KLDH262

The problem

A folder of ten thousand AI generations named 00042_.png is write-only memory. KLDH262 indexes those folders in place — never moving or copying a file — by reading the generation metadata that ComfyUI, Automatic1111, and Civitai embed in each image: prompt, negative prompt, seed, steps, sampler, checkpoint, and the full LoRA stack.

All derived data — tags, ratings, characters, albums — lives in one JSON catalogue outside the image folders. No account, no server, no telemetry.

Hand-rolled binary parsers

The metadata layer reads PNG tEXt/iTXt/zTXt chunks directly, walks ComfyUI's node graph backward from the sampler through the LoRA chain to the checkpoint, and parses A1111 parameter blocks from both PNG text and EXIF. On a real mixed library it recovers a full generation record for roughly 96% of files.

/**
 * Walks a PNG's chunk list and pulls out IHDR dimensions plus every text
 * chunk. Both generators stash their parameters here: ComfyUI writes the
 * node graph as `prompt`, A1111 writes its parameter string as `parameters`.
 */
function readPng(buf: Buffer): { w: number; h: number; texts: PngText[] } | null {
  if (buf.length < 8 || buf.readUInt32BE(0) !== 0x89504e47) return null;
  const texts: PngText[] = [];
  let w = 0, h = 0, off = 8;

  while (off + 8 <= buf.length) {
    const len = buf.readUInt32BE(off);
    const type = buf.toString('ascii', off + 4, off + 8);
    const start = off + 8;
    const end = start + len;
    if (end > buf.length) break;

    if (type === 'IHDR') {
      w = buf.readUInt32BE(start);
      h = buf.readUInt32BE(start + 4);
    } else if (type === 'tEXt') {
      const data = buf.subarray(start, end);
      const nul = data.indexOf(0);
      if (nul > 0) {
        texts.push({
          keyword: data.toString('latin1', 0, nul),
          text: data.toString('latin1', nul + 1),
        });
      }
    } else if (type === 'iTXt') {
      /* keyword \0 compressed? \0 langTag \0 translated \0 text —
         malformed chunks are skipped rather than failing the file */
      // …
    }
    off = end + 4; // skip CRC
  }
  return { w, h, texts };
}

Small ML, no cloud

Two features most apps would ship as API calls are computed locally. A naive-Bayes-style 18+ classifier learns from the user's own labels — combining prompt terms, LoRA identity, and a skin-chroma histogram — and deliberately under-calls, because a human confirms each batch. "Find more like these" builds a TF-IDF centroid over prompt terms, LoRA overlap, and color composition, on the argument that for synthetic images the generation record is a stronger signal than any face embedding.

Duplicate detection uses two distinct perceptual hashes for two distinct questions: a 64-bit difference hash for "is this the same picture" and a color signature for "does this look alike".

Engineering discipline

The renderer has no filesystem or network access at all: 51 named IPC channels, media served over a custom protocol addressed by ID so no path ever crosses the boundary, and API tokens encrypted via the OS keychain. State lives in a 1,000-line pure reducer handling 107 action types with no Date.now() or Math.random() inside — React StrictMode's double-invoke is a true no-op.

Zero native modules — image decoding, thumbnailing, and hashing all go through Chromium — so CI builds installers for Windows, Linux, and macOS from one codebase with nothing to compile.

Built with

Every technology this project runs on, and the job it does here. Each one links to the other projects it appears in.

By the numbers