WAN 2.5 Text-to-Video
WAN 2.5 text-to-video with one-pass audio/video sync, lip-sync, multilingual prompts and custom voice upload. 5–10s at 20–40 credits per second.
WAN 2.5 Text-to-Video generates a talking or narrated clip from a single prompt with one-pass audio/video synchronization. The model produces the voiceover and lip movement together rather than generating picture and sound separately, and it handles Chinese and other languages more reliably than earlier WAN versions. On Conut.ai it renders 5 or 10-second clips at 480p, 720p or 1080p for 20, 30 or 40 credits per second, and you can upload a custom voice track up to 15 MB.
- Duration: 5s / 10s
- Resolution: 480p / 720p / 1080p
- Price: 20–40 credits/s
- Audio: One-pass A/V sync
How to use WAN 2.5 Text-to-Video
Write for one speaker and one runtime. The model generates voice and lip movement in the same pass.
- Write the scene and the line Name the speaker, the setting and the exact words. Keep the line short enough for five or ten seconds at a natural pace.
- Pick the size by destination 1280×720 or 1920×1080 for landscape, 720×1280 or 1080×1920 for vertical. The 832×480 options are the cheap draft sizes.
- Upload audio when timing is fixed If you already have the voice track, upload it. The model will sync the performance to the supplied audio instead of generating its own.
- Disable expansion for exact copy Prompt expansion can reword the script. Turn it off when the line is a tagline or legal copy.
- Review sync before resolution Check that the mouth and the voice agree at the start. A beautiful 1080p render with drifting lip sync is still unusable.
Settings reference
WAN 2.5 Text-to-Video is built around one-pass audio/video sync: the prompt produces the picture and the voice together.
| Setting | Value | Notes |
|---|---|---|
| Prompt | Required | Describe the scene, the speaker and what they say. |
| Audio upload | Optional, up to 15 MB | Supply a track when the voice or timing must be exact. |
| Duration | 5s / 10s | Default is 5 seconds. Match the line length to the runtime. |
| Size | 1280×720 / 720×1280 / 1920×1080 / 1080×1920 / 832×480 / 480×832 | Landscape and portrait at 480p, 720p and 1080p. Default is 1280×720. |
| Negative prompt | Optional | Useful for warped faces and lip artefacts in close-ups. |
| Prompt expansion | Optional toggle | Expands sparse prompts; disable it when the script is fixed. |
| Seed | Optional | Reuse a seed to keep a performance while adjusting the line. |
One-pass audio/video synchronization
Most video models generate silent pictures and leave audio to a second tool. WAN 2.5 generates both in one pass, so the voice, mouth movement and scene sound are produced together. That makes it the practical choice for talking-head clips, narrated product shots and social videos where a silent result would need a separate lip-sync pass.
The trade-off is control. Because the model invents the voice and performance along with the picture, you get less direct control over tone and timing than a dedicated voice tool plus a lip-sync tool. Uploading your own audio track is the way to keep a specific voice while still using WAN's synchronization.
- Voiceover and lip movement generated together
- Scene audio produced in the same pass
- Custom voice upload up to 15 MB
- Multilingual prompts, including Chinese
- Landscape and portrait size presets
WAN 2.5 pricing by resolution
Pricing is per second by resolution, and the spread between tiers is modest. A 5-second 480p clip costs 100 credits; the same clip at 1080p costs 200. Because the endpoint caps at 10 seconds, total cost stays predictable.
- Billed per second by resolution
- 10-second maximum
- Six size presets across landscape and portrait
- Audio generation and upload do not change the rate
| Resolution | Credits/s | 5s | 10s |
|---|---|---|---|
| 480p | 20 | 100 | 200 |
| 720p | 30 | 150 | 300 |
| 1080p | 40 | 200 | 400 |
Custom voice and multilingual prompts
Uploading a voice track is useful when the clip must match an existing brand voice, a narrator or a script that has already been recorded. Keep the file under 15 MB and make sure it fits the clip length; the model uses it as the audio target while generating the picture.
WAN 2.5 is notably better than earlier WAN versions at Chinese prompts and at switching between languages. Write the spoken line directly in the prompt when you want the model to generate the voice, and keep the visual description separate so the performance is not diluted by staging details.
- Upload a voice track up to 15 MB for a specific voice
- Write spoken lines in the target language
- Keep dialogue and visual description in separate sentences
- Use portrait sizes for social talking-head formats
- Test a 5-second line before committing to 10 seconds
Prompting a talking shot
A talking shot needs three things: a speaker, a line and a setting. Introduce the speaker and framing first, then the spoken line, then the environment and light. If the camera should stay locked, say so; a moving camera during speech makes lip-sync harder to judge.
Keep the line short enough to fit the clip. A 5-second clip holds roughly one sentence; 10 seconds holds two or three. For product narration, write the line as a benefit statement rather than a paragraph, and let the visual carry the detail.
- Speaker, line, setting: one sentence each
- Keep the camera locked for the clearest lip-sync
- One sentence for 5 seconds, two or three for 10
- Avoid overlapping dialogue and complex action
- Use a clean frontal or three-quarter angle for faces
- Lock the seed while testing delivery
- Use the negative prompt for concrete artifacts
- Turn off prompt expansion when the line must be verbatim
Limitations and alternatives
The 10-second cap is the main constraint. There is no 4K, and the model is not a precision dubbing tool: if the mouth must match a prerecorded track frame by frame, a dedicated lip-sync workflow is more reliable. Very fast speech or multiple speakers in one shot also reduce accuracy.
WAN 2.6 is the better choice for silent cinematic 1080p at a similar rate. Wan 3.0 adds 30-second durations but does not specialize in talking shots. Seedance 2.5 is the 4K option for cinematic work without dialogue.
- 10-second maximum
- No 4K
- Not a frame-accurate dubbing tool
- Fast speech and multiple speakers reduce lip-sync quality
- For silent 1080p, compare WAN 2.6
Prompt examples for WAN 2.5 Text-to-Video
- Product presenter: A presenter stands in a bright minimal studio, medium shot, looking directly at the camera. She says: This bottle keeps your coffee hot for twelve hours. Soft key light, clean background, locked camera. Clear room tone. One benefit statement fits a 5-second clip.
- Travel narration: A traveler walks along a coastal path at sunrise, wide shot from behind, then turns to the camera. He says: The best views are never at the end of the road. Warm morning light, gentle sea breeze. Waves and distant gulls. A turn toward camera gives the line a natural start.
- Chinese social clip: A young chef stands behind a counter in a modern restaurant, medium close-up. She says in Mandarin: 新鲜食材,每天现做。 Warm pendant light, shallow depth of field, locked camera. Kitchen ambience. Write the spoken line in the target language.
- Explainer shot: A scientist stands beside a clean lab bench, medium shot, holding a small device. He says: This sensor detects leaks before they become failures. Cool white lab light, neutral background, slight push-in. Quiet lab room tone. Keep the prop visible but avoid hand gestures near the face.
- Fashion voiceover: A model walks through a sunlit studio, then stops and faces the camera, medium shot. She says: Designed for movement, made to last. Warm editorial light, soft shadows, locked camera. Footsteps and quiet studio tone. Stop the walk before the line so lip-sync has a stable frame.
- Custom voice test: A narrator stands in a dark recording studio, close-up, facing the camera. The uploaded voice track speaks the line. Single warm key light, black background, locked camera. Studio room tone. Upload a voice track under 15 MB for this workflow.
Pro tips for WAN 2.5 Text-to-Video
2.5 is a talking-picture model. Direct the performance, not the lens.
- One speaker, one idea Splitting a short clip between two speakers produces rushed dialogue and weak sync. Use separate generations and cut them together.
- 480p is the draft tier 832×480 and 480×832 cost 20 credits per second. Use them to test scripts, then re-render the chosen take at 1080p.
- Write dialogue you could say in one breath Long, clause-heavy sentences are where sync breaks. Short sentences with clear pauses give the model room to animate.
- Name the emotion Warm, urgent, whispered: the prompt steers delivery as well as content. A neutral script produces a neutral read.
- Use 2.6 for silent cinematic work When there is no speech, WAN 2.6 Text-to-Video offers longer runtimes and more camera control without the speech-sync constraints.
Frequently asked questions
How much does WAN 2.5 text-to-video cost?
It costs 20 credits per second at 480p, 30 at 720p and 40 at 1080p. A 5-second 720p clip is 150 credits.
How long can a WAN 2.5 clip be?
5 or 10 seconds. Ten seconds is the maximum.
Does WAN 2.5 generate audio automatically?
Yes. It produces voiceover, lip movement and scene audio in one pass, synchronized with the picture.
Can I upload my own voice track?
Yes. The endpoint accepts an audio track up to 15 MB, which is the way to keep a specific voice or a prerecorded script.
Does it support Chinese prompts?
Yes. WAN 2.5 is multilingual-friendly and handles Chinese prompts more reliably than earlier WAN versions.
What resolutions and sizes are available?
480p, 720p and 1080p, with landscape and portrait presets for each.
Is WAN 2.5 a precise dubbing tool?
No. It synchronizes generated speech well, but frame-accurate dubbing of an existing track is better handled by a dedicated lip-sync workflow.
What is the difference from WAN 2.6?
WAN 2.6 is a silent cinematic model with seed, negative prompt, 1080p and audio upload, capped at 15 seconds. WAN 2.5 specializes in one-pass talking shots and caps at 10 seconds.
Compare with other text-to-video models
- WAN 2.5 Image-to-Video
- WAN 2.6 Text-to-Video
- Wan 3.0 Text-to-Video
- Kling Text-to-Video
- Seedance 2.0 Text-to-Video
- Browse all AI video models
Specifications
- Provider: Wavespeed
- Model type: Text to video
- Credit pricing: 480p: 20 credits · 720p: 30 credits · 1080p: 40 credits
- Capabilities: new, cinematic, starter, audio-sync, multi-shot
About Conut.ai
Conut.ai is an AI image, video and audio generation platform with 50+ creative tools. It supports text-to-image, image-to-video, face swap, motion control, relighting, inpainting and video effects through models including Nano Banana Pro, Seedream, GPT Image, Sora 2, Kling, WAN and Seedance. Start free — no credit card required.
Explore more AI tools
- Seedance 2.5 Turbo Text-to-Video
- Seedance 2.5 Image-to-Video
- Seedance 2.5 Turbo Image-to-Video
- Seedance 2.0 Mini Text-to-Video
- Seedance 2.0 Mini Image-to-Video
- MiniMax H3 Text-to-Video
- MiniMax H3 Image-to-Video
- MiniMax H3 Reference-to-Video
- Wan 3.0 Text-to-Video
- Virtual Try-On
- Photo Restore
- Cinema Studio