AcademiaSD adds Fibo 2.0 LoRA training with NF4 quantization, cutting VRAM from 16GB to 5.2GB
AcademiaSD · reddit · 2026-10-21
AcademiaSD's open-source LoRAlab Trainer Studio now trains LoRAs (and edit LoRAs) for BRIA's Fibo 2.0 Dev. NF4 quantization on attention/MLP/cross-attention of the 44 blocks cuts VRAM from 16GB to 5.2GB; the perceiver, modulation and I/O layers stay BF16. It uses Fibo 2.0's own structured JSON (VGL) caption format, offers 30-step/CFG5 previews and one-click ComfyUI export. Benchmarks on an RTX 5080 16GB (rank 8, all targets): 512² at 2.0 s/step (8.7GB), 1024² at 6.2 s/step (11.2GB); 1024² edit training needs 16GB. Still experimental, and Fibo 2.0 Dev is CC BY-NC 4.0 (non-commercial). Also newly supported: Ming-Image, FLUX.3 Image Klein, Z-Image, Qwen-Image 2512/Edit 2511 and more.
More from Multimodal
- Cosmic horror concept video made with Seedance 2.5 via DaVinci AI — Terry__Poppins · 2026-10-21
- Stunning AI-generated mech video wows Reddit viewers — Raykaix · 2026-10-21
- Screenwriter previzzes a screenplay scene with AI: plane, kid, 1950s air base — EnthusiasticNtrovert · 2026-10-21
- Kling 4 goes live on Runway with unlimited generation and 25% discount — aziz4ai · 2026-10-21
- Ideogram teases infinitely scalable designs, launching 9am PT tomorrow — aziz4ai · 2026-10-21
- Dev open-sources ComfyUI movie studio that turns one prompt into a fully editable film — Goat_bless · 2026-10-21