometimes the FF sheet doesn’t give images related to the story, it gives random images. (For example, Row 80 on the sheet shared.)

The prompt and image do not match 4/10 times.

There are images with distorted faces, eyes, hands, and fingers in the FF sheet, so we have to manually generate those.

I feel there’s a huge difference in the quality of manual generation and Auto AI. Manual generation is always better. However, it depends on what we want to focus on—quantity or quality.

Example:

Reference Sheet:

https://docs.google.com/spreadsheets/d/1I-O-38WQ7Ol6cu_3ieOkM7ujp7tHu_EiD5jKE5gU3s0/edit?gid=1484276044#gid=1484276044⁠

⁠

(This sheet has all the problems.)

CD3 - Shoyel

AutoAI Issue Summary Report

⁠

Problem: Inconsistent character generation (age, appearance, costume)

Examples:

Character descriptions from the canvas (e.g., Elena, 30-year-old with pale skin, thick body, chignon hairstyle) are frequently altered in AutoAI generations (e.g., slim woman, olive skin, wavy hair).

Elderly characters (e.g., Vivian, 55, wrinkled, botoxed) appear too young, often in the 30–35 range, missing key visual aging cues.

Costume changes mentioned in subtitles (e.g., black suit to hospital attire) are not reflected accurately — characters sometimes appear in unrelated outfits like lab coats or aprons.

Frequency: Very frequent.

Out of 10 promos, how often does this occur? 9–10

Within a promo, how many times does this occur? Almost every major character or scene

⁠

Problem: Location and context mismatch

Examples:

Scene described as "The meeting room erupted in applause" results in an empty, lifeless conference room.

"Logan is lying in bed and screaming" results in an image of Logan sitting upright, calmly dressed in a coat suit.

Frequency: High frequency, especially in emotion- or action-heavy scenes.

Out of 10 promos, how often does this occur? 7–8

Within a promo, how many times does this occur? 5–10 scenes per promo

⁠

Problem: Incomplete image sets (only one image generated per row)

Examples:

In 6 out of 10 episodes reviewed, over 300 of 388 rows had only a single image generated, instead of the usual 3–4 per row.

Frequency: Extremely frequent in larger datasets.

Out of 10 promos, how often does this occur? 6–7

Within a promo, how many times does this occur? Hundreds of rows in some cases

⁠

Problem: Repetitive framing and monotonous shot composition

Examples:

Especially in Drama genre: majority of generations are close-up portraits, with little to no variation in expressions or composition.

Lack of medium, wide, or dynamic shots. No clear visual storytelling progression.

Frequency: Consistently observed in entire genres.

Out of 10 promos, how often does this occur? 8–10

Within a promo, how many times does this occur? In most or all rows

⁠

Problem: Reference image mismatch

Examples:

Out of four reference images generated, often only one is accurate, with others diverging significantly in character appearance or environment.

Frequency: Common across promos with reference image needs.

Out of 10 promos, how often does this occur? 6–7

Within a promo, how many times does this occur? Affects almost every reference image set

CD4 - Shreya

Based on my basic limited experience with Auto Ai work, here's my observation:

Problems:

Inconsistency in Characters costume, lack of generating proper scar on faces, huge image disparity in how they are supposed to look based on the script & the auto ai generation. Character inconsistency (e.g., changing necklines, hair, sleeves, etc.)

Image quality and context often weak due to prompt overload

Continuity breaks — AI doesn’t track costume/location changes across subtitles

Repetitive frontal character shots with non-dramatic expressions

Requires human adjustments to fit narrative and make prompts concise without bloat — reduces automation value

Often generates unusable images for character shots (out-of-context, costume inconsistencies, continuity, expressionless, not in-action)

More suited for filler/background visuals than key narrative frames

For comparison check Auto AI promo 271 N:

271 N - Google Sheets⁠

and Manually made same story

NCPI1274 - Google Sheets⁠

. All the points made can be seen in the comparison of these two sheets.

Examples:

535 N⁠

(TAB, where Daemon is supposed to be shirtless while in the auto AI entirely he has come wearing a black shirt. All the Royal Stitches Auto Ai Sheets (the scar on the protagonist's face is missing)

Frequency: Out of 10 promos, all promos made by Auto AI have this, because it hasn't evolved yet to be used directly. Within a promo, it is a problem trying to merge Manual generations with Auto AI, because instead of generating consistent images, we have to generate manageable ones to blend with Auto AI. The targets becomes saving Auto AI, than being effective, inevitably becoming counter-productive.

CPI Fantasy

Column 1

Column 2

Column 3

Problem

Examples

Frequency (approx %)

Scene Setting Consistency Errors

→ Backgrounds change illogically, away from the world that was created → Scene transitions without logic due to something being take literally eg: "shivered" mentioned so showed a cat → During a conversation in the same setting, background looks different mid-scene

27.51% (Scene Setting Consistency)

Character Consistency Errors

→ Characters change post 3rd min, randomly generate new characters not in canvas → Character appearance inconsistent across slides even when name in mentioned in canvas → Female characters generated even when male specified, random females added out of context

19.30% (Character Consistency)

Prompt Misinterpretation

→ "Male martial artist" but shows female character → Bobby shown instead of Laila → Bobby shown Asian because "martial art" mentioned → Literal dragons shown even when mentioned as emotional metaphor →Random reuse of characters

11.85% (Out of Context)

Gender Biases / Costume Errors

→ Female visuals shown even when male prompted → Male characters shown wearing skirts

2.89% (Costume Consistency)

Weak Action Sequences

→ Static headshots instead of action → Distorted characters during action → Background continuity breaks → Mid scenes stay flat, no play or drama → Dark assets dont transalte well on Meta, so manually changed

1.06% (Style Consistency)

Character Canvas Issues

→ Descriptions like "attractive, confident" instead of clear physical traits → Spelling mistakes cause character mismatch → Forbidden words like "teenager" not removed early (Leonardo doesnt create images for words like teenage / small child - we should remove them using AI instead of CDs having to do it manually)

There are no rows in this table

⁠

Detailed Sheet:

Open link

⁠

In-app & WW

General Auto AI Concerns

Inconsistent models - Most of the versions we use except flow state, don't understand shots and lighting prompt, like low angle, OTS etc.

Narrator - Auto AI Character- The auto Ai sheets usually does not pickup the narrator, and most of the promos have a narrator narrating the story, so usually every row has a different character in auto ai as narrator.

Distortion- Overall with leonardo, shots of hands, fingers, action scenes or dual character in a frame brings distortion, or inaccuracy. Hence such frames require multiple attempts, moreover to give a diverse visual experience such shots are needed in promos..

Content gets flagged - Leonardo Ai has very strict guidelines for Content related to children indulged in violence and it keeps flagging the Content. Prompts with children are very sensitive to Ai. Most of them are reported as Inappropriate.

Limited options - Very small pallet of faces for certain prompts, like its almost regular to get the same images for two different characters with two different prompts

Inconsistency in Werewolf Elements (e.g., Tattoos, Statues).

Visual elements tied to the werewolf theme—such as werewolf tattoos or werewolf statues—appear differently in nearly every frame.These props or design motifs lack visual continuity in terms of size, shape, style, and placement.

Prompt - Suffix issue

Every Prompt has a suffix given below:

The characters should wear period-appropriate clothing (medieval, Victorian, or gothic), with subtle wolf-inspired jewelry or symbols indicating their werewolf heritage

This prompt further complicates and lengthens the process by adding a random character to an establishing or a background shot which then have to be corrected manually.

ACD : Neha Gupta

Worksheet:

The Dark Side ⁠

⁠

Inconsistence character of tamia skin color get changed in whole AUTO AI image few place out shows fair, few place dark in row 15 and many other rows.

There is a random character appearance in row 3 and many other rows.

Row 15 displays a distorted image that shows unclear fingers.

The AUTO AI-created tamia cloning reveals that she is holding herself in row 21.

Half of the sheet only has one image option instead of four in the AUTO AI column after row 47, which limits our selection area and forces us to create a multiple image Generation.

There is also an inconsistency in location with wrong Character representation in row 56.

ACD : Manmohita

Worksheet-

Dashing to my Alpha Twins⁠

⁠

Missing Backgrounds and Locations:

Several entries lack detailed backgrounds or specific locations, which are essential for contextualizing the scenes and enhancing the narrative.

Inconsistent Werewolf Characteristics:

Some images do not consistently portray werewolf traits such as fur, claws, or fangs, leading to ambiguity in character depiction.

Lack of Character Descriptions:

Certain entries are missing comprehensive character descriptions, making it challenging to visualize and differentiate characters effectively.

Ambiguous or Vague Details:

Some descriptions are too generic or lack specificity, which can hinder the creative process and result in less impactful imagery.

Worksheet:

Selling Myself To The Alpha⁠

⁠

Background and Location Mismatches

Around 9 images had to be completely redone due to unsuitable indoor settings (e.g., modern homes). These were replaced with more context-appropriate locations like cafés or outdoor scenes.

Missing Supernatural Elements

Some images lacked werewolf-themed visuals. Werewolf statues were added to enhance the supernatural atmosphere.

Character Inconsistencies

Character faces did not match the canvas in some cases, requiring full regeneration for consistency.

ACD : Kumar Aditya

Worksheet –

Rejected My Alpha Mate⁠

Rows 5 & 15: The characters and settings appear to be randomly generated, lacking alignment with the intended context. The locations and character depictions need to be more deliberate and consistent with the theme.

Subtitle Mismatch: In one instance (row marked with a sunglasses emoji), the subtitle does not correspond with the image, resulting in a disconnect between text and visual.

Worksheet:

The Guardian Wolf and Her Alpha Mate⁠

⁠

1. Most of the generated images feature costumes that are overly casual and not suitable for a werewolf-themed show. A more fitting wardrobe is recommended to maintain genre coherence.

2. Rows 15, 18, 21, 22: The character’s facial features in the generated images differ significantly from those shown in the reference canvas. As a result, all visuals were manually generated to preserve visual consistency.

ACD : Vaibhav Singh Tomar

Worksheet:

Feral Confessions⁠

⁠

Lead character inconsistency due to narrator conflict (constantly occurs). (row 3)

Story goes to flashback, Auto AI keeps the age and the costume of the character same, for instance in this promo the character is supposed to be 4 year old in the sequence. (row 26)

Worksheet:

Bullied By My Alpha Prince⁠

⁠

Inconsistent character ( both bully and lead) (row 4 and row 12)

Auto AI always unable to understand and generate actions that are significant to the sequence (row 18)

ACD : Hussain

Worksheet:

My Mate and Brother's Betrayal⁠

⁠

Complete Rework Needed:The original asset’s look and feel did not align with the intended visual style or theme.

Atmospheric Elements Missing:

Elements like mist and a red moon had to be added to create the appropriate mood.