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 | 🤗HF Demo | Demo | Project Page | 🤗HF Model
The human ability to easily solve multimodal tasks in context (i.e., with only a few demonstrations or simple instructions), is what current multimodal systems have largely struggled to imitate. In this work, we demonstrate that the task-agnostic in-context learning capabilities of large multimodal models can be significantly enhanced by effective scaling-up. 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, such as visual prompting and object-grounded generation. 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. Code and models are publicly available to facilitate future research.
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 | Weight |
|---|---|
| Emu2 | 🤗 HF link |
| Emu2-Chat | 🤗 HF link |
| Emu2-Gen | 🤗 HF link |
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.images[0].save("astronaut.png")
# image editing
image = Image.open("./examples/dog2.jpg").convert("RGB")
prompt = [image, "wearing a rad hat on the beach."]
ret = pipe(prompt)
ret.images[0].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>",
"A photo of",
"<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,
"on the grass",
]
ret = pipe(prompt)
ret.images[0].save("three_dogs.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},
}