Skip to content
Character inconsistency (40% of ACDs feedback) - resolving via LoRAs adoptionRepetitive or irrelevant images (26%), P0 fixDisfigured or blank frames (24%), P1 fix
Playground LimitationsVideo Quality Drops: Playground gives 720p by default; only Google Flow supports 1080p upscale.Fix: Flow API creating 1080p videos to be integrated into Playground.16:9 Aspect Ratio Restriction: Most of our promos are 1:1, while the default is 16:9 in Veo3. Scaling teams prefer using Kling because it supports 1:1 videos. Fix: Resizing API (Runway / Firefly) to be integrated within Playground.No history of past generations: Playground sometimes fails because of volume constraints - ACDs then have to start over and can’t access their previous generations. This slows down rework attempts for a video. Fix: UI changes to record generations historyKnowledge Gaps (Will be fixed by weekly training sessions)Context Gap: Videos often lose story or visual context if prompts or reference images aren’t well-structured.Fix: Releasing a prompt guide for Veo3 + sharing feedback with all contest entrants on their videosPrompt Sensitivity: Some prompts give weird or unexpected results, but this can be fixed with better prompt writing.Multi-Character Scenes: Hard to control who speaks or acts when more than one character is in the same shot.Tool LimitationsContent Flags: Action, gory, or sensual scenes often get blocked or fail silently due to policy filters.Some videos get auto-downgraded to Veo2 automaticallyAudio Issues: Dialogue sometimes drops - this is a known issue shared by Google in Google Flow updates. Too much reliance on sound and dialogues should be avoided until next update.
We ran a Veo 3 contest internally to increase familiarity with the tool.We’re using this contest to understand how ACDs are approaching Veo3 and offer customised feedback.
Veo3 Prompting (Live): Hands-on session to address current prompting challenges and reinforce visual storytelling basics.Scheduled for Fri, 20 Jan.Filmmaking Refresher (Live): A quick session on cinematography, shot types, and treatment to align on visual grammar. Planned for next Friday (will be discussing briefly during Veo 3 Prompting this week)Veo3 Prompting Guide (Async/ Offline Resource): A written and Loom-video-based guide to help teams prompt better.
Veo3 + AutoAI Integration: Needs another week to get to stable outputExperimented with multiple Veo 3 prompts and reference images - WIP but improving.Still images must be contextually strong for Veo 3 video generation to work effectively.Process: AutoAI image → used as reference in Veo 3 → generates final video.AutoAI + ChatGPT 4o Integration: MVP launch within next 2-3 daysNSFW image issue has been resolved.Ongoing issues: character and outfit consistency still lacking for long promos. Another challenge: high-quality but repeated images due to mid-sequence context lossLoRA Integration: 100% adoption by next FridayLoRAs are implemented, but adoption was slow.After syncs with CDs, execution has now picked up with visible improvements in the past few days.SlugLines Execution: 100% adoptedAll teams are actively using them for new scripts.ComfyUI workflows: TBDFeedback from external agencies suggest that using ComfyUI workflows to access Flux’s API directly is cost-effective and controls image consistency.We haven’t fully explored the potential of this workflow - doing more research to understand if it will help us speed up AutoAI.
Character Inconsistency (40%)Caused by canvas mismatches. LoRA was not previously integrated but now should help stabilise appearances.Costume InconsistencyCurrently only one outfit used per character. But, scripts are long and characters change costumes during different scenes. If costume inputs are integrated into sluglines (like scene labels), this can be solved.Style InconsistencyPrompt suffixes are the fix; need to ensure better adoption and team-wide communication.Image Relevance (26%)Repetitive Shots / Limited Framing:Distortions (24%)Disfigured/Blurred Images & Failed Two-Character Frames:
These are model-level issues. Can be fixed through manual replacement for now - exploring if new models (e.g. Flux Contekt, ChatGPT 4o) are better at handling these issues.Blank images:Image Enhancements (11%)Lack of Visual Dynamism & Missed Themes:
Prompt suffixes can solve both - pushing for proper adoption by CDs.
Bytedance’s Seedance 1.0 - Trending on Twitter for beating Veo3 in a video-to-video benchmarkUsage reports are mixed - evaluating if this is a viable option for Scaling teamLumaAI - Transform video art stylesManus & Genspark - Automating Veo3 video creation
18th June, 2025
Executive Summary
Veo3 is now used daily by 40–45 creatives across teams, but full adoption is hindered by default 720p quality (vs. 1080p in Google Flow, Kling), the inability to create 1:1 videos (preferred by Scaling), and the lack of a history feature in Playground. We’re addressing all of these issues + knowledge gaps via training.
AutoAI is improving, with Sluglines at 100% adoption. Next set of issues to fix:
The Veo3-AutoAI integration is close to stable, and the ChatGPT 4o integration has resolved NSFW filtering but still struggles with mid-sequence context and outfit consistency. Both are 6-10 days away from release.
New tools under evaluation include Seedance 1.0 by Bytedance, which recently outperformed Veo3 in benchmarks, LumaAI for visual style transfers, and Manus and Genspark for automating Veo3 video generation.
Veo3 Adoption & Challenges
What’s stopping us from achieving 100% adoption?
Veo3 Internal Contest
Training Sessions on Best Practices
AutoAI Updates
AutoAI Issues to Resolve in the Next 30 Days
(% values based on ACD+CD poll feedback & conversations with all CDs)
New tools being tested
Want to print your doc?
This is not the way.
This is not the way.
Try clicking the ··· in the right corner or using a keyboard shortcut (
CtrlP
) instead.