LoRA Dataset Studio Adds Qwen-Image 2.1 LoRA Training via AI Toolkit
Ill-Ant-9489 · reddit · 2026-09-23
The developer of LoRA Dataset Studio (LDS), a free self-hosted app for dataset prep and LoRA training, has added Qwen-Image 2.1 training via Ostris' AI Toolkit: pick the model family, keep RGB images and prose captions, then launch locally or via the Cloud Training plugin — LDS handles config, progress, previews and checkpoint downloads.
- Validated end-to-end with a 27-image character dataset over 500 steps (a functionality check, not an ideal recipe); preset uses rank 32, convrot8 quantization and Qwen's shifted flow-matching setup
- To try: update LDS and the Cloud Training plugin, then select Qwen-Image 2.1 in the dataset's Training panel; cloud GPU rental billed separately, local training stays free
- Not yet included: reference-image editing, RGBA datasets, Qwen generation inside LDS Test Studio
Code: github.com/perfectgf/lora-dataset-studio and github.com/ostris/ai-toolkit
More from Multimodal
- Solo creator builds AI action short with MiniMax, Tripo and Blender — shares full workflow — Spoonman915 · 2026-09-23
- Game-to-CGI transition cuts with MiniMax H3 on an RTX 4060 — seed management tips included — Far_Cast_Far_Wide · 2026-09-23
- Pexo's conversational AI video agent churns out product ads for pennies — SimplyAnnisa · 2026-09-23
- Refining MiniMax H3 720p video to 2K on an 8GB GPU: a ComfyUI workflow discussion — ashendonep · 2026-09-23
- Tencent Hunyuan's Hy Image 3.5 preview lands in ComfyUI with +30% human eval win rate — TencentHunyuan · 2026-09-23
- Anyone can now generate videos like these — full prompt shared — umesh_ai · 2026-09-23