Quan Sun1*, Yufeng Cui1*, Xiaosong Zhang1*, Fan Zhang1*, Qiying Yu2,1*, Zhengxiong Luo1, Yueze Wang1, Yongming Rao1,
Jingjing Liu2, Tiejun Huang1,3, Xinlong Wang1†
1 BAAI, 2 THU, 3 PKU
* equal contribution † project lead
| Paper | Fast Demo | 🤗HF Demo | 🤗HF Models | Project Page | Video Demo |
We introduce Emu2, a generative multimodal model with 37 billion parameters, trained on large-scale multimodal sequences with a unified autoregressive objective. Emu2 exhibits strong multimodal in-context learning abilities, even emerging to solve tasks that require on-the-fly reasoning. The model sets a new record on multiple multimodal understanding tasks in few-shot settings. When instruction-tuned to follow specific instructions, Emu2 further achieves new state-of-the-art on challenging tasks such as question answering benchmarks for large multimodal models and open-ended subject-driven generation. These achievements demonstrate that Emu2 can serve as a base model and general-purpose interface for a wide range of multimodal tasks.
Zero-shot subject-driven generation
We supplemented the results of Refer Expression Comprehension on RefCOCO, RefCOCO+, RefCOCOg, and compared them with generalist models.
| Model | RefCOCO val |
RefCOCO testA |
RefCOCO testB |
RefCOCO+ val |
RefCOCO+ testA |
RefCOCO+ testB |
RefCOCOg val |
RefCOCOg test |
|---|---|---|---|---|---|---|---|---|
| OFA-L | 79.96 | 83.67 | 76.39 | 68.29 | 76.00 | 61.75 | 67.57 | 67.58 |
| Shikra-13B | 87.83 | 91.11 | 81.81 | 82.89 | 87.79 | 74.41 | 82.64 | 83.16 |
| Qwen-VL-7B | 89.36 | 92.26 | 85.34 | 83.12 | 88.25 | 77.21 | 85.58 | 85.48 |
| Emu2-Chat | 90.40 | 93.88 | 85.97 | 87.05 | 91.43 | 80.47 | 87.64 | 88.11 |
Clone this repository and install required packages:
git clone https://github.com/baaivision/Emu
cd Emu/Emu2
pip install -r requirements.txt| Model name | HF Weight |
|---|---|
| Emu2 | 🤗 HF link |
| Emu2-Chat | 🤗 HF link |
| Emu2-Gen | 🤗 HF link |
- Model type: An auto-regressive multimodal model based on the transformer architecture.
- License: Non-commercial license
- Initialized from model: LLaMA.
You can also access the native PyTorch models from Link
- Emu2 and Emu2-Chat can be loaded using
emu.chat.EmuChatGeneration - Emu2-Gen can be loaded using
emu.diffusion.EmuVisualGeneration
To facilitate the local usage, we provide the demo codes for both huggingface version and native PyTorch version models.
# Before using the HF demo tools, please download Emu2-Chat or Emu2-Gen from
# Emu2-Chat: https://huggingface.co/BAAI/Emu2-Chat
# Emu2-Gen: https://huggingface.co/BAAI/Emu2-Gen
# launch the frontend
cd demo/frontend
python frontend.py
# launch the backend
cd demo/backend/hf_model
# huggingface version of demo supports both multi-GPU deployment and quantization
# launch both generation and chat demo
python backend.py --model-path ${PATH_TO_Emu2-Chat_and_Emu2-Gen}
# launch chat only demo
python backend.py --disable-generate --model-path ${PATH_TO_Emu2-Chat}
# with N gpus
python backend.py --disable-generate --chat-gpu-per-instance ${N} --model-path ${PATH_TO_Emu2-Chat}
# quantize
python backend.py --disable-generate --chat-quantize --model-path ${PATH_TO_Emu2-Chat}
# launch generation only demo
python backend.py --disable-chat --model-path ${PATH_TO_Emu2-Gen}
# with N gpus
python backend.py --disable-chat --generate-gpu-per-instance ${N} --model-path ${PATH_TO_Emu2-Gen}
# quantize
python backend.py --disable-chat --generate-quantize --model-path ${PATH_TO_Emu2-Gen}# Before using the native PyTorch demo tools, please download Emu2-Chat or Emu2-Gen from
# https://model.baai.ac.cn/model-detail/220122
# launch the frontend
cd demo/frontend
python frontend.py
# launch the backend
cd demo/backend/pytorch_model
# native PyTorch version of demo only supports a simple multi-GPU deployment strategy
# launch both generation and chat demo
MODEL_PATH="DIR_CONTAINS_Emu2-Gen_pytorch_model_AND_Emu2-Chat_pytorch_model"
python backend.py --model-path ${MODEL_PATH}
# launch chat only demo
python backend.py --disable-generate --model-path ${DIR_CONTAINS_Emu2-Chat_pytorch_model}
# with N gpus
python backend.py --disable-generate --chat-gpu-per-instance ${N} --model-path ${DIR_CONTAINS_Emu2-Chat_pytorch_model}
# launch generation only demo
python backend.py --disable-chat --model-path ${DIR_CONTAINS_Emu2-Gen_pytorch_model}
# with N gpus
python backend.py --disable-chat --generate-gpu-per-instance ${N} --model-path ${DIR_CONTAINS_Emu2-Gen_pytorch_model}from PIL import Image
import requests
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("BAAI/Emu2") # "BAAI/Emu2-Chat"
model = AutoModelForCausalLM.from_pretrained(
"BAAI/Emu2", # "BAAI/Emu2-Chat"
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
trust_remote_code=True).to('cuda').eval()
# `[<IMG_PLH>]` is the image placeholder which will be replaced by image embeddings.
# the number of `[<IMG_PLH>]` should be equal to the number of input images
query = '[<IMG_PLH>]Describe the image in details:'
#image = Image.open(requests.get('https://github.com/baaivision/Emu/Emu2/examples/blue_black_1_top_left.jpg?raw=true',stream=True).raw).convert('RGB')
image = Image.open("./examples/blue_black_1_top_left.jpg").convert('RGB')
inputs = model.build_input_ids(
text=[query],
tokenizer=tokenizer,
image=[image]
)
with torch.no_grad():
outputs = model.generate(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
image=inputs["image"].to(torch.bfloat16),
max_new_tokens=64,
length_penalty=-1)
output_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)Interleaved image and text
from PIL import Image
import requests
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("BAAI/Emu2") # "BAAI/Emu2-Chat"
model = AutoModelForCausalLM.from_pretrained(
"BAAI/Emu2", # "BAAI/Emu2-Chat"
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
trust_remote_code=True).to('cuda').eval()
# `[<IMG_PLH>]` is the image placeholder which will be replaced by image embeddings.
# the number of `[<IMG_PLH>]` should be equal to the number of input images
query = "[<IMG_PLH>][red, white, 3, bottom left].[<IMG_PLH>][yellow, white, 2, top left].[<IMG_PLH>][green, black, 4, bottom right].[<IMG_PLH>]"
images = [
Image.open("./examples/red_white_3_bottom_left.jpg").convert('RGB'),
Image.open("./examples/yellow_white_2_top_right.jpg").convert('RGB'),
Image.open("./examples/green_black_4_bottom_right.jpg").convert('RGB'),
Image.open("./examples/blue_black_1_top_left.jpg").convert('RGB')
]
inputs = model.build_input_ids(
text=[query],
tokenizer=tokenizer,
image=images
)
with torch.no_grad():
outputs = model.generate(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
image=inputs["image"].to(torch.bfloat16),
max_new_tokens=64,
length_penalty=-1)
output_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)from PIL import Image
import requests
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from accelerate import init_empty_weights, infer_auto_device_map, load_checkpoint_and_dispatch
tokenizer = AutoTokenizer.from_pretrained("BAAI/Emu2") # "BAAI/Emu2-Chat"
with init_empty_weights():
model = AutoModelForCausalLM.from_pretrained(
"BAAI/Emu2", # "BAAI/Emu2-Chat"
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
trust_remote_code=True)
device_map = infer_auto_device_map(model, max_memory={0:'38GiB',1:'38GiB',}, no_split_module_classes=['Block','LlamaDecoderLayer'])
# input and output logits should be on same device
device_map["model.decoder.lm.lm_head"] = 0
model = load_checkpoint_and_dispatch(
model,
'local/path/to/hf/version/Emu2/model',
device_map=device_map).eval()
# `[<IMG_PLH>]` is the image placeholder which will be replaced by image embeddings.
# the number of `[<IMG_PLH>]` should be equal to the number of input images
query = '[<IMG_PLH>]Describe the image in details:'
image = Image.open("./examples/blue_black_1_top_left.jpg").convert('RGB')
inputs = model.build_input_ids(
text=[query],
tokenizer=tokenizer,
image=[image]
)
with torch.no_grad():
outputs = model.generate(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
image=inputs["image"].to(torch.bfloat16),
max_new_tokens=64,
length_penalty=-1)
output_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)Interleaved image and text
from PIL import Image
import requests
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from accelerate import init_empty_weights, infer_auto_device_map, load_checkpoint_and_dispatch
tokenizer = AutoTokenizer.from_pretrained("BAAI/Emu2") # "BAAI/Emu2-Chat"
with init_empty_weights():
model = AutoModelForCausalLM.from_pretrained(
"BAAI/Emu2", # "BAAI/Emu2-Chat"
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
trust_remote_code=True)
device_map = infer_auto_device_map(model, max_memory={0:'38GiB',1:'38GiB',}, no_split_module_classes=['Block','LlamaDecoderLayer'])
# input and output logits should be on same device
device_map["model.decoder.lm.lm_head"] = 0
model = load_checkpoint_and_dispatch(
model,
'local/path/to/hf/version/Emu2/model',
device_map=device_map).eval()
# `[<IMG_PLH>]` is the image placeholder which will be replaced by image embeddings.
# the number of `[<IMG_PLH>]` should be equal to the number of input images
query = "[<IMG_PLH>][red, white, 3, bottom left].[<IMG_PLH>][yellow, white, 2, top left].[<IMG_PLH>][green, black, 4, bottom right].[<IMG_PLH>]"
images = [
Image.open("./examples/red_white_3_bottom_left.jpg").convert('RGB'),
Image.open("./examples/yellow_white_2_top_right.jpg").convert('RGB'),
Image.open("./examples/green_black_4_bottom_right.jpg").convert('RGB'),
Image.open("./examples/blue_black_1_top_left.jpg").convert('RGB')
]
inputs = model.build_input_ids(
text=[query],
tokenizer=tokenizer,
image=images
)
with torch.no_grad():
outputs = model.generate(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
image=inputs["image"].to(torch.bfloat16),
max_new_tokens=64,
length_penalty=-1)
output_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)Check quantization guidance at transformers
from PIL import Image
import requests
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("BAAI/Emu2") # "BAAI/Emu2-Chat"
model = AutoModelForCausalLM.from_pretrained(
"BAAI/Emu2", # "BAAI/Emu2-Chat"
load_in_4bit=True,
trust_remote_code=True,
bnb_4bit_compute_dtype=torch.float16).eval()
query = '[<IMG_PLH>]Describe the image in details:'
image = Image.open("./examples/blue_black_1_top_left.jpg").convert('RGB')
inputs = model.build_input_ids(
text=[query],
tokenizer=tokenizer,
image=[image]
)
with torch.no_grad():
outputs = model.generate(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
image=inputs["image"].to(torch.float16), # should be torch.float16
max_new_tokens=64,
length_penalty=-1)
output_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)import cv2
from diffusers import DiffusionPipeline
import numpy as np
from PIL import Image
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
# For the first time of using,
# you need to download the huggingface repo "BAAI/Emu2-GEN" to local first
path = "path to local BAAI/Emu2-GEN"
multimodal_encoder = AutoModelForCausalLM.from_pretrained(
f"{path}/multimodal_encoder",
trust_remote_code=True,
torch_dtype=torch.bfloat16,
use_safetensors=True,
variant="bf16"
)
tokenizer = AutoTokenizer.from_pretrained(f"{path}/tokenizer")
pipe = DiffusionPipeline.from_pretrained(
path,
custom_pipeline="pipeline_emu2_gen",
torch_dtype=torch.bfloat16,
use_safetensors=True,
variant="bf16",
multimodal_encoder=multimodal_encoder,
tokenizer=tokenizer,
)
# For the non-first time of using, you can init the pipeline directly
pipe = DiffusionPipeline.from_pretrained(
path,
custom_pipeline="pipeline_emu2_gen",
torch_dtype=torch.bfloat16,
use_safetensors=True,
variant="bf16",
)
pipe.to("cuda")
# text-to-image
prompt = "impressionist painting of an astronaut in a jungle"
ret = pipe(prompt)
ret.image.save("astronaut.png")
# image editing
image = Image.open("./examples/dog.jpg").convert("RGB")
prompt = [image, "wearing a red hat on the beach."]
ret = pipe(prompt)
ret.image.save("dog_hat_beach.png")
# grounding generation
def draw_box(left, top, right, bottom):
mask = np.zeros((448, 448, 3), dtype=np.uint8)
mask = cv2.rectangle(mask, (left, top), (right, bottom), (255, 255, 255), 3)
mask = Image.fromarray(mask)
return mask
dog1 = Image.open("./examples/dog1.jpg").convert("RGB")
dog2 = Image.open("./examples/dog2.jpg").convert("RGB")
dog3 = Image.open("./examples/dog3.jpg").convert("RGB")
dog1_mask = draw_box( 22, 14, 224, 224)
dog2_mask = draw_box(224, 10, 448, 224)
dog3_mask = draw_box(120, 264, 320, 438)
prompt = [
"<grounding>",
"An oil painting of three dogs,",
"<phrase>the first dog</phrase>"
"<object>",
dog1_mask,
"</object>",
dog1,
"<phrase>the second dog</phrase>"
"<object>",
dog2_mask,
"</object>",
dog2,
"<phrase>the third dog</phrase>"
"<object>",
dog3_mask,
"</object>",
dog3,
]
ret = pipe(prompt)
ret.image.save("three_dogs.png")
# Autoencoding
# to enable the autoencoding mode, please pull the latest pipeline_emu2_gen.py first,
# and you can only input exactly one image as prompt
# if you want the model to generate an image,
# please input extra empty text "" besides the image, e.g.
# autoencoding mode: prompt = image or [image]
# generation mode: prompt = ["", image] or [image, ""]
prompt = Image.open("./examples/doodle.jpg").convert("RGB")
ret = pipe(prompt)
ret.image.save("doodle_ae.png")from emu.chat import EmuChatGeneration
# Emu2
pipe = EmuChatGeneration.from_pretrained(
"Path to Emu2_pytorch_model.bf16.bin",
dtype=torch.bfloat16,
)
# Emu2-Chat
pipe = EmuChatGeneration.from_pretrained(
"Path to Emu2-Chat_pytorch_model.bf16.bin",
instruct=True,
dtype=torch.bfloat16,
)
# Single GPU, e.g. cuda:0
pipe = pipe.multito(["cuda:0"])
# Multi GPU, e.g. cuda:0 and cuda:1
pipe = pipe.multito(["cuda:0", "cuda:1"])
# In the context of chat, input must be List[List[str | Image.Image]]
# The length of outer list must be odd, which represents the ROLEs, like [USER, ASSISTANT, USER, ...],
# The first and the last ROLE must denote USER.
# The content in the innter list is the input/output of USER/ASSISTANT
kwargs = {
"do_sample": False,
"max_new_tokens": 1024,
"temperature": 0.7,
"top_k": 3,
"top_p": 0.9,
"length_penalty": 2.0,
"num_beams": 5,
"repetition_penalty": 1.0,
}
# image caption case
image = Image.open("./examples/blue_black_1_top_left.jpg").convert('RGB')
user_input = [image, "describe the image in details:"]
output = pipe([user_input], **kwargs)
# video case
# suppose that you have already extracted the video frames
frames: List[Image.Image] = []
user_input = ["[VIDEO]"] + frames + ["[/VIDEO]", "What happened in the video?"]
output = pipe([user_input], **kwargs)
# Emu2-Chat also supports generation under the grounding senario
# to enable the grounding generation, please pass
# is_grounding=True and skip_special_tokens=False
# to the generation function
# NOTE: Our grounding coordinates range from 0 to 224
image = Image.open("./examples/squirrel.jpeg").convert("RGB")
user_input = [image, "Where is the squirrel?"]
output = pipe([user_input], is_grounding=True, skip_special_tokens=False, **kwargs)
# In the context of non-chat, input must be List[str | Image.Image]
# which represents the input content
# interleaved case
img1 = Image.open("./examples/red_white_3_bottom_left.jpg").convert('RGB')
img2 = Image.open("./examples/yellow_white_2_top_right.jpg").convert('RGB')
img3 = Image.open("./examples/green_black_4_bottom_right.jpg").convert('RGB')
img4 = Image.open("./examples/blue_black_1_top_left.jpg").convert('RGB')
input = [
img1, "[red, white, 3, bottom left].",
img2, "[yellow, white, 2, top left].",
img3, "[green, black, 4, bottom right].",
img4,
]
output = pipe(input, max_new_tokens=15)from emu.diffusion import EmuVisualGeneration
pipe = EmuVisualGeneration.from_pretrained(
"Path to Emu2-Gen_pytorch_model.bf16.safetensors",
dtype=torch.bfloat16,
use_safetensors=True,
)
# Single GPU, e.g. cuda:0
pipe = pipe.multito(["cuda:0"])
# Multi GPU, e.g. cuda:0 and cuda:1
pipe = pipe.multito(["cuda:0", "cuda:1"])
# text-to-image
prompt = "impressionist painting of an astronaut in a jungle"
ret = pipe(prompt)
ret.image.save("astronaut.png")
# image editing
image = Image.open("./examples/dog.jpg").convert("RGB")
prompt = [image, "wearing a red hat on the beach."]
ret = pipe(prompt)
ret.image.save("dog_hat_beach.png")
# grounding generation
def draw_box(left, top, right, bottom):
mask = np.zeros((448, 448, 3), dtype=np.uint8)
mask = cv2.rectangle(mask, (left, top), (right, bottom), (255, 255, 255), 3)
mask = Image.fromarray(mask)
return mask
dog1 = Image.open("./examples/dog1.jpg").convert("RGB")
dog2 = Image.open("./examples/dog2.jpg").convert("RGB")
dog3 = Image.open("./examples/dog3.jpg").convert("RGB")
dog1_mask = draw_box( 22, 14, 224, 224)
dog2_mask = draw_box(224, 10, 448, 224)
dog3_mask = draw_box(120, 264, 320, 438)
prompt = [
"<grounding>",
"An oil painting of three dogs,",
"<phrase>the first dog</phrase>"
"<object>",
dog1_mask,
"</object>",
dog1,
"<phrase>the second dog</phrase>"
"<object>",
dog2_mask,
"</object>",
dog2,
"<phrase>the third dog</phrase>"
"<object>",
dog3_mask,
"</object>",
dog3,
]
ret = pipe(prompt)
ret.image.save("three_dogs.png")
# Autoencoding
# to enable the autoencoding mode, you can only input exactly one image as prompt
# if you want the model to generate an image,
# please input extra empty text "" besides the image, e.g.
# autoencoding mode: prompt = image or [image]
# generation mode: prompt = ["", image] or [image, ""]
prompt = Image.open("./examples/doodle.jpg").convert("RGB")
ret = pipe(prompt)
ret.image.save("doodle_ae.png")We thank the great work from LLaMA, BLIP-2, Stable Diffusion, and FastChat.
If you find Emu useful for your research and applications, please consider starring this repository and citing:
@article{Emu2,
title={Generative Multimodal Models are In-Context Learners},
author={Quan Sun and Yufeng Cui and Xiaosong Zhang and Fan Zhang and Qiying Yu and Zhengxiong Luo and Yueze Wang and Yongming Rao and Jingjing Liu and Tiejun Huang and Xinlong Wang},
publisher={arXiv preprint arXiv:2312.13286},
year={2023},
}