Alpaca

Integrates Stable Diffusion's inpainting model directly into Photoshop's native editing canvas, allowing users to select regions and generate photorealistic content that blends with existing image context. The plugin marshals Photoshop's selection masks as inpainting prompts, processes them through a local or cloud-hosted Stable Diffusion inference endpoint, and composites results back into the active layer while preserving non-selected pixels. This approach eliminates context-switching between applications and maintains Photoshop's non-destructive editing paradigm through layer-based composition.

Enables users to generate new images from text descriptions using Stable Diffusion's text-to-image pipeline, with iterative prompt refinement and parameter tuning (guidance scale, sampling steps, seed control) exposed through Photoshop's UI. The plugin tokenizes text prompts, encodes them through CLIP text encoder, and passes embeddings to the diffusion model's UNet for iterative denoising. Users can regenerate with different seeds, adjust guidance strength to balance prompt adherence vs. creativity, and preview variations before committing to canvas.

Scales generated or existing images to higher resolutions using Stable Diffusion's upscaling pipeline or latent-space super-resolution techniques. The plugin encodes the input image into latent space, applies upscaling operations (2x, 4x, or custom factors), and decodes back to pixel space while optionally applying detail refinement through diffusion-based enhancement. This preserves image coherence better than naive interpolation and can add fine details consistent with the original content.

Applies artistic styles or visual aesthetics to images using Stable Diffusion's img2img pipeline with style-specific prompting or LoRA (Low-Rank Adaptation) fine-tuned models. The plugin encodes the input image into latent space, applies noise injection at a configurable strength (denoise parameter), and guides denoising toward a target style through prompt conditioning. Users can select from preset styles (oil painting, watercolor, anime, photorealism, etc.) or provide custom style descriptions, with control over how strongly the style is applied.

Enables processing multiple images or generating multiple variations in sequence through a batch queue system. The plugin accepts a list of prompts, images, or parameters, processes them serially or in parallel (if cloud-based), and outputs results as separate layers or files. This capability abstracts away manual iteration, allowing users to generate 10+ variations or process an entire folder of images without manual triggering for each operation.

Abstracts the underlying inference provider (local GPU, cloud APIs like Replicate or RunwayML, or self-hosted servers) behind a unified plugin interface. Users can configure which backend to use, switch providers without changing workflows, and optionally fall back to alternative providers if one is unavailable. The plugin handles API authentication, request marshaling, and response parsing for each provider, allowing seamless switching between local and cloud inference based on performance, cost, or availability constraints.