ExpSoft ArckheensToolzBasic

1. Renamer

Bulk-renames a folder of files using <Subject>-<Concept>-NNNN.ext. Random shuffle by default to break lexicographical order bias during training. Recursive scan, configurable digit width, copy-vs-move toggle, live preview count.

Use case — 800 photos called IMG_20231104_205411.jpg become bjorkstyle-0001.jpg … bjorkstyle-0800.jpg, randomized, in a sibling renamed/ folder. The Kohya / OneTrainer / SimpleTuner trainer can now find them and won't infer dumb temporal patterns from the original timestamps.

2. Captions

For each image in a folder, creates a matching .txt sidecar (same basename). Optional fill-with-text for a common trigger word; existing .txt are backed up with a timestamp suffix before being overwritten. Top-level only (no recursion — one dataset folder at a time).

Use case — the renamed 800 images need 800 caption sidecars for Kohya. Click once: 800 empty .txt appear next to them. Or fill them with "bjorkpost, 1woman" as a starting point you'll refine in the next step.

3. Tag Editor

Search-and-replace (literal or regex) across every .txt in a folder, recursively. Backs up the original as .YYYYMMDDHHMMSS.bak before each pass. Auto-cleanup of orphan separators ("tag1, , tag2" → "tag1, tag2"). Live count of matches before you commit. Preview button = no-op dry run.

Use case — auto-generated captions from WD14 / JoyTag / Florence-2 are full of noise: "watermark", "signature", "blurry". Run three replaces with empty target, dataset becomes signal-only. Or normalise "blue eyes" / "blue_eyes" across 800 files in one pass. Indispensable for any auto-captioned dataset.

4. Prompts

Walks a folder, reads every .txt, flattens each to a single line (whitespace collapsed), concatenates into one prompts_made.txt. Optional Prefix / Suffix wrap each line. Recursive option, anti-recursion guard (output file is auto-skipped during scan).

Use case — after curating 800 captions in Tag Editor, generate one big prompts.txt ready for an A1111 / Forge / ComfyUI batch queue. Or feed it to a benchmark script to compare two checkpoints on the SAME captions your dataset uses.

5. Resizer

Visual gallery: 200 px thumbnails of every image in a folder, click-to-toggle selection, "Select All / None" for fast triage. Target size + longest-side / shortest-side mode. Live recompute of target dimensions under each thumb when you change settings. Output is always PNG (no lossy re-encoding) using WPF Fant interpolation (Lanczos-equivalent).

Use case — 800 reference photos at 4032×3024. Click to deselect the 12 blurry ones you spot in the thumbnails. Set "1024 longest side", click Resize. Done — the SDXL / Flux trainer eats them straight.

6. Video Merge

FFmpeg -f concat -safe 0 -c copy — lossless stream-copy concatenation. List of clips with Move Up / Move Down to set the final order. Min 2 clips. Generates a temporary list file in %TEMP%, cleans up after. Live ffmpeg / ffprobe log during the run.

Use case — 8 short clips of someone speaking on a podcast (~30 s each, all H.264 1080p 30fps from the same source) → one continuous 4-min video for the next step (Split, Frame Extractor, or the Diarizer's audio track). Note: lossless concat requires identical codec/resolution/fps; mismatched inputs need re-encoding (planned in V1.1).

7. Video Split

Two modes — By duration (segments of N seconds via -segment_time) or By parts (N equal slices computed from the source duration probed by ffprobe). Optional output framerate, optional audio drop. Live readout of source duration / resolution / fps / codec at browse time.

Use case — a 2-hour interview or lecture video. Split into 60×120-second chunks for AI video pipelines that cap at ~2 minutes (Wan, Hunyuan, AnimateDiff). Or into 10 equal parts for distributed processing across multiple machines.

8. Frame Extractor

Four modes: Keyframes only (select=eq(pict_type,I) — best for diverse, scene-cut frames), All frames (massive output), Interval (one frame every N seconds via -vf "fps=1/N"), or Custom -vf (power-user filter chain). Output PNG / JPEG / BMP / TIFF, configurable PNG compression level.

Use case — a music video of an actor or a stock footage reel becomes an image dataset. Keyframes-only on a 3-minute clip yields 80-200 visually distinct frames — feed those to the Renamer + Captions + Resizer pipeline above and you have a face / style LoRA training set in 10 minutes flat.

System

Windows 10/11 (64-bit). No GPU needed. Minimal RAM usage.

Setup

One-click FFmpeg download on first launch (~100 MB). After that, fully offline.

Supported Formats

Images: JPG, PNG, BMP, GIF, WebP, TIFF. Videos: MP4, AVI, MKV, MOV, WMV, FLV — anything FFmpeg supports.