Ultra Selection
Stop guessing which model to use. Ultra Selection is WaveSpeed's intelligent routing layer that analyzes your prompt complexity and performance requirements in real-time. We automatically select and execute the optimal model—whether it's FLUX for detail, SDXL for speed, or a specialized LoRA for style—ensuring the best possible result every time.
How Ultra Selection Works
A three-step process to optimize every generation request.
1
Intent Analysis
The system analyzes your input prompt to understand the specific requirements (e.g., "photorealistic portrait," "anime line art," "vector logo"). This ensures the right model family is identified before any generation begins.
2
Dynamic Benchmarking
We compare available models against real-time performance metrics including current load, inference speed, and historical quality scores for similar tasks. This includes all major models like those in our Best Open Source Image Models collection.
3
Smart Routing
Your request is routed to the best-fit model. For example, a simple sketch request might go to SDXL Turbo (Low Cost / High Speed), while a complex architectural render routes to FLUX.1 [dev] (High Quality). Custom LoRA adapters are also factored into routing decisions when style-specific generation is detected.
Why Use Ultra Selection?
Compare the manual model selection process with our automated approach.
- Need highest visual fidelity? → Ultra Selection routes to FLUX.1 [dev]
- Need fastest inference? → Routes to SDXL Turbo or LCM
- Need lowest cost? → Routes to the most economical model for your task
- Need video generation? → Beta routing between Wan 2.1 and Kling
- Want all options in one API? → WaveSpeed handles it automatically
Q & A
What is Ultra Selection?
Ultra Selection is an API feature that acts as a "Meta-Model." Instead of specifying a model ID (like flux-dev), you specify a strategy (like best-quality or fastest-response), and WaveSpeed handles the rest.
Can I set specific criteria?
Yes. You can configure the selection logic based on three priorities: Quality (always pick the best visual fidelity), Speed (always pick the fastest inference), or Cost (always pick the most economical model).
Does it work for video generation?
Currently, Ultra Selection is optimized for Image Generation. Video support is in beta, routing between models like Wan 2.1 and Kling based on prompt duration and motion complexity.
How much does it cost?
Ultra Selection itself is free to use. You are billed based on the actual model that ends up processing your request. If the system routes to a cheaper model, you pay less.
Can I see which model was used?
Yes. The API response metadata always includes the model_id of the model that actually performed the generation, ensuring transparency.