akhaliq/Fibo-2.0-edit-IDM-1step

Fibo 2.0 Edit — IDM 1-Step

Bria Fibo 2.0 Edit Turbo image editing in a single network evaluation, no CFG — one LoRA adapter. Community implementation of Intrinsic Distribution Matching (IDM) (Liu et al., arXiv:2610.39633), compressing the 4-step distilled editing model to 1 step.

	DiT forwards / edit
Fibo 2.0 Edit Turbo (4 steps, no CFG)	4
This adapter (1 step, no CFG)	1

Method

• Internal distribution matching — real score from the frozen 4-step teacher (adapters disabled), fake score from a LoRA branch trained with flow matching on detached generator outputs; DMD direction sg[clamp(v_real - v_fake, ±1)] as in APEX.
• Feature moment loss (L_pop, λ = 0.5) — mean and per-coordinate std matching in the generator-EMA feature space (mean-pooled last-block hidden states at high noise), two FIFO queues K = 16384. Reference: ground-truth edited images.
• Generator: strict 1-NFE map x = z − v_gen(z, t=1) throughout training.

Recipe: LoRA r=64 α=64 all-linear ×3 branches (gen/fake/ema) on one frozen base, AdamW lr 1e-4, bf16, grad checkpointing, 1024×1024, batch 8, 550 steps on 1×H200. EMA decay 0.99. Zero-LoRA init on briaai/Fibo-2.0-edit-turbo. Moment-loss features are extracted at 512px (consistent on generator and reference sides) to fit single-GPU memory.

Data: 9,000 editing triplets from UltraEdit_500k (FreeForm, clip_sim_dir ≥ 0.1). Eval: MagicBrush test split, first-turn rows (never trained on). The naive_1step/ folder keeps the pre-training naive 1-step probe of the teacher for comparison.

Usage

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-1step", weight_name="pytorch_lora_weights.safetensors")

image = load_image("photo.jpg")
edited = pipe("make the jacket red leather", image=image,
              num_inference_steps=1, guidance_scale=1.0).images[0]
edited.save("edited.png")

Eval results

MagicBrush test, 128 first-turn rows at 1 NFE, unshifted t=1→0 Euler for both rows (L1 ↓, CLIP-I ↑, CLIP-T ↑ vs ground-truth target; CLIP-T = output vs instruction, ViT-B/32):

Model	Steps	L1 ↓	CLIP-I ↑	CLIP-T ↑
Fibo 2.0 Edit Turbo, naive	1	0.2198	0.5430	0.2080
IDM 1-step (this adapter)	1	0.2624	0.5312	0.2002

Reference from the sibling 2-step repo (different protocol, 2 NFE): teacher 4-step L1 0.0479 / CLIP-I 0.9180 / CLIP-T 0.2066.

Limitations

• 1-step editing favors attribute-level edits (color, expression, object swap, style); structural edits and in-image text remain hardest at 1 NFE.
• Trained on square 1024px crops of UltraEdit free-form edits; masked and multi-reference editing are inherited from the base model but were not distilled explicitly.
• License follows the base model: CC BY-NC 4.0 (non-commercial), per Bria's terms.

Citation

@misc{liu2026intrinsicdistributionmatching,
  title={Intrinsic Distribution Matching},
  author={Deyuan Liu and Yansen Han and Kaiwen Zheng and Jun Xie and Jiacheng Liu and Zhenglin Cheng and Peng Sun and Jintao Zhang and Jianguo Li and Tao Lin},
  year={2026}, eprint={2610.39633}, archivePrefix={arXiv}, primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.39633},
}

Base model: Fibo 2.0 (Bria AI). Method grounding: DMD2, APEX/TwinFlow internal-score construction.