Create AI videos with 240+ avatars in 160+ languages.
I’ve spent my whole career working in video, and even then it’s hard to keep up with how fast AI video is evolving. New tools are launching every month, each promising cinematic results, but in reality the quality varies a lot.
Instead of relying on demos or hype, I tested the leading AI video generators myself. I ran the same prompt through each tool to see how they actually perform in terms of realism, motion, consistency, and overall output.
Check out the video below to quickly see the 13 best AI video generators compared side by side with the same prompt, and read on for my full review of each tool.
The best AI video generators
For cinematic videos:
- Veo 3.2: Most realistic visuals with strongest audio integration
- Seedance 2.0: Best motion physics and cinematic camera movement
- Kling 3.0: Most stable, controllable, production-ready cinematic generator
- Sora 2: Most advanced storytelling and narrative generation
- Runway Gen-4.5: Strong camera motion but weak detail stability
- Luma Ray3: Most beautiful UX with elegant visual output
- PixVerse 5.5: Best for short, dynamic, social-ready videos
- Grok Imagine: Most creative, imaginative, emotionally driven visuals
- WAN 2.6: Most reliable flexible tool for unrestricted prompts
- Pika 2.5: Playful, low-res tool for viral-style content
- Adobe Firefly: Strong image engine, weak video generation
- Hailuo 2.3: Outdated video quality, not competitive
For work:
- Synthesia: Best for avatar-driven internal comms and training videos
How I tested these AI video generators
To compare these AI video generators fairly, I ran the same set of prompts through each one, testing both text-to-video and image-to-video.
My test prompt
A cinematic aerial shot of a floating city above glowing clouds at golden hour. The camera follows a cloaked traveler riding a hovering speeder, weaving between crystal towers suspended in the air. Below, rivers of light flow between floating islands with waterfalls pouring into mist. Airships drift past in the background, their metallic hulls reflecting sunlight and lens flare. Magic runes flicker on tower surfaces as drones circle. The traveler’s cape ripples in the wind while the camera performs a smooth tracking orbit with natural motion blur and shallow depth of field. Soft volumetric rays pierce through the clouds, creating prismatic reflections. Hyper-realistic textures (metal, glass, fog), cinematic teal-orange color grading, and a warm, atmospheric tone.
My evaluation criteria
To evaluate each AI video generation model, I assessed performance across the following criteria:
- Accuracy (prompt fidelity): Does it follow the prompt correctly without missing elements or hallucinations
- Realism (visual + physics): Lighting, textures, motion, and whether it feels believable
- Consistency (stability): Do objects, motion, and details hold together across frames
- Creativity (interpretation): Does it add anything interesting beyond the prompt or feel flat
- Audio quality: How well sound is generated and synced (if supported)
- Performance: Generation time, reliability, and ease of use
- Constraints: Clip length, cost, access limitations, and control over output
Comparison table
1. Veo
- Best for: Cinematic realism + integrated audio
- Max resolution: Up to 4K
- Max clip length: ~6–8 sec
- Generation time: ~5–7 min
- Audio: Yes (clean, well-synced)
Accuracy: Follows prompts very closely with no visible hallucinations or drift. In my test, all key elements were rendered as expected and stayed consistent throughout the shot.
Realism: Lighting, atmosphere, and spatial depth are where Veo really stands out. Effects like light passing through structures and reflections feel physically believable, with only minor instability in fast-moving details like propellers.
Consistency: The scene holds together well across frames, with stable motion and fewer artifacts than faster versions of the model. Objects, lighting, and environment remain coherent even during movement.
Creativity: More controlled than experimental. Veo executes the prompt accurately and delivers strong cinematic results, but doesn’t add unexpected creative variations or stylization.
My verdict
Veo 3.2 produced one of the most polished outputs in my testing. The combination of accurate prompt execution, strong lighting, and stable motion makes it feel closer to a finished cinematic shot than most other tools.
Audio is a major differentiator. Compared to tools like Kling, where sound can feel flatter or less connected, Veo’s audio was clean, well synced to the visuals, and added a noticeable layer of realism without needing post-production.
The image-to-video mode performed better than text-to-video in my tests, with stronger depth, atmosphere, and physical plausibility. Motion and lighting held together more convincingly, especially in more complex scenes.
The trade-offs are clear. Clips are short, generation takes several minutes, and the cost is high. Camera movement also feels more limited compared to tools like Kling, which offer more flexibility for dynamic shots.
Bottom line: If you want realistic, cinematic video with integrated sound, Veo 3.2 is one of the strongest options available. It’s best suited for high-quality, narrative content rather than fast, high-volume workflows.
2. Seedance
- Best for: Cinematic motion and camera dynamics
- Max resolution: 720p
- Max clip length: ~15 sec
- Generation time: ~2–3 min
- Audio: Yes (environment + music, well synced)
Accuracy: Strong overall. It captures the intent of the prompt well, though it occasionally interprets elements instead of following them exactly.
Realism: Motion physics are one of the best I tested. Drone movement, propellers, and object interactions feel natural and stable. Lighting reacts dynamically across surfaces. The main limitation is resolution, as 720p reduces fine detail.
Consistency: Very reliable. Objects remain stable during movement and the scene holds together well across frames, even with multiple moving elements.
Creativity: Focused on cinematic execution rather than experimentation. The strength is in how scenes are staged. Camera movement, composition, and flow feel intentional and directed.
My verdict
Seedance 2.0 was one of the most impressive tools I tested from a filmmaking perspective. The camera actively follows the subject and builds a sense of progression instead of just animating a static scene. Both text-to-video and image-to-video felt closer to a film sequence than a generated clip.
The biggest strength is motion. Compared to Kling and Grok, Seedance handles dynamic movement and object physics more convincingly, especially with complex elements like drones. Lighting also evolves naturally during the shot, which adds to the cinematic feel.
Audio is another strong point. Environmental sounds, moving objects, and background music are layered and synchronized in a way that improves immersion without extra work.
The trade-offs are practical. Access is restricted to business plans or aggregators, resolution is capped at 720p, and prompt control is slightly less precise than more rigid tools.
Bottom line: Seedance 2.0 is one of the strongest tools for cinematic motion and scene composition. It is ideal when movement and shot design matter most, but resolution and access limits make it less flexible for everyday use.
3. Kling
- Best for: Controlled cinematic video generation
- Max resolution: 1080p
- Max clip length: ~15 sec
- Generation time: ~4–7 min
- Audio: Yes (improved, well synced)
Accuracy: Extremely strong. In my test, Kling followed the prompt closely with no unexpected elements, missing details, or hallucinations. It feels very controlled and predictable.
Realism: High level of realism across lighting, physics, and environmental behavior. Motion, reflections, and material response feel grounded, especially compared to more stylized tools.
Consistency: One of the most stable models I tested. Objects, characters, and environments remain coherent across frames, even in more complex scenes with multiple moving elements.
Creativity: More restrained than earlier versions. Kling 3.0 prioritizes stability and prompt accuracy over unexpected or experimental outputs, which makes it reliable but slightly less expressive.
My verdict
Kling 3.0 feels like a shift from experimental AI video toward something closer to controlled production. The output is clean, stable, and cinematic, with smooth camera motion and a clear sense of scene structure.
Compared to tools like Runway, Kling delivers stronger realism and more believable physics. Compared to Veo, it offers more flexibility in camera movement and scene control, even if the audio and emotional realism are not as strong.
The image-to-video mode is a standout. Kling extends environments beyond the original frame while maintaining spatial consistency, which makes scenes feel larger and more immersive rather than just animated.
There are still some limitations. Complex moving objects like drones can show slight distortion over time, and the model is slightly less visually bold than earlier versions. Generation times are also longer when working with more detailed scenes.
Bottom line: Kling 3.0 is one of the most reliable and production-ready AI video tools available. It is a strong choice when you need predictable, high-quality cinematic output with good control, even if it sacrifices some creative flair.
4. Sora
- Best for: Narrative storytelling and multi-scene video
- Max resolution: Up to 1080p
- Max clip length: ~4–12 sec
- Generation time: ~9–10 min
- Audio: Yes (dialogue + sound generation)
Accuracy: Very strong. Sora translates prompts into coherent scenes with a clear understanding of tone, structure, and intent, even with minimal input.
Realism: High. Lighting, motion, and environmental depth feel natural, with believable physics and strong sense of space, even at lower resolutions.
Consistency: Strong across scenes. Sora maintains continuity and structure better than most tools, especially when generating multi-shot or narrative sequences.
Creativity: Outstanding. This is where Sora clearly stands out. It does not just follow prompts but expands them into full scenes with emotion, pacing, and narrative flow.
My verdict
Sora 2 feels fundamentally different from other AI video tools. Instead of generating isolated clips, it creates scenes that feel directed, with a sense of story, emotion, and progression. In my testing, character motion, lighting, and world-building all worked together in a way that felt closer to a short film than a generated output.
The biggest strength is storytelling. Sora understands context and tone in a way that goes beyond visual accuracy, and it can generate dialogue, sound, and multi-scene sequences that feel cohesive. This makes it far more powerful for narrative content than tools focused purely on visuals.
That said, it comes with trade-offs. Generation is slow, costs are high, and access is still limited. There is also very little manual control after generation, as most of the output is driven entirely by the prompt.
Bottom line: Sora 2 is the most advanced tool I tested for storytelling and narrative video. It is ideal for cinematic concepts and creative direction, but less suited to fast workflows or users who need precise control.
5. Runway
- Best for: Camera motion and creative workflows
- Max resolution: 720p
- Max clip length: ~10–12 sec
- Generation time: ~5+ min (text-to-video)
- Audio: No native audio in video generation
Accuracy: Decent. The model follows the general idea of the prompt, but finer details and motion-specific instructions are not always executed correctly.
Realism: Weak in motion. While composition and lighting look good at first glance, details break down quickly during camera movement, with blur, distortion, and simplified physics.
Consistency: Mixed. Scenes can start strong, but objects lose clarity and stability once motion increases, especially in image-to-video.
Creativity: Strong in direction and composition. Camera movement feels intentional and cinematic, giving the impression of a directed shot even if the underlying realism is lacking.
My verdict
Runway Gen-4.5 stands out most for its camera work. In my testing, subject tracking and shot composition felt deliberate and well-directed, which gives the output a strong cinematic feel at a glance.
The problem appears as soon as motion increases. Fine details collapse, objects blur, and spatial consistency breaks down, especially compared to tools like Veo or Kling which maintain structure more reliably.
Image-to-video is noticeably weaker than text-to-video. While the initial still image can look highly detailed, that quality does not carry through once animation begins, and the scene loses much of its original atmosphere.
Runway works best as part of a broader creative workflow rather than a standalone production tool. Its ecosystem, editing tools, and flexibility are strong, but the core video output is not yet competitive for high-end or realism-driven projects.
Bottom line: Runway Gen-4.5 is a strong creative tool with excellent camera motion and workflow flexibility, but limited realism and stability make it better suited for experimentation than final production.
6. Luma Dream Machine
- Best for: Aesthetic visuals and artistic scenes
- Max resolution: Up to 4K (tested via Firefly)
- Max clip length: ~5–10 sec
- Generation time: ~2–7 min depending on mode
- Audio: No native audio
Accuracy: Strong. It interprets prompts well and captures both visual direction and overall mood with high precision.
Realism: Good in calm scenes, weaker in motion. Lighting, color, and composition look natural, but physics break down during fast movement or complex action.
Consistency: Better in image-to-video than text-to-video. Slower motion holds together well, but faster sequences can cause objects to blend or lose structure.
Creativity: Strong on atmosphere and visual style. Luma produces elegant, cinematic-looking scenes, especially in nature or fantasy settings, with a clear artistic direction.
My verdict
Luma stands out immediately through its design and overall experience. The interface is clean, minimal, and genuinely enjoyable to use, with well-organized controls and helpful visual presets that guide creative direction.
In terms of output, the model performs best in slower, more atmospheric scenes. Landscapes, lighting, and textures feel refined and visually polished, especially in fantasy or nature environments.
Where it struggles is motion. During faster camera moves or more dynamic action, physics become unstable and objects can blend together unnaturally. Compared to tools like Veo or Kling, it is less reliable for complex movement.
Image-to-video performs more consistently than text-to-video, with steadier motion and better object coherence. The tool also offers useful creative controls like start and end frames, presets, and a modify editor for refining outputs.
Bottom line: Luma is a beautifully designed tool that excels at mood, atmosphere, and artistic visuals. It is a great choice for aesthetic content and concept work, but less suited for fast motion or realism-driven scenes.
7. PixVerse
- Best for: Short-form cinematic and social content
- Max resolution: 1080p
- Max clip length: ~5–6 sec
- Generation time: ~1–2 min
- Audio: Yes (basic, built-in)
Accuracy: Good. It follows the general prompt well, especially in mood and composition, though fine details and motion instructions are not always precise.
Realism: Moderate. Scenes look cinematic at a glance, but detail drops during fast camera movement, with blur and weaker physics compared to top-tier tools.
Consistency: Decent for short clips. The scene holds together over a few seconds, but instability appears during sharper motion or more complex sequences.
Creativity: Strong. PixVerse prioritizes mood, style, and visual storytelling, often producing engaging and dynamic shots even if technical accuracy is not perfect.
My verdict
PixVerse V5.5 is clearly designed for short, engaging content. In my tests, camera movement felt dynamic and the overall scene had a strong cinematic atmosphere, making it well suited for social media or quick promotional clips.
The biggest limitation is duration. With a maximum of around 5–6 seconds, clips end just as they start to develop, which makes it difficult to build narrative or pacing.
Compared to tools like Kling or Veo, PixVerse sacrifices realism and detail for speed and accessibility. However, it is significantly faster and easier to iterate with, which makes it practical for high-volume content creation.
Audio is included but basic. It supports the mood of the scene but lacks clarity and depth compared to stronger implementations like Veo.
Bottom line: PixVerse V5.5 is a strong option for fast, short-form video creation with good visual style and accessibility. It works best for social content, but limited duration and weaker realism make it less suitable for cinematic projects.
8. Grok Imagine
- Best for: Creative concepts and cinematic mood
- Max resolution: 720p
- Max clip length: ~15 sec
- Generation time: ~1.5–2 min
- Audio: Yes (music + ambient, well integrated)
Accuracy: Good, but stylized. It follows the core idea of the prompt, though it interprets it more creatively rather than strictly reproducing every detail.
Realism: Not the focus. Physics can break intentionally, but the world feels coherent because it leans into a fantasy aesthetic rather than realism.
Consistency: Mixed. Scenes generally hold together, but there are occasional issues with motion stability and fine detail, especially in more complex elements.
Creativity: Excellent. This is where Grok stands out. It produces visually bold, imaginative scenes with strong lighting and a clear artistic identity.
My verdict
Grok Imagine feels very different from most AI video tools. Instead of trying to replicate reality, it leans into interpretation and mood. In my tests, the outputs felt expressive and cinematic, with strong lighting, unusual environments, and a clear sense of visual identity.
One of the most impressive aspects is audio. Grok generates both ambient sound and music that align closely with the tone of the scene, creating a more emotional and immersive result than many other tools.
The image-to-video mode also stands out for storytelling. Rather than simply animating a frame, it extends the scene and creates a sense of narrative progression, which adds depth to the output.
The main limitation is technical quality. Resolution is capped at 720p, fine details can become soft, and motion can feel unstable at times. Compared to tools like Kling or Veo, it is less reliable for precise or realistic outputs.
Bottom line: Grok Imagine is one of the most creatively expressive tools I tested. It is best used for concept development, mood exploration, and experimental storytelling rather than high-precision or production-ready video.
9. Wan
- Best for: Reliable generation and flexible prompts
- Max resolution: 1080p
- Max clip length: ~5–15 sec
- Generation time: ~1–5 min
- Audio: Yes (basic, with artifacts)
Accuracy: Good. It understands prompt intent well, even in more sensitive or unusual scenarios where other tools may fail.
Realism: Moderate. Lighting and scene structure are solid, but visuals often lean slightly stylized or cartoon-like rather than fully cinematic.
Consistency: Strong. Outputs are stable and reliable, with coherent scenes and predictable results across generations.
Creativity: Practical rather than expressive. It focuses on delivering usable results instead of pushing visual boundaries or cinematic style.
My verdict
WAN 2.6 works best as a dependable, everyday tool. In my testing, it consistently delivered usable results with clear scene structure, good lighting, and intentional camera movement.
The key advantage is flexibility. Unlike stricter tools like Sora, WAN handled prompts that were rejected elsewhere, including more dramatic or sensitive scenes, which makes it useful when other models block generation.
The trade-off is visual quality. Compared to tools like Veo or Kling, outputs feel more stylized and less physically accurate, and object motion can appear slightly unstable. Audio is included but contains noticeable artifacts and lacks depth.
Despite that, WAN is reliable. It generates consistently, handles prompts well, and offers longer clip durations than many competitors, which makes it practical for regular use.
Bottom line: WAN 2.6 is a dependable and flexible AI video tool that prioritizes usability over cinematic quality. It is a strong option for everyday workflows and situations where other tools fail, but not the best choice for high-end visual realism.
10. Pika
- Best for: Viral clips and social content
- Max resolution: 480p on free plan
- Max clip length: ~5–10 sec
- Generation time: ~2–4 min
- Audio: No
Accuracy: Basic. It captures the general idea of the prompt, but execution lacks precision and depth, especially in more complex scenes.
Realism: Weak. Motion physics are simplified, objects blur during movement, and the overall scene lacks believable interaction with light and environment.
Consistency: Limited. Some improvements over earlier versions, but scenes can still feel fragile, especially in image-to-video where motion breaks down more easily.
Creativity: Strong for playful visuals. The tool focuses on expressive, attention-grabbing outputs rather than realism or cinematic quality.
My verdict
Pika 2.5 feels more like a creative playground than a production tool. In my testing, it delivered visually engaging results with improved color grading and atmosphere compared to earlier versions, but it still falls short in realism and depth.
Text-to-video performs better than image-to-video. The latter struggles with motion stability and physical coherence, with objects blurring and interactions feeling artificial.
Compared to tools like Kling or Veo, Pika is far behind in lighting, physics, and overall cinematic quality. However, it is faster, more accessible, and better suited for quick, experimental content.
The biggest limitations are resolution and lack of audio. At 480p on the free plan, output quality is noticeably lower than most competitors, and the absence of built-in sound reduces immersion.
Bottom line: Pika 2.5 is best used for fast, playful, and social-first content. It works well for experimentation and viral clips, but does not meet the standard for cinematic or professional video production.
11. Adobe Firefly
- Best for: Adobe ecosystem and concept visuals
- Max resolution: 1080p
- Max clip length: ~5 sec
- Generation time: ~60–75 sec
- Audio: No (separate generation only)
Accuracy: Inconsistent. The model follows the general prompt, but details are often simplified or missing, and results lack depth.
Realism: Weak. Motion, physics, and lighting do not feel cinematic, and outputs often resemble basic animation rather than realistic footage.
Consistency: Moderate. The main elements are present, but the scene lacks cohesion and natural behavior, especially in motion.
Creativity: Strong for images, limited for video. Image generation is detailed and varied, but video feels less expressive and less developed.
My verdict
In my testing, Firefly performed better in text-to-video than image-to-video, but neither mode delivered competitive results. The outputs felt flat, with weak motion and limited cinematic depth.
The main advantage is speed. Generation is fast and consistent, which makes it useful for quick iterations or concept testing.
However, compared to tools like Kling or Veo, Firefly is clearly behind in realism, motion quality, and overall visual impact. The interface also feels cluttered and harder to navigate than most modern AI tools.
Bottom line: Firefly is useful for quick concept visuals and Adobe-based workflows, but its video generation is still behind the leading tools and not suitable for cinematic output.
12. Hailuo
- Best for: Image concepts and stylized visuals
- Max resolution: 1080p (lower for longer clips)
- Max clip length: ~6–10 sec
- Generation time: ~3–12 min depending on mode
- Audio: No
Accuracy: Basic. The model follows the prompt at a surface level, but the visual interpretation feels outdated and lacks depth.
Realism: Weak. Motion, lighting, and overall scene quality feel behind current standards, with outputs resembling older game-like visuals rather than cinematic footage.
Consistency: Limited. Scenes are stable enough to complete, but lack detail and dynamic behavior, especially in video generation.
Creativity: Mixed. Some elements like fabric simulation and color grading stand out, but overall results feel uninspired compared to competitors.
My verdict
In my testing, Hailuo 2.3 was one of the weakest video generators. Text-to-video results felt flat and outdated, with minimal camera movement and weak visual direction.
Image generation is the stronger side of the tool. The atmosphere, color grading, and certain elements like fabric movement can look interesting, but that quality does not carry over into video. Once animated, scenes feel static and lack realism.
Another major limitation is the lack of audio, which makes the output feel incomplete compared to most modern tools.
Compared to tools like Kling, Veo, or even PixVerse, Hailuo is clearly behind in motion quality, realism, and overall cinematic feel.
Bottom line: Hailuo 2.3 has some creative potential in image generation, but its video capabilities feel outdated and are not suitable for professional or cinematic use.
Synthesia
- Best for: Training, onboarding, internal communications, sales enablement, and localized content
- Output types: Avatar-led videos
- Languages: 160+ with built-in translation
Output quality: Very high for its category. In my testing, English avatar realism was especially strong, with natural facial expressions, stable lip-sync, and controlled gestures that feel appropriate for professional delivery.
Control & editing: Structured and predictable. The slide-based editor makes it easy to control scenes, layouts, and messaging, but it is designed for presentation-style content rather than dynamic editing. The AI video assistant lets you create videos by iterating with a chatbot, and includes motion graphics and AI-generated B-roll options using models like Veo 3 and Sora (which they offer free access to via their AI video generator tool).
Workflow & usability: Clear and linear. You typically start by uploading an existing document or a script, then you select an avatar and voice, edit your video and then generate. I think this works well for business users, and the integrated translation workflow is one of the strongest parts of the platform.
Scalability: Strong for enterprise teams. Collaboration tools, brand kits, and version control support team workflows.
Business features: Built for corporate use. Includes multilingual translation, secure video sharing, stock media integrations, and workspace management aligned with enterprise needs.
Synthesia is built specifically for business video at scale. You can turn scripts, documents, webpages, or slides into presenter-led videos without filming, and keep everything consistent and on-brand across teams.
The trade-off is flexibility and speed. Rendering takes several minutes, and the platform is clearly optimized for structured corporate content rather than dynamic or high-volume marketing workflows.
Bottom line: Synthesia is one of the strongest tools for enterprise video production. It is ideal for training, onboarding, and multilingual communication, but I probably wouldn't recommend it for social media or other marketing-style videos.
Kyle Odefey is a London-based filmmaker and Video Editor at Synthesia. His content has reached millions across TikTok, LinkedIn, and YouTube, even inspiring an SNL sketch, and has been featured by CNBC, BBC, Forbes, and MIT Technology Review.
Frequently asked questions
What’s the best AI video generator for business use cases like training, onboarding, and internal comms?
Synthesia. It turns scripts and docs into presenter-led videos with realistic avatars, 1-click translation, LMS exports, brand kits, and team workflows. If you want extra B-roll, pair Synthesia with Veo 3 or Sora 2 clips inside the same project.
What’s the best AI video generator for cinematic short films and emotional storytelling?
Veo 3 and Seedance 2 for the most natural acting, lighting, and camera language. If you have access, Sora 2 is excellent for multi-scene narrative flow. For strong results at a saner price, Kling is the practical alternative.
What’s the best AI video generator for fast social ads with sound in one tool?
PixVerse. Quick renders, built-in audio and optional speech, solid prompt control, and handy features like Fusion and Swap. Runners-up: Runway (great polish and 4K upscale) and Seedance for clean, stable motion.
What’s the best budget-friendly AI video generator for quick, reliable output?
Wan. Very low cost for short 720p/1080p clips, fast, and stable. Consider Seedance for similarly clean, dependable motion, and PixVerse off-peak pricing when you also want audio.
What’s the best AI video generator for product demos and app promos?
Runway. Excellent UI, strong image-to-video, scene expansion, and 4K upscale. If you’re starting from high-quality stills, Seedance or Kling add smooth motion and good physics.
What’s the best AI video generator for fashion, perfume, or mood-driven brand visuals?
Hailuo for gorgeous lighting, texture, and cinematic feel when atmosphere matters most. Luma Dream Machine is a close second for elegant, dreamy aesthetics and a great UX. For fast, artsy sketches, Grok Imagine is interesting.
What’s the best AI video generator for YouTube explainers and tutorials?
Synthesia. Presenter-led formats, clear voice options, templates, on-brand visuals, and translations make repeatable explainer production easy. Add Runway or PixVerse for quick B-roll, motion accents, and sound.
What’s the best AI video generator for multilingual localization at scale?
Synthesia. It handles 160+ languages with 1-click translation, natural voices, localized avatars, and LMS-friendly exports—perfect for turning one master video into many regional versions. For on-brand visuals, layer in Veo 3 or Sora 2 B-roll where needed.
