Wan Image: specs, benchmarks, and how to run it on Lyceum

Get started: call Wan Image on Lyceum

You can generate high-quality assets with Wan Image using Lyceum Technology's OpenAI-compatible API. Because this is a dedicated image generation model, you must use the images/generations endpoint with standard Bearer authentication, rather than the standard chat completions endpoint used for text models.

Integrating the model into your existing application requires zero new SDKs. You point your HTTP client to the Lyceum endpoint and pass the correct model string. Here is the exact code required to generate an image:

import requests
 
 response = requests.post(
 BASE_URL + "/images/generations",  # BASE_URL: the base URL shown in your Lyceum dashboard
 headers={"Authorization": "Bearer <your lyceum api key>"},
 json={"model": "lyc-wan-image", "prompt": "a sunset over the ocean", "aspect_ratio": "1:1"},
 )
 print(response.json["image_url"])
 

Pricing and region for Wan Image

Wan Image is available on Lyceum's serverless inference API, which is specifically optimized for cost-efficient, high-throughput generation tasks. The model is priced at a flat rate of $0.005 per image, making it highly scalable for bulk generation workflows.

To meet the strict data residency requirements of our European customers, all API requests for Wan Image are processed exclusively in our eu-north1 hosting region. Your prompts, reference data, and generated images are processed within the European Union, with GDPR-compliant processing in European data centres.

What Wan Image is good at

Photorealistic portraits and skin textures

Wan Image excels at generating highly realistic human subjects, setting a strong baseline for commercial photography replacement. It produces natural skin textures, accurate lighting interactions, and structural stability that avoids the artificial, plastic look common in earlier diffusion models. This makes the model highly effective for lifestyle photography, fashion mockups, and character design where authenticity is paramount.

Multi-language text rendering

Unlike many Western-centric image models that struggle with typography, Wan Image renders text inside the image rather than leaving it to a separate design step. Alibaba Cloud's Model Studio documentation for the Wan image models specifies prompt support for Chinese and English of up to 5,000 characters, and publishes no wider language list, so treat any broader figure as untested until you have run it yourself. The capability is still valuable for enterprise teams creating localized marketing assets, academic charts, and typography-heavy designs, because it reduces the need for post-production text overlays in graphic design software.

Precise color and composition control

The model offers fine-grained control over the visual output, which is critical for professional workflows. It adheres strictly to composition prompts and can accurately render specific color palettes based on hex codes or reference ratios. This is especially important for enterprise users who need to match strict brand guidelines. Furthermore, its underlying architecture utilizes a shared latent space for text and visual semantics. This ensures that complex, multi-element prompts are rendered logically, with objects placed correctly in relation to one another without blending concepts inappropriately.

Limitations and trade-offs

Where Wan Image falls short

While Wan Image is a highly capable all-rounder for commercial and creative tasks, it is not the absolute leader in strict prompt adherence for highly complex, surreal compositions. Users pushing the boundaries of spatial reasoning may occasionally need two or three prompt iterations to land the exact arrangement described in a dense, multi-subject brief. Models specifically tuned for extreme instruction following might require fewer attempts for these edge cases.

Additionally, because Wan Image is optimized for high-fidelity output, its generation latency can be slightly higher than heavily quantized, lower-resolution models. If your application requires sub-second, real-time image generation, such as live interactive avatars or instant visual feedback loops, a smaller distilled model might be a more appropriate choice.

Finally, while its multi-language text rendering is excellent and highly versatile, its English typography consistency can sometimes be edged out by models that have been fine-tuned exclusively on English-only datasets. Teams generating assets entirely in English might notice occasional spelling artifacts in very long sentences, requiring a quick regeneration. Despite these trade-offs, the model remains an exceptionally strong choice for general-purpose, high-resolution visual content creation.

Benchmarks and how it compares

Wan Image benchmark results

Alibaba's Wan image models are scored in public blind-vote arenas, where raters compare two generations of the same prompt without knowing which model produced which. On that evidence the family sits mid-field on raw aesthetic preference, and its case rests on what a generation costs to reach that quality rather than on topping the table.

Elo ratings and provider list prices from the Artificial Analysis Text to Image Leaderboard, read 3 August 2026.

Comparison to sibling models

The models above Wan in that arena are priced accordingly. GPT Image 2 (high) tops it at 1,339 Elo but lists at $211.00 per 1,000 images, and Reve 2.1 follows at 1,299 Elo and $24.00 per 1,000. The Wan image models land between 1,109 and 1,147 Elo at $21.00 to $64.00 per 1,000 at their source providers, and on Lyceum this endpoint is $0.005 per image, or $5.00 per 1,000. That is the trade being made: you give up the top of the aesthetic table and pay a fraction of the leader's rate, which is the right call for bulk catalogue work, localization and variant generation, and the wrong one when a single hero image has to be the best available.

Using it in production

Production configuration for Wan Image

When deploying Wan Image via Lyceum Technology, the model operates on our serverless inference API, which balances output quality with efficient processing speeds. At a flat rate of $0.005 per image, cost forecasting is straightforward and highly predictable. For a production workload generating 10,000 images per month, such as an e-commerce catalog update or a dynamic marketing personalization platform, the total inference cost is $50.

The API accepts standard parameters such as prompt and aspect_ratio, allowing you to integrate it into automated content pipelines. Because the endpoint is fully managed by Lyceum, your engineering team does not need to worry about VRAM allocation, batch sizing, or managing cold starts on dedicated GPU instances. The infrastructure scales automatically to meet your concurrent request volume.

Furthermore, because the model natively outputs high-resolution images, you can often bypass the need for secondary upscaling steps in most web and digital print workflows. This reduces the overall complexity of your pipeline and decreases the time-to-delivery for visual assets. For best results in production, we recommend passing detailed, descriptive prompts that specify lighting, camera angles, and color hex codes to fully leverage the model's precise control capabilities.

Running Wan Image on EU-sovereign infrastructure

Why run Wan Image on Lyceum

For European enterprises and AI startups, data sovereignty is a hard requirement, not an optional feature. Most major API providers route image generation prompts through US-based servers, creating significant compliance risks when processing proprietary product designs, internal mood boards, or sensitive marketing briefs. Lyceum Technology solves this fundamental issue by hosting Wan Image entirely within our eu-north1 region.

By running Wan Image on Lyceum, your requests are processed in European data centres in Spain, Paris and the Nordics, with this model served from eu-north1. Billing is per image with no base fee, so a bulk generation run costs what the arithmetic says it costs. You get the simplicity of an OpenAI-compatible endpoint, requiring no code changes beyond updating the base URL and API key, combined with GDPR-compliant processing in European data centres.

With per-image billing and no minimum commitments, Lyceum allows you to scale your creative workflows securely and cost-effectively. Whether you are building an AI application on EU infrastructure or transitioning away from expensive hyperscaler credits, our platform provides the reliability and transparency needed for production deployments. You maintain complete control over your data while leveraging state-of-the-art generative capabilities.

Sources

[1] arXiv: Wan, Open and Advanced Large-Scale Video Generative Models; [2] Alibaba Cloud Model Studio: Text-to-image with Qwen-Image and Wan models; [3] Alibaba Cloud Model Studio: Wan-Image generation and editing API reference; [4] Artificial Analysis: Text to Image Leaderboard; [5] arXiv: High-Resolution Image Synthesis with Latent Diffusion Models