I've been working on Wan 2.6 - an AI video generator that creates 1080p videos from text and images. The thing that kept me up at night? Making the audio actually sync properly with the visuals.
Let me share the journey, the challenges, and what I learned building this.
Why I Built This
Here's what frustrated me about existing AI video tools:
The audio sync was awful. Generate a video of someone talking, and their lips move like a badly dubbed movie. It just looked... wrong.
Quality was all over the place. Your character would morph halfway through. One frame they're a young woman, next frame they're somehow a different person.
Limited control. You'd get what you get, no way to fine-tune or adjust.
I wanted to build something that actually worked well. Something I'd want to use myself.
What Wan 2.6 Does
Let me break down the core features:
Text-to-Video
Type a description, get a video.
Example: "A chef flipping a pancake in a sunny kitchen" → You get a 15-second video of exactly that in 1080p.
Image-to-Video
Got a static image? Bring it to life.
Upload a photo and describe what should happen. "Make her wave at the camera" or "zoom into the product" - that kind of thing.
Text-to-Image
Need custom visuals? Generate images to use in your videos or standalone.
Everything outputs at 1080p resolution, 24fps, with native audio synchronization.
The Audio Sync Nightmare
This was the hardest part by far.
When you generate video with AI, you're creating each frame. But when someone speaks, their mouth needs to match the sounds they're making. Not just roughly - it needs to be precise.
The Challenge
Think about it: when you say "P" or "B", your lips close. When you say "O", your mouth forms a circle. Every sound has a specific mouth shape, and it happens at exact milliseconds in the audio.
Getting an AI to:
Understand the audio timing
Generate the right mouth shapes
Keep the face consistent
Make it look natural
...is incredibly complex.
What Didn't Work
Attempt 1: Generate video first, add audio later.
Result: Looked like a ventriloquist dummy. Terrible.
Attempt 2: Generate audio first, then create video to match.
Result: Better, but timing was always slightly off. Still weird.
Attempt 3: Generate both simultaneously with shared information.
Result: Finally! This worked.
The breakthrough was realizing audio and video can't be separate processes. They need to be generated together, each informing the other in real-time.
Took months to get right, but now the lip sync actually looks believable.
Keeping Characters Consistent
Another major challenge: making sure your subject doesn't morph into a different person.
Early versions would do this thing where the character would slowly change. Ask for "a woman reading a book" and by the end, she's somehow a completely different person.
Not ideal for any kind of storytelling.
The Solution
The system now "remembers" what your subject looks like in the first frame and maintains those features throughout. It tracks key characteristics - facial features, clothing, style - and keeps them consistent.
It's not perfect (AI never is), but it's way better than the morphing mess we started with.
The 1080p Challenge
Here's the thing: generating high-quality video is computationally expensive.
1080p at 24fps means generating tons of pixels. And each frame needs to be:
High quality
Consistent with previous frames
Generated in reasonable time
We had to get creative:
Smart upscaling: Generate at a lower resolution first, then intelligently upscale. The trick is making the upscaling look natural, not artificial.
Frame interpolation: Generate key frames, then create smooth transitions between them. Cuts computational load in half while keeping motion smooth.
Optimization everywhere: Batch processing, smart caching, and tons of other tweaks to make it actually usable.
Currently, a 5-second video takes about 45 seconds to generate. Not instant, but way better than the 10+ minutes early versions took.
Making Static Images Move
One of my favorite features is image-to-video. Upload a static image, describe what should happen, and watch it animate.
The challenge? Making the motion look natural.
You can't just randomly move pixels around. The system needs to understand:
What objects are in the image
How those objects should move realistically
What motion makes sense for the prompt
A person waving should look like a natural wave. A car driving should follow physics. A product rotating should maintain its shape.
This took a lot of iteration, but when it works well, it's pretty magical.
Real-World Use Cases
I built this thinking about content creators and marketers. But people use it for all sorts of things:
Educators creating teaching materials and explainer videos
Small businesses making product demos without expensive video production
Authors creating book trailers on limited budgets
Social media managers generating quick content for posts and stories
Marketers testing video concepts before investing in full production
Hobbyists just making cool stuff for fun
The variety of use cases has been surprising and awesome.
What Works Well
Let me be honest about what Wan 2.6 does really well:
✅ Audio sync - This is our strong point. Lip movements actually match speech naturally.
✅ Quality - 1080p output looks professional, not AI-generated garbage.
✅ Consistency - Characters stay recognizable throughout the video.
✅ Ease of use - No complex settings or technical knowledge needed.
✅ Multiple workflows - Text-to-video, image-to-video, text-to-image all in one place.
Current Limitations (Being Real)
Nothing's perfect. Here's what we're still working on:
Video length: Currently capped at 15 seconds. Longer videos while maintaining quality is technically challenging.
Processing time: 45 seconds per 5-second video isn't bad, but faster would be better.
Fine control: Users want more precise control over specific elements. Working on it.
Edge cases: Weird or complex prompts sometimes produce unexpected results.
Hardware requirements: Quality generation needs decent computing power.
I'm not hiding these - they're just the reality of current AI video technology.
Lessons Learned
Solve The Hardest Problem First
I wasted time on UI before tackling audio sync. Should've solved the toughest technical challenge first, then built around it.
Quality > Speed (Usually)
I could've launched with 720p and saved on compute. But in video, quality is immediately noticeable. People care about how it looks.
Users Surprise You
I thought this would mostly be for marketing videos. The actual use cases are way more diverse and creative than I imagined.
Iteration Is Everything
The first version was terrible. The tenth version was better. The hundredth version actually worked. Keep iterating.
Listen To Feedback
Users find problems you'd never spot. They want features you'd never think of. Pay attention.
What's Next
We're actively working on:
Longer videos (30+ seconds)
More control over specific elements and scenes
Faster generation through better optimization
Better motion in image-to-video
More customization options
The roadmap is driven by what users actually need, not just what's technically cool.
Try It Out
Wan 2.6 is live at wan26.io.
Whether you're creating content for social media, making educational materials, or just experimenting with AI video - give it a shot and see what you can create.
The interface is straightforward: enter your prompt or upload an image, hit generate, and get your video. No complex setup, no technical knowledge required.
What would you create with AI video generation? Any specific use cases you'd love to see supported? Drop your thoughts in the comments - I'm genuinely curious what the dev community thinks! 💬
Top comments (0)