Undersketch is a private desktop system for training custom image models on an artist's own work. An artist assembles a dataset from their own images, trains a LoRA adapter over FLUX, then generates new images in their style, including img2img.
Undersketch
Undersketch is a private desktop system for training custom image-generation models on an artist's own work, designated GSV-A9 (pace-layer A) spanning the media-ecology, life-systems-heart, and infrastructure domains, currently in-flight (last updated 2026-05-04). Mechanically, an artist assembles a training dataset by dragging their own images into the app, trains a LoRA adapter over FLUX via Replicate's FLUX LoRA pipeline, and then generates new images in their own established style, with prompt and strength controls plus img2img generation from a selected input image; trained models are kept in a personal library rather than published to a shared public model zoo. The application is built as a Tauri (Rust) desktop shell wrapping a TypeScript/Vite frontend; datasets and generated outputs are stored in Backblaze B2, while model training itself is delegated entirely to Replicate rather than run locally. A deliberate architecture decision routes all data access through a single app-owned Hono/Node backend over one /v1 API surface: the frontend holds no direct provider logic, authenticates via Clerk, attaches the resulting bearer token to every call, and never reaches a backend-as-a-service URL directly. This boundary is the product of a real migration: the project began on Supabase, calling Auth, PostgREST, and Edge Functions directly from the client, and has since moved identity and persistence onto Neon behind the app-owned server, with Clerk taking over session handling; the legacy Supabase schema remains as source material during the cutover. The single-server-boundary discipline (centered on src/api.ts) is what the given material credits with making that migration tractable. Within the GSV portfolio, Undersketch was surfaced by the same filesystem-indexing pass that identified Thisbe, Typhon, and other infrastructure-tier projects, placing it in that same discovered cluster, though no specific cross-project integration is described. Current status: LoRA training and generation pipeline working, B2 storage configured, mid-migration from Supabase to Neon/Clerk/Hono.