Training Krea2 LoRA on 16GB VRAM without block swapping, down to 5-7s/it

Corgice · reddit · 2026-10-19

A Redditor details how to fine-tune a Krea2 LoRA on a single 16GB card without block swapping:

• Model hack: Modified musubi-tuner to train LoRA directly on the int4 convrot quantized checkpoint (no existing tool supported this); code is open-sourced.
• Optimization journey: 13s/it on RTX 4080 Super, a worse 18s/it on RTX 5060 Ti, and mixing int4 convrot + nvfp4 barely helped.
• Breakthrough: Latest ComfyUI with cu130 gave a huge boost — inference 2s/it (no cfg, 1k res) and training down to 9s/it; adding Flash Attention + torch.compile brought training to 5-7s/it.
• Footprint & settings: 12-13GB VRAM, 2GB extra RAM; 1024 resolution, batch size 1, AdamW, grad accum 1. 1200 steps ≈ 1.6h.
• Open issue: A V-shaped loss curve on some datasets — possibly quantization-related (suspected QWT method).

Code: github.com/IceSandwich/musubi-tuner (supportint4convrotkrea2 branch).

Original post →

More from Multimodal

• Researcher: even odds of a famous AI-generated singer or TV show by 2027 — Afinetheorem · 2026-10-19
• BMW i7 TV ad uses a 3D digital twin to seamlessly match the real car — chris_zerolight · 2026-10-19
• Glass Spider-Man bust with lightning nebula half-face wows AI art timeline — egeberkina · 2026-10-19
• 4D Gaussian splatting demo runs in your browser, Chrome ready — janusch_patas · 2026-10-19
• Four favorite Midjourney V6.1 --sref style codes, with working parameters — michaelrabone · 2026-10-19
• Prompt share: a fill-in-the-blank template for quirky gothic cartoon character illustrations — azed_ai · 2026-10-19

Multimodal channel →