Free open-source LoRA trainer fits 9 models including MiniMax-H3 into 4GB consumer GPUs
AcademiaSD · reddit · 2026-10-04
A Reddit user released AcademiaSD LoRAlab Trainer Studio, a free all-in-one LoRA training suite for consumer GPUs (one-click Windows installer, brand-new Linux support). Highlights:
- 9 trainers, one web UI: Qwen-Image 2.1, FLUX.2 Klein 9B, Krea 2, Z-Image, Ideogram 4, Anima, SDXL/Pony/Illustrious, LTX 2.3 video, and MiniMax-H3 (video + audio), with minimum VRAM from 4GB to 12GB.
- Extreme VRAM savings: everything loads in 4-bit NF4, with text encoder and VAE run once in a pre-cache stage; SDXL trains in 3.5GB, so 4GB laptops can train SDXL/Pony/Illustrious LoRAs.
- MiniMax-H3: the official 33B checkpoint is 500GB; the trainer uses a 41GB NF4 version with block swap to fit 8GB VRAM. Supports mixed image/video/audio datasets, plus RefMods — encode reference images/clips (with audio) for ComfyUI's reference-to-video node in seconds, no training needed.
- Extras: built-in Qwen3-VL auto-captioning (natural language, Danbooru tags, or JSON with bounding boxes), live training previews, exact-step resume, and one-click export to ComfyUI/Forge.
More from coding & agent
- Letting Claude prompt Midjourney: machines telling machines how to make pictures — technollama · 2026-10-04
- Why Agents Don't Need GUIs: Shell, Pipes and Zellij Beat Any App — Liu_eroteme · 2026-10-04
- MCP Guardrails: Gate Tool Calls at the Method Layer With Agentgateway's ExtMCP — bibryam · 2026-10-04
- Scoble's vacation AI agent reads X's AI community and declares 'The Harness Is the Product Now' — Scobleizer · 2026-10-04
- Sandboxing AI agents with bubblewrap and tmpfs so they can't touch your secrets — Liu_eroteme · 2026-10-04
- Instinct and Muse become the first AI agents a daily user actually relies on — nateliason · 2026-10-04