Wan 3.0 Reference-to-Video Spicy

Wan 3.0 Reference-to-Video Spicy is a multimodal video generation model developed by Alibaba (Tongyi Wanxiang), released in August 2026. Designed for high-energy visual dynamics and stylized action generation, it leverages extended spatiotemporal attention mechanisms and reference feature anchoring to synthesize large-scale physical movements, aggressive camera trajectories, and dramatic temporal transitions while maintaining strict subject identity and clothing texture fidelity. The Spicy version is an unrestricted global variant that relaxes content limitations while preserving the original model's generation capabilities.

On the iCreat platform, this model operates via the https://api.icreat.ai/v1/task/submit/aliyun/wan3-0/reference-to-video-global endpoint using an asynchronous two-step API mode (submit task, then query result). It supports three resolution tiers: 480P, 720P, and 1080P. Pricing is calculated per second, with both reference video input duration and output video duration being billable. The Spicy version shares identical capabilities, parameters, and pricing with the standard R2V version, differentiated only by the endpoint path, the X-ICREAT-AI-GROUP: wan-spicy request header, and its relaxed content restrictions.

Model Positioning

Wan 3.0 Reference-to-Video Spicy is positioned as a high-dynamic video synthesis model tailored for global users requiring unrestricted content generation. It is ideal for producing dynamic action videos, film and television special effects pre-visualization, interactive game visual assets, and high-impact commercial advertising pipelines. By supporting combined inputs of reference images, videos, and audio, it provides creators with flexible multimodal control while pushing the boundaries of visual dynamics and dramatic transitions.

Core Capabilities

High-Energy Visual Dynamics The model excels at synthesizing large-scale physical movements, aggressive camera trajectories, and dramatic temporal transitions. It is built on extended spatiotemporal attention mechanisms that allow for complex, high-impact visual sequences.

Reference Feature Anchoring It maintains strict subject identity and clothing texture fidelity through reference feature anchoring. Users can input reference images, videos, and audio simultaneously to guide the generation process with high precision.

Unrestricted Global Generation As the Spicy variant, it relaxes content limitations for global users, enabling the creation of unrestricted and edgy content that standard models may filter, without compromising on generation quality.

Native Audio Support The model natively supports audio generation, producing video content with synchronized audio and visual elements by default.

Pricing

Note: The above are iCreat platform prices, consistent with Alibaba Cloud Bailian official pricing (¥0.45/0.9/1.8 per second for 480P/720P/1080P). iCreat platform prices shall prevail.

Application Scenarios

• Dynamic action video production
• Film and television special effects pre-visualization
• Interactive game visual asset generation
• High-impact commercial advertising pipelines
• Stylized action and high-energy visual content creation

Model Comparison

Comparison Table 1: Wan 3.0 R2V Versions (iCreat Endpoint)

Note: Spicy and Standard versions share identical pricing, differentiated by endpoint path and request header; Prime accelerated version is priced at approximately 1.4 times the standard version.

Comparison Table 2: R2V vs T2V vs I2V (Global Endpoint Series)

Note: R2V, T2V, and I2V share the same unit price per second, but total task costs differ. For R2V, when a reference video is used, its duration is included in the billable total duration (Total Cost = Unit Price × (Reference Video Input Duration + Output Video Duration)). T2V and I2V are billed only based on output duration. For precise cost calculation, please use total task cost as the basis for comparison.

Why Choose Wan 3.0 Reference-to-Video Spicy?

• Unrestricted content generation for global users, relaxing standard model limitations
• High-energy visual dynamics with aggressive camera trajectories and dramatic transitions
• Strict subject identity and clothing texture fidelity via reference feature anchoring
• Combined multimodal reference inputs (image, video, audio) for precise control
• Native audio generation for synchronized audiovisual content
• Flexible resolution (480P/720P/1080P) and aspect ratio options

Specifications

Architecture

The model is built on extended spatiotemporal attention mechanisms and reference feature anchoring. It processes multimodal reference inputs (images, videos, and audio) alongside textual prompts to anchor subject identity and stylistic elements. The architecture is optimized for synthesizing high-energy physical movements and complex temporal transitions while maintaining high fidelity to the provided references. It operates via an asynchronous task-based API flow where requests are submitted to the task endpoint and results are polled via the task result endpoint.

Notes

• The ratio parameter is required; an aspect ratio must be specified when submitting a task.
• The duration parameter ranges from 4 to 30 seconds; values outside this range will be rejected.
• Reference media (images/videos/audio) must be publicly accessible HTTPS URLs; private or authenticated addresses may cause task failure.
• The model natively supports audio generation; generated videos include synchronized audiovisual content by default.
• The Spicy global version relaxes content restrictions compared to the standard version.
• Both reference video input and output are billable: Total Cost = Applicable Unit Price × (Reference Video Input Duration + Output Video Duration); when no reference video is used, the input duration is counted as 0.
• Please keep your API Key secure and avoid hardcoding it in client-side code or public repositories.

Frequently Asked Questions

How do I call Wan 3.0 Reference-to-Video Spicy? The model uses an asynchronous two-step API. First, submit the task via a POST request to the endpoint https://api.icreat.ai/v1/task/submit/aliyun/wan3-0/reference-to-video-global, including the Authorization: Bearer <API Key> header and the X-ICREAT-AI-GROUP: wan-spicy header. Then, query the task result using the task ID via the https://api.icreat.ai/v1/task/result endpoint.

What is the difference between the Spicy global version and the standard version? They share identical model capabilities, parameters, and pricing. The differences lie in the endpoint path (reference-to-video-global vs reference-to-video) and the required X-ICREAT-AI-GROUP: wan-spicy request header. Additionally, the Spicy global version relaxes content restrictions for global users.

How is billing calculated? Billing is based on the formula: Total Cost = Applicable Unit Price × (Reference Video Input Duration + Output Video Duration). Both reference video input and generated video output are billable. When no reference video is used, the input duration is counted as 0. Worked examples: For 720P, generating a 5-second video without a reference video costs $0.50 ($0.10 × 5). Using a 5-second reference video and generating a 5-second output costs $1.00 ($0.10 × (5 + 5)).

What are the differences between R2V, T2V, and I2V? They share the same unit price per second, but total task costs differ. For R2V, when a reference video is used, its duration is included in the billable total duration. T2V and I2V are billed only based on output duration. R2V supports combined inputs of text, images, videos, and audio, whereas T2V supports only text, and I2V supports text and images.

What reference types does the media array support? The media array supports three reference types: reference_image, reference_video, and reference_audio. These can be used in combination to provide multimodal guidance for the video generation process.

What is the maximum single duration? The duration ranges from 4 to 30 seconds. The duration parameter must be an integer, and values outside this range will be rejected.