RunDiffusion

Executes Stable Diffusion and related generative models on cloud-provisioned GPU infrastructure (likely NVIDIA A100/H100 or similar), abstracting away local hardware requirements. The workspace likely maintains persistent GPU instances or on-demand allocation pools to minimize cold-start latency, with request queuing and load balancing across multiple inference nodes. Users submit prompts via web UI and receive generated images within seconds to minutes depending on model size and queue depth.

Interactive UI for composing text prompts, adjusting numerical hyperparameters (sampling steps, guidance scale, seed, resolution), and selecting model variants without command-line or code interaction. The interface likely includes prompt syntax highlighting, parameter sliders with real-time preview updates, and a history/favorites system for reproducible generations. Changes to parameters trigger immediate re-queuing of inference jobs with new settings.

Accepts multiple generation requests (either via UI form submission or API) and manages them through a priority queue with fair scheduling across concurrent users. The system likely implements backpressure handling, job status tracking, and result delivery via webhooks or polling. GPU resources are allocated dynamically based on queue depth and user tier, with estimated completion times provided upfront.

Provides a curated catalog of Stable Diffusion checkpoints (v1.5, v2.1, XL, community fine-tunes) with version pinning and automatic model loading into GPU memory. The platform abstracts model selection via a dropdown or tag system, handling model weight downloads, VRAM allocation, and compatibility checks transparently. Users can lock generations to specific model versions for reproducibility across time.

Accepts uploaded images as conditioning input for img2img workflows, with optional mask-based inpainting to regenerate specific regions. The system encodes input images into latent space, applies noise based on a strength parameter, and denoises with the prompt as guidance. Masking is likely implemented via alpha channel or separate mask image, with feathering to blend inpainted regions smoothly.

Maintains a persistent database of all user-generated images with associated metadata (prompt, parameters, model version, timestamp, seed). The system indexes this data for full-text search on prompts and tags, with filtering by date range, model, or parameter ranges. Users can organize generations into projects/folders, favorite results, and export generation logs for external analysis.

Enables multiple users to access shared projects with role-based access control (view-only, editor, admin). The system maintains a shared generation queue and result storage, with audit logs tracking who generated what and when. Permissions are enforced at the project level, with granular controls over image deletion, parameter modification, and member management.

Exposes HTTP endpoints for submitting generation requests, polling job status, retrieving results, and managing projects programmatically. The API uses JSON payloads for request/response, with standard HTTP status codes and error messages. Authentication is likely via API keys with rate limiting per tier, and responses include job IDs for asynchronous tracking.

Applies neural upscaling models (likely Real-ESRGAN or similar) to increase image resolution and reduce artifacts. The system accepts images at any resolution and upscales to 2x, 4x, or 8x the original size while preserving detail. Upscaling is performed on cloud GPU infrastructure, with optional post-processing (sharpening, denoising) applied before delivery.