Photosonic AI

Converts natural language text prompts into images by processing descriptions through a diffusion-based generative model (likely Stable Diffusion or proprietary variant) with style tags embedded in the prompt pipeline. The system interprets style keywords (photorealistic, oil painting, anime, etc.) and applies them as conditioning parameters during the diffusion sampling process, allowing users to steer artistic direction without manual model fine-tuning.

Implements a token-based consumption model where free-tier users receive 10 monthly image generation credits, each credit consumed per image request regardless of resolution or style complexity. The system tracks credit usage per account via a database-backed quota manager, enforcing hard limits at the API gateway level and preventing generation requests when credits are exhausted until the monthly reset cycle.

Embeds Photosonic as a native module within Writesonic's copywriting platform, allowing users to generate images directly from within content creation sessions without context switching. The integration exposes a unified API surface where generated images are automatically linked to associated copy, enabling batch workflows where marketing copy and supporting visuals are created in a single session with shared metadata (campaign name, brand guidelines, etc.).

Parses natural language prompts to extract style directives (photorealistic, oil painting, anime, watercolor, sketch, etc.) and encodes them as conditioning vectors that guide the diffusion model's sampling trajectory. The system maintains a curated taxonomy of supported styles with associated embedding representations, allowing the model to blend multiple style descriptors (e.g., 'photorealistic oil painting') into a composite conditioning signal that influences both aesthetic and structural aspects of generation.

Supports sequential generation of multiple images within a single session, with each request consuming one credit from the user's monthly quota. The system queues generation requests, processes them serially (or with limited parallelism), and aggregates results into a downloadable collection. Quota deduction happens atomically per request, with failed generations (timeouts, errors) typically not consuming credits, though this behavior may vary by plan tier.

Generates images at fixed resolutions (typically 512x512 or 1024x1024 pixels) and exports in PNG or JPEG formats with configurable compression. The system does not perform post-generation upscaling; resolution is determined at generation time by the underlying diffusion model's configuration. Export format selection affects file size and quality characteristics but not the underlying image content.

Optimizes end-to-end generation latency (typically 15-30 seconds from prompt submission to image delivery) through model quantization, inference batching, and GPU resource allocation strategies. The system likely uses a lighter diffusion model variant or reduced sampling steps compared to competitors, trading some quality for speed. Latency varies based on queue depth and server load, with peak hours potentially extending generation time to 45+ seconds.

Tracks generation history per user account, storing metadata about each image generated (timestamp, prompt used, style applied, resolution, credit cost). The system provides a dashboard view of usage patterns, remaining credits, and generation history with filtering/search capabilities. Analytics data is persisted in a user-scoped database and accessible via the web dashboard; no API export of analytics is mentioned.

Manages multiple subscription tiers (free, paid plans) with different monthly credit allocations and feature access levels. The system enforces tier-based limits at the API gateway, preventing free-tier users from accessing paid features and throttling generation speed for lower tiers. Billing is handled via Stripe or similar payment processor, with automatic monthly credit reset tied to subscription renewal date.