ExpSoft — Products

UGPUM™ (Unified GPU Memory) — pool two consumer GPUs into one usable VRAM space. Currently in test validation.

Train LoRAs on big diffusion models (Wan 2.1 14B, Flux 2 Klein 9B, Flux 2 Dev) that don't fit on a single 16 GB card. UGPUM™ splits the DiT across two GPUs by compute time, not block count — heterogeneous-aware. The certain win today is the additional 16 GB of pooled VRAM; for the full benefit, two PCIe ×16 slots are strongly recommended (PCIe ×4 second slot is slower than DDR5 system memory bandwidth). Two demo datasets bundled for a five-minute first-try. No NCCL, no DDP, no WSL2.

Run the biggest diffusion & video models on a single 16 GB GPU. Currently in test validation.

fp8-on-read streaming + GGUF let one ordinary 16 GB card run models that normally need 24–80 GB — Flux 2 Dev (106 GB), Qwen-Image 20B, Wan 2.1/2.2 video (text-to-video, image-to-video, first-last-frame), HunyuanVideo, plus SD 1.5 / SDXL / Flux / Z-Image / HiDream. The full bf16 weights are never materialized; layers stream onto the GPU one at a time so peak VRAM stays ~4–7 GB. Ships with a reproducible validation grid comparing every model at every quant (bf16 / NF4 / fp8 / GGUF Q4–Q8).

Take real control of Windows Update.

Three-layer neutralizer (Group Policy + Scheduled Tasks + Services) with automatic backup and one-click restore. Tiny native exe, no Python, no GPU. For Windows 11 Pro users tired of forced updates breaking workflows mid-render.

One-click installer & launcher for OneTrainer.

Fine-tune AI models with 28 downloadable base models and 7 pre-configured profiles. Portable, GPU-accelerated, no command line required. Wraps the OneTrainer engine in an installer that handles everything — Python, CUDA, models, configs.

Batch image cropping with interactive zones.

Drop a folder of images, set a square or free rectangle crop, drag to position, export to PNG / JPEG or ZIP. Built for dataset prep and rapid asset extraction.

7-in-1 dataset & media toolkit.

Rename, caption, prompt, resize images. Merge, split videos, extract frames. Portable, FFmpeg-powered. The Swiss army knife of dataset prep.

AI image captioning, no guardrails.

Abliterated QwenVL models + WD14 tagger. Single, batch & tag-forced captioning. 10 prompt presets, 100% local, portable. RTX 40xx / 50xx.

Rip audio CDs with metadata, art, and synced lyrics.

MP3, FLAC, WAV, OGG, M4A output. Auto MusicBrainz metadata, Cover Art Archive, LRCLIB synced lyrics. Portable, no account, no GPU.

Generate full-length songs locally on your GPU.

Text-to-music, REMIX from audio (Cover / Inspire / Extend), LoRA training, offline lyrics. Portable, self-contained.

Speaker diarization & audio splitting.

Powered by pyannote.audio. Drop an audio file, get one track per speaker. GPU or CPU, your choice.

Browse, download, and manage AI models from CivitAI.

Organize your collection, preview galleries, sync downloads, mass-tag, dedupe. The catalog you wish CivitAI had on your local disk.

Personal media archiver — one URL, one click, one playable file.

Paste a single video, a full playlist or a batch list and SaveVideos analyses, queues and downloads each item in best quality (MKV merge via FFmpeg). Three parallel downloads. 1800+ platforms via yt-dlp. Multilingual Legal & Compliance section (EN / FR / ES / DE).

Frame-perfect video chunker — turn raw footage into ML-ready annotated clips.

A focused Windows editor that slices long videos into clean, cropped, labelled chunks for HuggingFace / LeRobot / WebDataset training. Zoomable timeline with waveform, I / O / Ctrl+B keyboard workflow, square ML crop presets (480–2048), FPS retiming with stretch slow-motion, JSON + CSV manifests. License-clean LGPL FFmpeg + LibVLC, fully portable.

One-click NVIDIA RTX Video Super Resolution.

Upscale images and videos 2x or 4x with the NVIDIA Video Effects SDK. Quality presets BICUBIC / BALANCED / HIGH / ULTRA. Two builds: 40xx for RTX 20/30/40 series, 50xx for Blackwell. Portable, self-contained.