Intro
Hi! Welcome and thanks for reading! I'll keep this short. :)
After making my first LoRA (Suito), I realized I needed better tools. I made some simple scripts to help with my workflow. I had starting building up many of them, so I bundled them into a Tauri app. Added more features and QoL improvements to them.
At first I made this for me and it was tightly coupled to my system. I decided I could open it up and post it for others. Some extra effort invested in easy setup and use of the app. Should work fine on linux/windows.
I use Kohya so the training section was designed around that. You can also just use this for the tools if you train in other ways.
Im still new to LoRA creation and image generation in general. Feedback would be great.
If you dont like my theme you can change the app colours in the settings.
Click here for a few screenshots
General Overview
Theres a handful or features so far. This is a overview of the main ones.
Data prep: Image cropping (center crop, square pad, trim, interactive manual), batch conversion to webp, and background removal. There's also a video frame scrubber that extracts frames at various quality presets, which is handy for building datasets from video sources.
Dataset Gallery & Stats: aggregate dataset analysis, image grid with thumbnails, caption overlays, and resolution badges, single image caption or batch editing.
Captioning & tagging: Caption file generation/repair, a bulk tag editor with find/replace across all files (with backup/restore), and an auto-tagger. The "tagger" itself started as a port of SmilingWolf's WD Tagger but I've since rebuilt most of it, added various QoL, batch processing, tag blacklisting, custom sort modes, and trigger word prefixing. It includes 3 models (SwinV2 v3, ViT Large v3, EVA02 Large v3) which download from HuggingFace only when selected and used.
Analysis tools: A tokenizer that checks how your trigger words tokenize against your base model, a LoRA tensor scanner that auto-detects format and computes effective weights and norms, stats, metadata, most active layers, full layer analysis, and graph viewer that parses TensorBoard training logs and plots loss curves with moving averages.
LoRA utilities: Merge multiple LoRAs with per-file weight control, SVD merge with rank truncation, resize/change rank without retraining (SVD or simple truncation), a precision converter (fp16/bf16/fp32), and extract a LoRA from the difference between a base model and a fine-tuned model.
Training: It hooks directly into Kohya as a training manager. Full configuration. You can kick off training runs without leaving the app. Kohya is not included, you point the app to your folder/venv.
Current Status
Still a work in progress but it's been useful for my own training sessions. I feel its ready enough to be shared. Probably a few rough edges, Id like to continue improving it for myself and if others find it useful then all the better. Im open to feedback/collaboration.
Source & Download
Its available in rpm, deb, flatpak, appimage, exe. Requires python.
Repo: Click here
Releases: Click here