Views
No views yet
1# clone our repo
2git clone https://github.com/wendell0218/FocusDiff.git
3cd FocusDiff
4
5# prepare python environment
6conda create -n focus-diff python=3.10
7conda activate focus-diff
8pip install -r requirements.txtJanus-Pro-7B as the pretrained model for subsequent supervised fine-tuning. You can download the corresponding model using the following command:1GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/deepseek-ai/Janus-Pro-7B
2cd Janus-Pro-7B
3git lfs pull1import os
2import torch
3import PIL.Image
4import numpy as np
5from transformers import AutoModelForCausalLM
6from janus.models import MultiModalityCausalLM, VLChatProcessor
7
8@torch.inference_mode()
9def generate(
10 mmgpt: MultiModalityCausalLM,
11 vl_chat_processor: VLChatProcessor,
12 prompt: str,
13 temperature: float = 1.0,
14 parallel_size: int = 4,
15 cfg_weight: float = 5.0,
16 image_token_num_per_image: int = 576,
17 img_size: int = 384,
18 patch_size: int = 16,
19 img_top_k: int = 1,
20 img_top_p: float = 1.0,
21):
22 images = []
23 input_ids = vl_chat_processor.tokenizer.encode(prompt)
24 input_ids = torch.LongTensor(input_ids)
25 tokens = torch.zeros((parallel_size*2, len(input_ids)), dtype=torch.int).cuda()
26 for i in range(parallel_size*2):
27 tokens[i, :] = input_ids
28 if i % 2 != 0:
29 tokens[i, 1:-1] = vl_chat_processor.pad_id
30 inputs_embeds = mmgpt.language_model.get_input_embeddings()(tokens)
31 generated_tokens = torch.zeros((parallel_size, image_token_num_per_image), dtype=torch.int).cuda()
32 for i in range(image_token_num_per_image):
33 outputs = mmgpt.language_model.model(inputs_embeds=inputs_embeds, use_cache=True, past_key_values=outputs.past_key_values if i != 0 else None)
34 hidden_states = outputs.last_hidden_state
35 logits = mmgpt.gen_head(hidden_states[:, -1, :])
36 logit_cond = logits[0::2, :]
37 logit_uncond = logits[1::2, :]
38 logits = logit_uncond + cfg_weight * (logit_cond-logit_uncond)
39 if img_top_k:
40 v, _ = torch.topk(logits, min(img_top_k, logits.size(-1)))
41 logits[logits < v[:, [-1]]] = float("-inf")
42 probs = torch.softmax(logits / temperature, dim=-1)
43 if img_top_p:
44 probs_sort, probs_idx = torch.sort(probs,
45 dim=-1,
46 descending=True)
47 probs_sum = torch.cumsum(probs_sort, dim=-1)
48 mask = probs_sum - probs_sort > img_top_p
49 probs_sort[mask] = 0.0
50 probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True))
51 next_token = torch.multinomial(probs_sort, num_samples=1)
52 next_token = torch.gather(probs_idx, -1, next_token)
53 else:
54 next_token = torch.multinomial(probs, num_samples=1)
55 generated_tokens[:, i] = next_token.squeeze(dim=-1)
56 next_token = torch.cat([next_token.unsqueeze(dim=1), next_token.unsqueeze(dim=1)], dim=1).view(-1)
57 img_embeds = mmgpt.prepare_gen_img_embeds(next_token)
58 inputs_embeds = img_embeds.unsqueeze(dim=1)
59 dec = mmgpt.gen_vision_model.decode_code(generated_tokens.to(dtype=torch.int), shape=[parallel_size, 8, img_size//patch_size, img_size//patch_size])
60 dec = dec.to(torch.float32).cpu().numpy().transpose(0, 2, 3, 1)
61 dec = np.clip((dec + 1) / 2 * 255, 0, 255)
62 visual_img = np.zeros((parallel_size, img_size, img_size, 3), dtype=np.uint8)
63 visual_img[:, :, :] = dec
64 for i in range(parallel_size):
65 images.append(PIL.Image.fromarray(visual_img[i]))
66
67 return images
68
69
70if __name__ == "__main__":
71 import argparse
72 parser = argparse.ArgumentParser()
73
74 parser.add_argument("--model_path", type=str, default="deepseek-ai/Janus-Pro-7B")
75 parser.add_argument("--ckpt_path", type=str, default=None)
76 parser.add_argument("--caption", type=str, default="a brown giraffe and a white stop sign")
77 parser.add_argument("--gen_path", type=str, default='results/samples')
78 parser.add_argument("--cfg", type=float, default=5.0)
79 parser.add_argument("--parallel_size", type=int, default=4)
80
81 args = parser.parse_args()
82 vl_chat_processor: VLChatProcessor = VLChatProcessor.from_pretrained(args.model_path)
83 vl_gpt: MultiModalityCausalLM = AutoModelForCausalLM.from_pretrained(args.model_path, trust_remote_code=True)
84 if args.ckpt_path is not None:
85 state_dict = torch.load(f"{args.ckpt_path}", map_location="cpu")
86 vl_gpt.load_state_dict(state_dict)
87 vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()
88 prompt = f'<|User|>: {args.caption}\n\n<|Assistant|>:<begin_of_image>'
89 images = generate(
90 vl_gpt,
91 vl_chat_processor,
92 prompt,
93 parallel_size = args.parallel_size,
94 cfg_weight = args.cfg,
95 )
96 if not os.path.exists(args.gen_path):
97 os.makedirs(args.gen_path, exist_ok=True)
98 for i in range(args.parallel_size):
99 img_name = str(i).zfill(4)+".png"
100 save_path = os.path.join(args.gen_path, img_name)
101 images[i].save(save_path)1@article{pan2025focusdiff,
2 title={FocusDiff: Advancing Fine-Grained Text-Image Alignment for Autoregressive Visual Generation through RL},
3 author={Pan, Kaihang and Bu, Wendong and Wu, Yuruo and Wu, Yang and Shen, Kai and Li, Yunfei and Zhao, Hang and Li, Juncheng and Tang, Siliang and Zhuang, Yueting},
4 journal={arXiv preprint arXiv:2506.05501},
5 year={2025}
6}