← sources

I trained a Fibo 2.0 LoRA a day after release and put my plush dragon into famous pictures

source: reddit (reddit:StableDiffusion)published 2026-10-21 12:17ingested 2026-10-22 02:40

retained text

Bria released Fibo 2.0 yesterday and I wanted to see how it does with a character LoRA and create something that I can actually post (all my other datasets are not royalty-free).

The dragon is a plush mascot I use for my LoRA portfolio. 15 training images.

What I used:

* Bria's train\_lora.py from the Fibo-2.0-dev repo
* 1024px, rank 64, alpha 64, lr 1e-4 constant, AdamW, batch 1, grad accum 2
* 400 steps, the likeness was already there at 200, and 300 was the best (this might hugely differ for human characters but I did not want to do a synthesized human)
* Rented an A100 40GB on Vast. It peaked at about 23 GB VRAM, ran at 6.2 s/step, and training took about 42 min. Each edit took about 80-90 seconds
* About $2 in GPU time in total

The lora works really well with a plain edit instruction and doesn't change much else

If you add a reference image from the dataset, it can give better results (It drew the outfits like the space suit.) but it often changes the whole picture

On Abbey Road I asked it to add the dragon as a fifth person but it just replaced one of the Beatles (I don't know which one, sorry.) I kept it because I found it funny.

The full settings and VRAM numbers are on my tracker, link in the comments. Fully free, no ads, just my referral links. 

document fetched 2026-10-22 00:34 · canonical https://reddit.com/r/StableDiffusion/comments/1xblkin/i_trained_a_fibo_20_lora_a_day_after_release_and