---
license: mit
base_model:
- Shitao/OmniGen-v1
pipeline_tag: text-to-image
tags:
- image-to-image
---


This repo contains bitsandbytes 4bit-NF4 model weights for [OmniGen-v1](https://huggingface.co/Shitao/OmniGen-v1). For info about OmniGen see the [original model card](https://huggingface.co/Shitao/OmniGen-v1).


- 8-bit weights: [gryan/OmniGen-v1-bnb-8bit](https://huggingface.co/gryan/OmniGen-v1-bnb-8bit)
- 4-bit (fp16, nf4) weights: [gryan/OmniGen-v1-fp16-bnb-4bit](https://huggingface.co/gryan/OmniGen-v1-fp16-bnb-4bit) -- for older GPUs (< Ampere/RTX 30xx) / Colab users.


## Usage
Set up your environment by following the original [Quick Start Guide](https://huggingface.co/Shitao/OmniGen-v1#5-quick-start) before getting started.

> [!IMPORTANT]
> NOTE: This feature is not officially supported yet. You'll need to install the repo from [this pull request](https://github.com/VectorSpaceLab/OmniGen/pull/151).

```python
from OmniGen import OmniGenPipeline, OmniGen

# pass the quantized model in the pipeline
model = OmniGen.from_pretrained('gryan/OmniGen-v1-bnb-4bit')
pipe = OmniGenPipeline.from_pretrained("Shitao/OmniGen-v1", model=model)

# proceed as normal!

## Text to Image
images = pipe(
    prompt="A curly-haired man in a red shirt is drinking tea.", 
    height=1024, 
    width=1024, 
    guidance_scale=2.5,
    seed=0,
)
images[0].save("example_t2i.png")  # save output PIL Image

## Multi-modal to Image
# In the prompt, we use the placeholder to represent the image. The image placeholder should be in the format of <img><|image_*|></img>
# You can add multiple images in the input_images. Please ensure that each image has its placeholder. For example, for the list input_images [img1_path, img2_path], the prompt needs to have two placeholders: <img><|image_1|></img>, <img><|image_2|></img>.
images = pipe(
    prompt="A man in a black shirt is reading a book. The man is the right man in <img><|image_1|></img>.",
    input_images=["./imgs/test_cases/two_man.jpg"],
    height=1024, 
    width=1024,
    guidance_scale=2.5, 
    img_guidance_scale=1.6,
    seed=0
)
images[0].save("example_ti2i.png")  # save output PIL image
```

## Image Comparisons
<img src="./assets/text_only_1111_4bit_bf16.png" alt="Text Only Comparison">
<img src="./assets/single_img_1111_4bit_bf16.png" alt="Single Image Comparison">
<img src="./assets/double_img_1111_4bit_bf16.png" alt="Double Image Comparison">

## Performance
For 4bit-nf4 quantized model on RTX 3090 GPU(24G):
| Settings     |  Only Text | Text + Single Image |  Text + Two Images    |
|:-------------|:----------:|:-------------------:|:---------------------:|
| use_kv_cache=False                                              | 6.8G, 1m16s  | 7.2G, 3m30s   | 7.7G, 5m47s   |
| use_kv_cache                                                    | 9.9G, 1m14s  | 20.4G†, 8m5s  | OOM (36.7G†, >1h10m)  |
| use_kv_cache,offload_kv_cache                                   | 6.8G, 1m16s  | 7.2G, 2m49s   | 8.4G, 4m3s  |
| use_kv_cache,offload_kv_cache,separate_cfg_infer                | 6.8G, 1m20s  | 7.0G, 2m31s   | 7.4G, 3m31s  |
| use_kv_cache,offload_kv_cache,offload_model*                    | 5.0G, 1m35s  | 6.0G, 3m7s    | 8.0G, 4m21s  |
| use_kv_cache,offload_kv_cache,separate_cfg_infer,offload_model* | 5.0G, 1m58s  | 5.3G, 3m29s   | 5.6G, 4m19s |

- † - memory_reserved > 24gb, RAM spillover
- \* - only VAE offload. Model loaded in 4bit cannot be offloaded.  

See original [inference settings table](https://github.com/VectorSpaceLab/OmniGen/blob/main/docs/inference.md#requiremented-resources) for bfloat16 performance.