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

Get started: call Image Ultra on Lyceum

Integrating Image Ultra into your application requires minimal setup. Because Lyceum provides an OpenAI-compatible API, you can generate images using standard HTTP requests without managing underlying GPU infrastructure. For image generation models, you must use the dedicated images/generations endpoint rather than the standard chat completions endpoint.

Unlike text models where you might use the OpenAI Python client with a custom base URL, image generation on Lyceum requires a direct HTTP POST request. You must pass your Lyceum API key as a Bearer token in the authorization header.

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-image-ultra", "prompt": "a sunset over the ocean", "aspect_ratio": "1:1"},
 )
 print(response.json()["image_url"])

Pricing and region for Image Ultra

Image Ultra is available exclusively in the eu-north1 region, ensuring all data processing remains within European borders. It operates on a strict pay-per-use model priced at $0.005 per image. There are no service tiers to choose between: Lyceum sells one Serverless Inference product, and latency and capability follow from the model you pick. There are no minimum commitments, no subscription fees, and no idle compute costs. You only pay for the exact number of images you successfully generate, making it highly predictable for production workloads.

What Image Ultra is good at

Sub-second generation latency

The primary advantage of Image Ultra is its extreme speed. Lyceum publishes no architecture detail for Image Ultra, but the sub-second class it belongs to rests on step distillation. Adversarial Diffusion Distillation, the method behind SDXL Turbo, samples in one to four steps, against the 50-step undistilled SDXL baseline it is measured on in the paper. Whatever the internals, Lyceum's published figure for Image Ultra is a result in under one second. For engineering teams, this means Image Ultra is viable for real-time applications where traditional diffusion models would introduce unacceptable delays and break the user experience.

Cost-efficient scaling for high-volume workloads

Because the model requires significantly fewer compute cycles per generation, it is highly cost-effective at scale. At $0.005 per image, applications that require high-volume generation can operate with predictable, low unit economics. This is particularly valuable for dynamic e-commerce cataloging, personalized marketing asset generation, or user-generated content moderation pipelines where thousands of images must be processed daily without inflating cloud infrastructure budgets.

Interactive and real-time workflows

Image Ultra excels in environments where users expect immediate visual feedback. This includes live UI mockups, real-time gaming asset generation, and interactive AI chat interfaces. The sub-second response time prevents workflow interruption, enabling a fluid user experience that heavier, multi-step models cannot support. Developers can wire the API directly to user input fields, generating visual concepts on the fly as the user types, which fundamentally changes how end-users interact with generative design tools.

Limitations and trade-offs

Where Image Ultra falls short

The aggressive distillation required to achieve sub-second generation times comes with inherent trade-offs in image fidelity. Image Ultra is not designed for hyper-realistic, high-resolution commercial photography. When compared to heavier, non-distilled models, it may struggle with intricate micro-details, complex multi-subject compositions, and precise text rendering. If your application requires flawless typography inside the generated image, a slower, high-parameter model is a better choice.

Resolution constraints and scaling

Ultra-fast models are typically optimized for standard resolutions, such as 512x512 or 1024x1024 pixels. Pushing the model to generate ultra-high-resolution outputs natively can degrade both speed and structural coherence. Attempting to force a 4K print-ready asset directly from the API will likely result in visual artifacts or repeated patterns. Teams needing high-resolution outputs should pair Image Ultra with a secondary upscaling model, though this negates the sub-second latency advantage.

Prompt adherence on complex instructions

While Image Ultra handles standard, descriptive prompts effectively, it has a lower capacity for deep semantic understanding compared to models with massive parameter counts. Highly specific prompts requiring exact spatial relationships (e.g., "a red cube exactly three inches to the left of a blue sphere"), specific lighting setups, or nuanced architectural details may require multiple generation attempts to achieve the desired result. Engineers should design their prompt templates to be concise and keyword-heavy rather than conversational.

Benchmarks and how it compares

Sub-second image model benchmarks

Image Ultra competes in the ultra-fast, distilled diffusion category. No independent ELO or latency score is published for the lyc-image-ultra endpoint, so the table below sets Lyceum's own published figure beside vendor-published numbers for two open models in the same class, SDXL Turbo and FLUX.1 Schnell. Read it as a class comparison rather than a head-to-head benchmark: the three figures were not measured on the same hardware, at the same resolution, or with the same harness.

Sources: Stability AI's SDXL Turbo announcement for the 207ms A100 figure, and the FLUX.1 [schnell] model card for the 1 to 4 step range. Licensing differs across the three: SDXL Turbo is released under a Stability AI non-commercial research license, FLUX.1 [schnell] under Apache 2.0, and Image Ultra is offered only as a hosted Lyceum endpoint.

Comparison to sibling catalogue models

When evaluating Image Ultra against standard diffusion models in the Lyceum catalogue, the primary differentiator is the speed-to-quality ratio. The heavier image models in the catalogue take longer per image, and Lyceum publishes no generation time for them, but they deliver superior photorealism, accurate anatomy, and reliable text rendering. Image Ultra sacrifices that top-tier fidelity to guarantee sub-second delivery. For infrastructure leads, this means Image Ultra is the default choice for interactive applications, while heavier models should be reserved for asynchronous batch processing or final-asset production where latency is not a constraint.

Using it in production

Production configuration for Image Ultra

Deploying Image Ultra in production requires understanding its specific API mechanics. Unlike text generation models, image generation does not use token streaming. You send a single synchronous POST request to the images/generations endpoint, and the API returns a JSON payload containing the image_url. Your application must be prepared to handle this synchronous block, though the sub-second latency of Image Ultra minimizes thread-locking issues.

Key parameters to configure in your production payload:

Prompt engineering

Keep descriptions concise. Distilled models respond better to direct, keyword-rich prompts rather than lengthy conversational instructions. Place the most important subjects at the beginning of the prompt string.

Aspect Ratio

The model supports standard ratios (e.g., 1:1, 16:9, 9:16). Stick to these defaults to avoid structural warping or unexpected cropping.

Error handling

Implement retry logic for network timeouts, even though the model itself processes requests in under one second.

Cost estimation at scale

Because Image Ultra is billed at a flat rate of $0.005 per image, forecasting production costs is highly predictable. If your application generates 100,000 images per month, your total compute cost is exactly $500.00. Hosted in the eu-north1 region, this price includes all infrastructure overhead. Furthermore, Lyceum charges no egress fees for retrieving the generated images, ensuring that high-traffic applications do not incur hidden network costs.

Running Image Ultra on EU-sovereign infrastructure

Why run Image Ultra on Lyceum

For European teams building AI applications, data residency is often a strict regulatory requirement. Many hyperscalers and US-based API providers route inference traffic through North American data centers, creating compliance risks for enterprise workloads. Lyceum solves this by hosting Image Ultra entirely within the eu-north1 region. This means processing in European data centres (eu-north1 for this model), with no training on customer data and no prompts or outputs retained after processing, which is a clear path to GDPR-compliant inference for European teams.

European infrastructure advantage

Unlike API wrappers that rent compute from larger cloud providers and pass the markup on, Lyceum prices Image Ultra without that margin stacking. Inference for the model runs in European data centres in Spain, Paris and the Nordics, and you are billed per image with no base fee. This approach avoids the capacity bottlenecks and block-reservation requirements common on public clouds.

Frictionless integration

Switching to Lyceum requires almost no engineering effort. Because the API is fully OpenAI-compatible, developers can migrate existing image generation pipelines by updating the base URL and API key. Whether you are building an interactive design tool or a serverless GPU inference pipeline, Lyceum provides the speed of Image Ultra with the security of EU-sovereign infrastructure. You avoid vendor lock-in while paying per image, with no subscription and no minimum commitment.

Sources

[1] arXiv: Adversarial Diffusion Distillation; [2] Hugging Face: stabilityai/sdxl-turbo Model Card; [3] Hugging Face: black-forest-labs/FLUX.1-schnell Model Card