Image generation looks simple from the outside: you write a prompt, you get an image. The reality involves a six-stage pipeline with multiple quality checks, style normalization, and fallback handling. Here's the full flow.
Stage 1: Prompt enhancement
User prompts are short, ambiguous, and often lack the specificity that diffusion models need to produce high-quality output. Before we send anything to the generation model, we run the prompt through Suzhou 3.1 with a prompt engineering instruction that adds lighting descriptions, composition guidance, style keywords, and technical parameters.
A prompt like "a cat on a desk" becomes "a tabby cat resting on a wooden desk, soft natural light from the left, shallow depth of field, lifestyle photography style, sharp focus on the cat's face, muted earth tones." The enhancement takes 0.3-0.5 seconds and improves output quality significantly on short or vague inputs.
Stage 2: Style resolution and generation
Users can select from style presets (photorealistic, artistic, diagrammatic, etc.) or leave style unspecified. Style selection modifies the enhanced prompt with additional conditioning tokens and adjusts the negative prompt to suppress conflicting aesthetics.
We run generation on a fine-tuned diffusion model hosted on our own infrastructure. Generation takes 4-8 seconds depending on resolution and aspect ratio. For 1:1 at standard resolution, median generation time is 5.2 seconds.
Stage 3: Quality scoring and fallback
After generation, we run the output through a quality scorer that evaluates sharpness, composition balance, and adherence to the prompt. If the score falls below our threshold, we automatically regenerate with adjusted parameters — typically a modified seed and slightly tweaked negative prompt.
About 12% of generations trigger a fallback regeneration. Users don't see this; they just receive the higher-quality result. The total pipeline time for fallback cases averages 9 seconds.
Generated images are stored in your account and accessible from the Images tab. You can regenerate with variations, refine with a follow-up prompt, or download the original at full resolution.