Executive Summary
Veo 3: Quality tops Runway and Kling, but costs are higher. Next steps: Start Veo3 adoption with Scaling pods immediately, train ACDs to generate motion videos, and experiment with Google Flow to further improve promo quality. AutoAI: Three major fixes shipped - prompt suffixes now standardise look across movie styles; the “Hooks” experiment boosts relevance with multi-character frames (seductive, dramatic, even humiliation beats); 4o-images MVP rolls out by Friday next week. Other updates: Chinese agency work has gotten stalled - will pick up on priority this week; ACD hiring is almost complete; CPI ops revamp underway to cut TATs. Veo3 Update
2-Minute Promo Experiment
Learnings from His Heavenly Mate (HHM)
Veo3 is best-in-class for video generation It should be even better with Flow (Full veo3 access + Imagen 4) Prompt adherence in Veo3 is very high - we’re getting the videos we want within 1-2 attempts Consistency needs to improve - we’ll test if images generated via Imagen 3/4 (Google’s models) work better with Veo3 Workflow used for HHM
Image generation using GPT-4o Uploaded selected stills as reference in Veo3 Used the following structured prompt format Veo3 Base Prompt Structure
[Perspective], [Shot Style + Details], [Subject Details + Environment], [Scene Details], [Lighting], [Film Style]
Perspective: Camera point of view (e.g., over-the-shoulder, wide shot) Shot Style + Details: Framing and motion (e.g., medium-wide, tracking forward) Subject + Environment: What’s in focus and where (e.g., a girl standing in a ruined temple) Scene Details: What’s happening in the scene (e.g., glowing runes appear around her) Lighting: Mood and lighting style (e.g., soft golden sunlight through broken windows) Film Style: Overall aesthetic (e.g., cinematic realism, soft grain, shallow depth of field) Example - Over-the-shoulder, medium-wide, tracking forward, a girl stands inside a ruined temple, glowing runes appear around her, soft golden sunlight through broken windows, cinematic realism with soft grain and shallow depth of field.
Veo3 Comparison against other video tools
Key takeaways:
Veo3 is outstanding on output quality and ease of use. With Flow access combining Imagen 4, Gemini and Veo3, it’s a one-stop solution to create videos. The introductory price where we get AI Ultra accounts at $249 / month with 100+ minutes of generation capability is very attractive. Kling is not enterprise-friendly, but delivers high quality promos. Hailuo is the most value-for-money tool ($95 / month, unlimited generations) Veo3 Roll-out
Options to access Veo3
Pollo API (Available today) More expensive than using Veo3 directly Leonardo API (ETA:~2 weeks) Veo3 using Flow via US email IDs + VPNs (Available today) Speed is slower as VPN throttles output after sustained usage Veo3 using Google’s APIs (Vertex AI) (ETA: Unknown) Best price + low latency; but not available in India yet Use cases where we can now integrate video
Timelines to execute Veo3 integrations
This week: Scaling teams to start promo generation with Veo3 using Pollo Next 2 weeks: Prepare training material & upskill ACDs on video generation Target: 1 / 4 ACDs in every pod to be upskilled for video generation by end of month By then, we’ll either have Google Flow unlocked in India or the APIs AutoAI Progress
Framework for AutoAI Issues
Sluglines & Character LoRAs [Fix for consistency]
Slugline implementation is 100% for all new scripts in CPI Fantasy & D/R Character LoRAs + tool to generate training datasets is live on Flux Dev model Moving all teams to Flux Dev for better consistency This also solves a low-hanging Character Canvas improvement piece 4o MVP will also solve consistency issues for us + Veo3 is expected to not face consistency issues Hooks experiment [Fix for image relevance]
Problem: Hook images fail to capture attention (first 3-4 frames)
AutoAI focuses on matching the literal context of the script, which leads to predictable images in the first 3–4 frames. At the hook level, we need scroll-stopper visuals - images that immediately catch attention and create curiosity. Solution: AutoAI Hook Experiment
To solve this, we changed our approach.
We used only the first 2–3 sentences of the story as input. Instead of asking for context-matching images, we asked AutoAI to extract the theme or emotional pulse of the hook. Based on that, we prompted it to generate exaggerated, unexpected images with heightened drama - still rooted in context. We’ve been able to reign in the NSFW content that was getting created by the prompts. This shift helped us create stronger, more attention-grabbing hooks without losing narrative relevance.
Next Steps:
Test Hooks v1 results at scale by end of the week. Add more hooks (sub-genre specific) Apply the rules/ instructions on first 3-4 images to the entire promo in v2. Prompt suffixes [Fix for image enhancements]
Problem:
Many stories operate in completely different worlds
a period drama set in 19th century London, a gritty modern crime thriller, a fantasy series in a mythical space, a werewolf romance set in a high school Each of these needs its own distinct visual style. But, AutoAI doesn’t understand story themes or time periods unless specified in the promo. Also, these styles tend to change within a 30-minute long promo.
Solution: Prompt Suffixes
We created Prompt Suffixes - short visual-style descriptors (under 1000 characters) that guide AutoAI to generate images that match the aesthetic, lighting, and tone of a show.
How We Built It:
We researched the visual language of critically acclaimed films and shows. Then, we translated those styles into prompt suffixes, genre by genre.
Examples from the repository:
: The Godfather, The Revenant, Moonlight, Blue Valentine + 20 more styles : In the Mood for Love, Portrait of a Lady on Fire, La La Land + 20 more styles : Teen Wolf, Werewolf by Night, An American Werewolf in London + 20 more styles : Se7en, Zodiac, Prisoners, Nightcrawler + 20 more styles Naruto, One Piece, Rick & Morty + 20 more styles [Researched by Parth + Ankit] Experiments:
Suffixes have already been created for:
ChatGPT 4o MVP on AutoAI [Fix for consistency + relevance]
ChatGPT 4o is now available on Leonardo Need to improvise/ change the current workflow which is designed for Leonardo AI. Other updates
Chinese agency work
We have two Chinese agency contacts from the Licensing team - progress has been slow last week. Picking up on priority this week. Training for ACDs (New & Existing)
Session 1: Parth (CPI Fantasy) The first 3 seconds decide performance
Visuals must act as scroll-stoppers. If you don’t grab attention early, users drop off — even before audio registers. Avoid literal visuals – aim for interpretive elevation
Don’t “translate” the voiceover word-for-word. Instead, add subtext, mood, and stakes visually. This contrast creates intrigue. Promos fail when they feel like image slideshows
Think in scenes with continuity and momentum, not standalone shots. Each image should build tension or emotional pull. Narrative + tension > clickbait
Use mystery, surprise, or dilemma early on to pull the viewer in. If the viewer crosses 20 seconds, they're invested in the story arc — this is the real retention trigger. Session 2: Faiz (Scaling Fantasy) Complete motion promo workflow
Fez walked us through the full process — from script to audio timeline, to image and motion generation, and final edit. Case study: MVS Weapons promo
Explained how combining strong still images with motion elements helped improve CPI performance. Hook and cliffhanger approach
Shared how to design bold, attention-grabbing hooks and end with cliffhangers that build curiosity and improve watch-through. Gen AI Tools Stack
Gave a clear view of how tools like MidJourney, Kling, Pika, ElevenLabs, and Higgsfield are used at each stage and how they all can be utilized to level-up the quality of promos. Creative habit building
Reinforced the importance of spending 2–3 hours daily exploring new GenAI tools — not for immediate output, but to sharpen creative instincts. Session 3: Greg (TBC) - next Wednesday
Hiring update
Nearly all gaps are closed. Last leg of hiring for the new WW Scaling + WW CPI pods is underway, plus some backfills in in-app team.
Ops update
We have faced capacity issues dealing with the volume of requests from CPI D/R and CPI WW. We are now actively evaluating every stage of the process to bring down the TAT + revamping the trackers to provide more visibility to stakeholders.
4th June, 2025
