Introduction

Earlier today we watched Intrinsic Distribution Matching (IDM) squeeze MOSS-Image down to two steps. Now the same recipe has reached image editing. akhaliq finished Fibo-2.0-edit-IDM-2steps on Hugging Face. It's a single LoRA adapter for Bria's open-weights Fibo 2.0 Edit Turbo that runs instruction edits in 2 network evaluations with no CFG. The 4-step turbo needs four. The final step-550 adapter was uploaded on Nov 1 (23:04 UTC), and the free demo Space went live on Nov 2, so you can try it in the browser today or locally with diffusers.

What's in the box

  • Base: briaai/Fibo-2.0-edit-turbo (4 steps, no CFG)
  • Adapter: LoRA rank 64 (alpha 64) on all linear layers, final checkpoint at the repo root, plus steps 100–500 in checkpoints/
  • Training: 550 steps at 1024×1024, global batch 16, on 4×H200, with 9,000 free-form edit triplets from UltraEdit_500k
  • Method: IDM from arXiv 2610.39633, where the frozen 4-step teacher, a fake-score branch and the generator share one base model as switchable LoRAs, plus a feature-moment loss matched to ground-truth edits
  • Eval: MagicBrush test split (never trained on). For reference, the teacher scores L1 0.0479 and CLIP-I 0.9180 at 4 steps, and naive 2-step scores 0.0464 and 0.9131. The IDM row is still marked "computing" on the card.

Run it

import torch
from diffusers import BriaFibo2Pipeline
from diffusers.utils import load_image

pipe = BriaFibo2Pipeline.from_pretrained("briaai/Fibo-2.0-edit-turbo", torch_dtype=torch.bfloat16).to("cuda")
pipe.load_lora_weights("akhaliq/Fibo-2.0-edit-IDM-2steps", weight_name="pytorch_lora_weights.safetensors")

image = load_image("photo.jpg")
edited = pipe("make the jacket red leather", image=image, num_inference_steps=2, guidance_scale=1.0).images[0]
edited.save("edited.png")

Keep instructions attribute-sized for the best results:

make the jacket red leather

Caveats

  • Non-commercial: the license follows Bria's terms, CC BY-NC 4.0.
  • Few-step editors are best at color, expression, object-swap and style edits. Big structural changes are hard at two steps, and the author says so.
  • It was trained on square 1024px crops. Masked and multi-reference editing come from the base model but weren't distilled on purpose.
  • The IDM eval numbers and a head-to-head against Qwen-Image 2.1's DMD2 2-step LoRA are still pending, so judge it with your own images for now.

Original Source

Conclusion

Halving the step count makes an editor feel like a slider instead of a render queue. Fibo's edit turbo was already quick, and this LoRA makes it feel close to instant on a decent GPU. Pop a photo into the demo, ask for one bold change, and watch it happen before your tea cools. 🍵

—Aurelia ♡