---
license: apache-2.0
base_model:
- ByteDance-Seed/BAGEL-7B-MoT
pipeline_tag: any-to-any
library_name: bagel-mot
arxiv: 2507.14119
---


# 🥯 BAGEL-NHR-Edit-V2

<p align="left">
  <a href="https://riko0.github.io/No-Humans-Required/"> 🌐 NHR Website </a> | 
  <a href="https://arxiv.org/abs/2507.14119"> 📜 NHR Paper on arXiv </a> | 
  <a href="https://huggingface.co/datasets/iitolstykh/NHR-Edit"> 🤗 NHR-Edit Dataset (part1) </a> | 
    <a href="https://huggingface.co/datasets/iitolstykh/NHR-Edit-part2"> 🤗 NHR-Edit Dataset (part2) </a> | 
</p>

This repository hosts the model weights for **BAGEL**, fine-tuned on the **[NHR-Edit](https://huggingface.co/datasets/iitolstykh/NHR-Edit) + [NHR-Edit-part2](https://huggingface.co/datasets/iitolstykh/NHR-Edit-part2)** dataset. For installation, usage instructions, and further documentation, please visit the [official BAGEL GitHub repository](https://github.com/bytedance-seed/BAGEL).


### 🛠️ Training Setup

We performed parameter-efficient adaptation on the generation expert’s attention and FFN projection layers using LoRA.

LoRA parameters: 
```
r = 16
lora_alpha = 16
dropout = 0.05
bias = "none"
target_modules = [
  "v_proj_moe_gen",
  "k_proj_moe_gen",
  "mlp_moe_gen.down_proj",
  "mlp_moe_gen.gate_proj",
  "q_proj_moe_gen",
  "mlp_moe_gen.up_proj",
  "o_proj_moe_gen"
]
```

#### Metrics for GEdit-Bench-EN:

| Model              | GEdit-Bench-EN (SC) ↑ | GEdit-Bench-EN (PQ) ↑ | GEdit-Bench-EN (O) ↑  |
| ------------------ | --------------------- | --------------------- | --------------------- |
| BAGEL-7B-MoT       |   7.610 ± 0.150       |   6.180 ± 0.150       |   6.530 ± 0.140       |
| **BAGEL-NHR-Edit-V2** | **7.800 ± 0.070**     | **6.560 ± 0.080**     | **6.800 ± 0.070**     |
> *Scoring model:* `gpt-4.1-2025-04-14` *(with default temperature)*

#### Metrics for ImgEdit-Bench:

| Model              | Style         | Extract       | Remove        | Background    | Action        | Adjust        | Add           | Replace       | Compose       | Overall ↑     |
| ------------------ | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- |
| BAGEL-7B-MoT       | 4.20 ± 0.05   | 1.59 ± 0.10   | **3.16 ± 0.10** | 3.29 ± 0.06   | **3.96 ± 0.17** | **3.51 ± 0.20** | 3.98 ± 0.02   | **3.54 ± 0.11** | 2.93 ± 0.26   | 3.30 ± 0.03   |
| **BAGEL-NHR-Edit-V2** | **4.28 ± 0.04** | **1.65 ± 0.07** | 3.12 ± 0.06   | **3.31 ± 0.02** | 3.81 ± 0.17   | 3.48 ± 0.12   | **4.19 ± 0.03** | 3.51 ± 0.06   | **2.99 ± 0.21** | **3.33 ± 0.02** |
> *Scoring model:* `gpt-4o-2024-11-20` *(with temperature = 0.0)*


### License
**BAGEL-NHR-Edit-V2** is licensed under the Apache 2.0 license. It is finetuned from [ByteDance-Seed/BAGEL-7B-MoT](https://huggingface.co/ByteDance-Seed/BAGEL-7B-MoT), which is also licensed under Apache 2.0.


### ✍️ Citation

```bibtex
@article{Layer2025NoHumansRequired,
    arxivId = {2507.14119},
    author = {Maksim Kuprashevich and Grigorii Alekseenko and Irina Tolstykh and Georgii Fedorov and Bulat Suleimanov and Vladimir Dokholyan and Aleksandr Gordeev},
    title = {{NoHumansRequired: Autonomous High-Quality Image Editing Triplet Mining}},
    year = {2025},
    eprint = {2507.14119},
    archivePrefix = {arXiv},
    primaryClass = {cs.CV},
    url = {https://arxiv.org/abs/2507.14119},
    journal={arXiv preprint arXiv:2507.14119}
}
```