DX Builder
DX Builder
BENCHMARKED DATA · DIRECT DASHSCOPE API · MULTIMODAL DIRECTING/Wan 3.0 (DashScope) Technical Reference & Production Guide
Open Engine Studio

Wan 3.0 Engine Manual: Measured Benchmarks & Prompt Engineering

Wan 3.0 (`wan3.0-video`) is Alibaba's flagship multimodal video generation model, delivering native 1080p rendering, full physical dynamics, native multi-language lip-sync dialogue, and multi-asset referencing up to 30 seconds per continuous take. All performance metrics below reflect direct DashScope API test runs conducted by DXBuilder.

Wan 3.0 Direct APIMeasured BenchmarksNative 1080pNative Lip-Sync & AudioUp to 30s TakesMulti-Asset Referencing
ALIBABAALIBABA· modelo oficial
Wan 3.0 Engine Manual: Measured Benchmarks & Prompt Engineering
Render Speed33–46m (720p) · 50m (1080p)
Native Resolution480p · 720p · 1080p Native
Max Take Length2s to 30s continuous
Supported Ratios16:9 · 9:16 · 1:1 · 4:3 · 3:4
Multimodal Input10 Img + 5 Vid + 5 Audio + Doc/Link
Audio PipelineNative Lip-Sync + Foley SFX
Commercial RightsFull Commercial License

1. Architecture, Verified Specs & Alibaba Announcements

Long Continuous Takes (2s to 30s) [Measured by us]

Generates continuous takes from 2 to 30 seconds per request without temporal breakdown when structured with timed shot lists.

Native Multimodal Conditioning [Measured by us]

Accepts up to 10 images, 5 video clips (≤15s total), 5 audio tracks (≤15s), and 1 document (100MB/50 pages) or web link per render job.

Native Dialogue, Lip-Sync & Foley [Measured by us]

Synthesizes timed dialogue enclosed in {brackets} with verified European Portuguese and regional dialect lip-sync accuracy plus atmospheric SFX.

Material & Brand Asset Fidelity [Measured by us]

Accurately maps reference logos and textures onto dynamic surfaces (cloth embroidery, neon signs, glowing embers) without graphic degradation.

Zero Wait-Time Multi-Reference [Measured by us]

Our benchmark tests confirm that scaling reference assets from 0 to 4 items adds zero additional GPU rendering latency.

Alibaba Roadmap Features [Announced by Alibaba, Untested]

Instruction-based video editing, chained temporal video extension, and live web-search context injection remain unverified in production.

30s · 720P · 1 foto + logo · falas PT-PT · 10 cortes aos tempos escritos
Vídeo vertical: gerado e medido pela DXBuilder · grelha: demos oficiais Alibaba (wan.video)

2. Benchmark Comparison: Wan 3.0 vs. Seedance 2.5

Production ParameterWan 3.0 (DashScope Direct)Seedance 2.5 PRO
Max Single Take Duration30 Seconds (Native continuous) [Measured]Up to 30 seconds per take
Native Resolutions480p · 720p · 1080p Native (No AI upscaling)480p · 720p (1080p/4K via Topaz Upscaling)
Audio & Dialogue EngineNative lip-sync + environmental SFX + dialects [Measured]Separate audio generation / post-sync pipeline
Render Duration (30s Take)33–46 min (720p) · 50 min (1080p) [Measured]2–5 min (fast pipeline)
Multimodal Input Capacity10 images + 5 videos + 5 audios + 1 doc/link50 reference images + reference video and audio
Cost per Second (DXBuilder)7.5 cr/s (480p) · 15 cr/s (720p) · 30 cr/s (1080p)25 cr/s (480p) · 35 cr/s (720p) · up to -25% discount at 30s

3. Directing Modes & Multimodal Input Limits

1 · Timed Shot-List Directing (Text + Native Audio)

Gatilho: Format prompts with absolute time markers and explicit dialogue tags: HARD CUT at Xs — [SCALE/ANGLE]: action {spoken dialogue}.

Regras: Continuous prose prompts yield 0–1 cuts in 15s; timed shot lists execute 4–12 cuts [Measured by us].

"00:00-00:05 CLOSE-UP: Merchant whispers {"O porto fecha à meia-noite."} [AUDIO: wind howling]. HARD CUT at 00:05 — WIDE SHOT: Storm surges over stone pier."

2 · Multimodal Asset Transfer (Image / Logo / Audio Locking)

Gatilho: Upload reference images (subject ID, wardrobe, logo patch) and audio clips with explicit identity locks.

Regras: Subject photos must be clean and upright (EXIF rotation ruins facial geometry). Reference asset count (0 to 4) does not alter render time [Measured by us].

"Apply character facial structure from @Image1 and wardrobe jacket patch from @Image2. Subject turns 45 degrees left. Material: heavy distressed leather."

3 · Rapid 5-Second Iteration Pipeline [Reported by third parties]

Gatilho: Draft and calibrate composition, lighting, and camera motion in 5-second test renders before committing to full 30-second 1080p final passes.

Regras: Saves 80%+ credit budget during framing calibration; upgrade to 30s/1080p only once timing and asset recognition lock.

"5s Test: 35mm low-angle tracking shot, protagonist walks 30% across frame left-to-right. Verify silhouette and backlight before full sequence render."

4. Measured Directing Rules & Prompt Syntax

Timed Hard Cuts [Measured by us]HARD CUT at 04:00 — MEDIUM SHOT: Actor turns to camera | HARD CUT at 09:00 — EXTREME CLOSE-UP: Hand grips copper dial.
Lip-Sync & Dialogue Tagging [Measured by us + AtlasCloud]Character speaks directly into lens: {"Temos apenas três minutos antes da ignição."} [AUDIO: heavy mechanical hum, no background music]
Look Lock Block [Measured by us][LOOK LOCK: overcast daylight 5600K, charcoal wool overcoat, wet cobblestone environment, muted cyan color grading]
Explicit Physical Grounding [Measured by us]Real gravitational weight, splashing water droplets with inertia, cloth drag resistance, no floating, no weightless wire-work.
Anti-IP Injection Block [Measured by us][NEGATIVE: no superhero emblems, no commercial franchise logos, no copyright costumes, no smooth slide floating motion]
Frame Percentage Vectoring [Reported by third parties]Subject accelerates from right edge to center, occupying 40% vertical frame height, moving 35% horizontal distance per second.

5. Production Prompt Blueprints

Cinematic Multi-Cut Narrative Take (15s — European Portuguese Lip-Sync)

Timed multi-shot scene with native dialogue lip-sync, ambient sound design, and verified physical weight [Measured by us].

[LOOK LOCK: Moody neon-lit alley, wet asphalt, 35mm anamorphic, heavy rain]
00:00-00:04 MEDIUM TRACKING SHOT: Detective walks forward with heavy footsteps, rain dripping from brimmed hat.
00:04-00:09 HARD CUT — CLOSE-UP: Detective stops under amber streetlight, looks straight into lens, speaking firmly: {"Não há mais tempo para negociações. O navio já partiu do cais."}
00:09-00:15 HARD CUT — OVER-THE-SHOULDER WIDE: A black cargo boat vanishes into foggy harbor waters. Real gravitational water displacement, heavy splashes.
[AUDIO: continuous heavy rainfall, distant foghorn, crisp vocal dialogue, no background synth music]
[NEGATIVE: no floating, no weightless wire-work, no superhero emblems, no frozen locked-off camera]

Brand Asset Materialization & Dynamic Physics (30s Native 1080p)

Full 30-second benchmarked sequence demonstrating reference logo reproduction onto physical surfaces with real mass dynamics [Measured by us].

[LOOK LOCK: Blacksmith forge workshop, warm 3200K tungsten backlight, volumetric smoke and rising embers]
00:00-00:08 LOW ANGLE SLOW DOLLY IN: Blacksmith hammers glowing hot steel bar on iron anvil with heavy impact shocks.
00:08-00:16 HARD CUT — MACRO SHOT: Hammer strikes molten metal surface, embossing the DXBuilder geometric logo from @Image1 into glowing orange iron, sparking embers scattering with authentic physics.
00:16-00:24 HARD CUT — MEDIUM PROFILE: Blacksmith plunges the branded metal blade into cold water vat. Instant dense steam eruption and boiling fluid turbulence.
00:24-00:30 HARD CUT — CLOSE-UP HERO SHOT: Water drips off the engraved brand logo on dark Damascus steel, camera maintains subtle hand-held organic drift.
[AUDIO: metallic anvil strikes, roaring forge fire, loud water quenching hiss, deep room tone]
[NEGATIVE: no static frozen camera, no synthetic plastic texture, no franchise trademarks]
Wan 3.0

6. Credit Calculator & Generation Economics

3s10s30s
Estimated Total150 créditos
Generate in Studio

7. Frequently Asked Questions

Based on our direct API measurements (August 2026), a 30-second 720p take averages ~40 minutes (33–46 min range). A 30-second 1080p take renders in ~50 minutes (+11% duration vs 720p). Queue wait times stabilized to ~0.5 min after launch scaling. You can close your browser; renders complete autonomously in your DXBuilder library.
👑 No prompts, no effort

Don't want to write prompts? Step into the movie yourself.

Premium presets run on the Seedance 2.5 engine — upload a selfie and an AI director shoots a 3-act blockbuster starring you. From 15 seconds to 1:30.

Browse all Premium presets
Free credits at sign-up No mandatory subscription Auto-refund on failures

Explore the other DXBuilder engines