README.md

---
base_model: black-forest-labs/FLUX.1-dev
library_name: diffusers
tags:
- lora
- text-to-image
- stable-diffusion
- flux
inference: true
pipeline_tag: text-to-image
license: mit
---

# ART - Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation

- Paper: *[Arxiv](https://arxiv.org/abs/2502.18364)*

- Official Repository: *[ART](https://github.com/microsoft/art-msra)*.

- Project Page: https://art-msra.github.io/

- Demo: [🤗demo](http://20.65.136.27:8060/) 

![teaser](./teaser.png)


## News🔥🔥🔥

* Feb, 26 2025. 💥💥💥 Weights uploaded.

## Abstract

Multi-layer image generation is a fundamental task that enables users to isolate, select, and edit specific image layers, thereby revolutionizing interactions with generative models. In this paper, we introduce the Anonymous Region Transformer (ART), which facilitates the direct generation of variable multi-layer transparent images based on a global text prompt and an anonymous region layout. Inspired by Schema theory suggests that knowledge is organized in frameworks (schemas) that enable people to interpret and learn from new information by linking it to prior knowledge.}, this anonymous region layout allows the generative model to autonomously determine which set of visual tokens should align with which text tokens, which is in contrast to the previously dominant semantic layout for the image generation task. In addition, the layer-wise region crop mechanism, which only selects the visual tokens belonging to each anonymous region, significantly reduces attention computation costs and enables the efficient generation of images with numerous distinct layers (e.g., 50+). When compared to the full attention approach, our method is over 12 times faster and exhibits fewer layer conflicts. Furthermore, we propose a high-quality multi-layer transparent image autoencoder that supports the direct encoding and decoding of the transparency of variable multi-layer images in a joint manner. By enabling precise control and scalable layer generation, ART establishes a new paradigm for interactive content creation.

##  Usage
Please first clone our repository and run the code below to have a try!
```python
import os
import math
import torch
import argparse
from PIL import Image
from multi_layer_gen.custom_model_mmdit import CustomFluxTransformer2DModel
from multi_layer_gen.custom_model_transp_vae import AutoencoderKLTransformerTraining as CustomVAE
from multi_layer_gen.custom_pipeline import CustomFluxPipelineCfg

def test_sample(pipeline, transp_vae, batch, args):

    def adjust_coordinate(value, floor_or_ceil, k=16, min_val=0, max_val=1024):
        # Round the value to the nearest multiple of k
        if floor_or_ceil == "floor":
            rounded_value = math.floor(value / k) * k
        else:
            rounded_value = math.ceil(value / k) * k
        # Clamp the value between min_val and max_val
        return max(min_val, min(rounded_value, max_val))

    validation_prompt = batch["wholecaption"]
    validation_box_raw = batch["layout"]
    validation_box = [
        (
            adjust_coordinate(rect[0], floor_or_ceil="floor"), 
            adjust_coordinate(rect[1], floor_or_ceil="floor"), 
            adjust_coordinate(rect[2], floor_or_ceil="ceil"), 
            adjust_coordinate(rect[3], floor_or_ceil="ceil"), 
        )
        for rect in validation_box_raw
    ]
    if len(validation_box) > 52:
        validation_box = validation_box[:52]
    
    generator = torch.Generator(device=torch.device("cuda", index=args.gpu_id)).manual_seed(args.seed) if args.seed else None
    output, rgba_output, _, _ = pipeline(
        prompt=validation_prompt,
        validation_box=validation_box,
        generator=generator,
        height=args.resolution,
        width=args.resolution,
        num_layers=len(validation_box),
        guidance_scale=args.cfg,
        num_inference_steps=args.steps,
        transparent_decoder=transp_vae,
    )
    images = output.images   # list of PIL, len=layers
    rgba_images = [Image.fromarray(arr, 'RGBA') for arr in rgba_output]

    this_index = batch["index"]
    os.makedirs(os.path.join(args.save_dir, this_index), exist_ok=True)
    for frame_idx, frame_pil in enumerate(images):
        frame_pil.save(os.path.join(args.save_dir, this_index, f"layer_{frame_idx}.png"))
        if frame_idx == 0:
            frame_pil.save(os.path.join(args.save_dir, this_index, "merged.png"))
    merged_pil = images[1].convert('RGBA')
    for frame_idx, frame_pil in enumerate(rgba_images):
        if frame_idx < 2:
            frame_pil = images[frame_idx].convert('RGBA') # merged and background
        else:
            merged_pil = Image.alpha_composite(merged_pil, frame_pil)
        frame_pil.save(os.path.join(args.save_dir, this_index, f"layer_{frame_idx}_rgba.png"))
    
    merged_pil = merged_pil.convert('RGB')
    merged_pil.save(os.path.join(args.save_dir, this_index, "merged_rgba.png"))


args = dict(
    save_dir="output/",
    resolution=512,
    cfg=4.0,
    steps=28,
    seed=41,
    gpu_id=1,
)
args = argparse.Namespace(**args)

transformer = CustomFluxTransformer2DModel.from_pretrained("ART-Release/ART_v1.0", subfolder="transformer", torch_dtype=torch.bfloat16)
transp_vae = CustomVAE.from_pretrained("ART-Release/ART_v1.0", subfolder="transp_vae", torch_dtype=torch.float32)
pipeline = CustomFluxPipelineCfg.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    transformer=transformer,
    torch_dtype=torch.bfloat16,
).to(torch.device("cuda", index=args.gpu_id))
pipeline.enable_model_cpu_offload(gpu_id=args.gpu_id) # Save GPU memory

sample = {
    "index": "case_0",
    "wholecaption": 'Floral wedding invitation: green leaves, white flowers; circular border. Center: "JOIN US CELEBRATING OUR WEDDING" (cursive), "DONNA AND HARPER" (bold), "03 JUNE 2023" (small bold). White, green color scheme, elegant, natural.',
    "layout": [(0, 0, 512, 512), (0, 0, 512, 512), (0, 0, 512, 352), (144, 384, 368, 448), (160, 192, 352, 432), (368, 0, 512, 144), (0, 0, 144, 144), (128, 80, 384, 208), (128, 448, 384, 496), (176, 48, 336, 80)],
}

test_sample(pipeline=pipeline, transp_vae=transp_vae, batch=sample, args=args)

del pipeline
if torch.cuda.is_available():
    torch.cuda.empty_cache()
```


## Citation
```bibtex
@article{pu2025art,
    title={ART: Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation}, 
    author={Yifan Pu and Yiming Zhao and Zhicong Tang and Ruihong Yin and Haoxing Ye and Yuhui Yuan and Dong Chen and Jianmin Bao and Sirui Zhang and Yanbin Wang and Lin Liang and Lijuan Wang and Ji Li and Xiu Li and Zhouhui Lian and Gao Huang and Baining Guo},
    journal={arXiv preprint arXiv:2502.18364},
    year={2025},
}
```