LoRA Training for Fibo 2.0 Added
retained text
The LoRAlab Trainer studio is a web app that trains LoRAs for a lot of models
(Qwen-Image, Krea, Z-Image, Anima, SDXL, LTX, MiniMax-H3 and so on) with
an automatic captioner, live previews, step-accurate resume and one-click
export to ComfyUI. Fibo 2.0 Dev joins it, with normal LoRAs and edit LoRAs
(before/after pairs).
GitHub:
https://github.com/AcademiaSD/AcademiaSD_LoRAlab-TrainerStudio
What you get with Fibo 2.0:
- NF4 on the attention, MLP and cross-attention of the 44 blocks, so the
transformer is 5.2 GB in VRAM instead of 16
- The pre-cache runs the model's own Qwen3-VL-4B once, NF4 by default.
Training never loads it
- Captions in the structured JSON (VGL) that Fibo 2.0 was trained with,
text rendering included
- BRIA's own recipe, and LoRA targets on all 415 layers by default (blocks
only is there too)
- Previews at 30 steps and CFG 5, plus a preview size option (768 or 1024)
to see them at the size you will use in ComfyUI
- The LoRA is exported in ComfyUI's format, fused layers included, so it
loads as is
On an RTX 5080 16 GB, rank 8, all targets:
- image at 512x512: 2.0 s/step, 8.7 GB
- image at 1024x1024: 6.2 s/step, 11.2 GB
- edit at 512x512: 3.1 s/step, 9.9 GB
- edit at 1024x1024: 14.9 s/step, 15.9 GB
Edit LoRAs are best trained at 512x512, at 1024x1024 you need 16 GB.
The NF4 build downloads the first time you use it:
https://huggingface.co/AcademiaSD/Fibo-2.0-NF4-for-LoRA-Training
It is still experimental. Pre-cache, training, previews, edit mode and
resume are all verified on Fibo 2.0 Dev, but it is young and the defaults
will probably move. And worth saying: Fibo 2.0 Dev is CC BY-NC 4.0, so
non-commercial use only.
While you are at it, the studio also got Ming-Image (transparent designs),
Boogu-Image-Edit 0.1, ERNIE-Image, FLUX.3 Image Klein, Z-Image, Anima,
MOSS-Image and Qwen-Image 2512 / Edit 2511.
document fetched 2026-10-22 00:34 · canonical https://reddit.com/r/StableDiffusion/comments/1xboo2u/lora_training_for_fibo_20_added