File size: 9,689 Bytes
4f9aeec
 
 
 
 
 
 
 
1835872
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c3afc00
 
 
1835872
c3afc00
1835872
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
---
license: apache-2.0
language:
- en
base_model:
- Qwen/Qwen2.5-VL-3B-Instruct
pipeline_tag: image-text-to-text
---
# 🩺 ImageDoctor: Diagnosing Text-to-Image Generation via Grounded Image Reasoning

**ImageDoctor** is a unified **evaluation model** for **text-to-image (T2I) generation**, capable of producing both **multi-aspect scalar scores** and **spatially grounded heatmaps**.  
It follows a **“look–think–predict”** reasoning paradigm that mimics human visual diagnosis — first localizing flaws, then reasoning about them, and finally producing an interpretable judgment.  


## 🔍 Key Features

- **🧠 Multi-Aspect Scoring**  
  Predicts four fine-grained quality dimensions:
  - *Plausibility*  
  - *Text–Image Alignment*  
  - *Aesthetics*  
  - *Overall Quality*

- **📍 Spatially Grounded Feedback**  
  Generates **heatmaps** highlighting potential artifacts and misalignments for more interpretable feedback.

- **🗣️ Grounded Image Reasoning**  
  Provides **step-by-step reasoning** that explains how the model arrives at its evaluation.  
  ImageDoctor not only describes *why* an image succeeds or fails but also **zooms in** on flawed regions—offering spatially grounded visual evidence that highlights artifacts or misalignments.  
  This localized reasoning enhances interpretability, transparency, and trustworthiness.

- **⚙️ Reinforcement-Finetuned with GRPO**  
  Trained with **Group Relative Policy Optimization (GRPO)** for richer, more stable preference alignment.

- **💡 Versatile Applications**  
  - As a **metric** for dense image–text evaluation  
  - As a **verifier** in test-time scaling setups  
  - As a **reward model** in reinforcement learning (e.g., DenseFlow-GRPO)

## 🖼️ Quick Start

```python
  from transformers import  AutoProcessor, AutoModelForCausalLM
  from PIL import Image
  import os
  import math
  from qwen_vl_utils import process_vision_info
  import torch
  import numpy as np
  
  def build_messages(image, task_prompt):
      return [{
          "role": "user",
          "content": [
              {"type": "image", "image": image},
              {
                      "type": "text",
                      "text": f"Given a caption and an image generated based on this caption, please analyze the provided image in detail. Evaluate it on various dimensions including Semantic Alignment (How well the image content corresponds to the caption), Aesthetics (composition, color usage, and overall artistic quality), Plausibility (realism and attention to detail), and Overall Impression (General subjective assessment of the image's quality). For each evaluation dimension, provide a score between 0-1 and provide a concise rationale for the score. Use a chain-of-thought process to detail your reasoning steps, and enclose all potential important areas and detailed reasoning within <think> and </think> tags. The important areas are represented in following format: \” I need to focus on the bounding box area. Proposed regions (xyxy): ..., which is an enumerated list in the exact format:1.[x1,y1,x2,y2];\n2.[x1,y1,x2,y2];\n3.[x1,y1,x2,y2]… Here, x1,y1 is the top-left corner, and x2,y2 is the bottom-right corner. Then, within the <answer> and </answer> tags, summarize your assessment in the following format: \"Semantic Alignment score: ... \nMisalignment Locations: ...\nAesthetic score: ...\nPlausibility score: ... nArtifact Locations: ...\nOverall Impression score: ...\". No additional text is allowed in the answer section.\n\n Your actual evaluation should be based on the quality of the provided image.**\n\nYour task is provided as follows:\nText Caption: [{task_prompt}]"
                  }
          ]
      }]
  
  checkpoint = "GYX97/ImageDoctor"
  image_path = "path_to_image"
  prompt = 'prompt_for_image_generation'
  img = Image.open(image_path).convert("RGB")
  r = math.sqrt(512*512 / (img.width * img.height))
  new_size = (max(1, int(img.width * r)), max(1, int(img.height * r)))
  img = img.resize(new_size, resample=Image.BICUBIC)
  
  
  processor = AutoProcessor.from_pretrained(checkpoint, trust_remote_code=True)
  model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto", trust_remote_code=True)
  
  messages = build_messages(img, prompt)
  text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
  image_inputs, video_inputs = process_vision_info(messages)
  output_dir = 'path_to_save_heatmaps'
  
  inputs = processor(
      text=[text],
      images=image_inputs,
      videos=video_inputs,
      padding=True,
      return_tensors="pt"
  ).to(model.device)
  
  # 2) Generate
  gen_kwargs = dict(
      max_new_tokens=20000,
      use_cache=True,
      return_dict_in_generate=True,
      output_hidden_states=True,
  )
  outputs = model.generate(**inputs, **gen_kwargs)
  
  # Decode assistant output (strip prompt tokens)
  generated_ids = outputs.sequences
  trimmed = [out[len(inp):] for inp, out in zip(inputs.input_ids, generated_ids)]
  decoded = processor.batch_decode(trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
  
  print(decoded.strip())
  
  has_tokens = all(
      hasattr(model.config, a) for a in ["image_token_id", "misalignment_token_id", "artifact_token_id"]
  )
  has_heads = all(
      hasattr(model, a) for a in ["text_hidden_fcs", "image_hidden_fcs", "prompt_encoder", "heatmap", "sigmoid"]
  )
  if has_tokens and has_heads and outputs.hidden_states is not None:
      true_generated = generated_ids[:, inputs.input_ids.shape[1]:]
  
      # Find special tokens
      misalignment_mask = (true_generated[:, 1:] == model.config.misalignment_token_id)
      artifact_mask     = (true_generated[:, 1:] == model.config.artifact_token_id)
  
      if misalignment_mask.any() or artifact_mask.any():
          # Gather final-layer hidden states across decoding steps
          step_states = []
          for step in outputs.hidden_states[1:]:  # skip encoder states at index 0
              step_states.append(step[-1])        # last layer [B, 1, H]
          all_gen_h = torch.cat(step_states, dim=1)  # [B, T, H]
  
          # Map text hidden → special token embeddings
          last_hidden_state = model.text_hidden_fcs[0](all_gen_h)  # [B, T, H’]
          # Index by masks (flatten batch/time)
          mis_tokens = last_hidden_state[misalignment_mask].unsqueeze(1) if misalignment_mask.any() else None
          art_tokens = last_hidden_state[artifact_mask].unsqueeze(1) if artifact_mask.any() else None
  
          # Visual embeddings (grid features)
          image_embeds = model.visual(
              inputs["pixel_values"].to(model.device),
              grid_thw=inputs["image_grid_thw"].to(model.device)
          )
          img_hidden = model.image_hidden_fcs[0](image_embeds.unsqueeze(0))  # [1, L, C]
          # reshape to low-res feature map (18x18 matches your original; change if your head differs)
          img_hidden = img_hidden.transpose(1, 2).view(1, -1, 18, 18)
  
          def run_heatmap(text_tokens):
              sparse_embeddings, dense_embeddings = model.prompt_encoder(
                  points=None, boxes=None, masks=None, text_embeds=text_tokens
              )
              low_res = model.heatmap(
                  image_embeddings=img_hidden,
                  image_pe=model.prompt_encoder.get_dense_pe(),
                  sparse_prompt_embeddings=sparse_embeddings.to(img_hidden.dtype),
                  dense_prompt_embeddings=dense_embeddings,
                  multimask_output=False
              )
              return model.sigmoid(low_res)  # [N, 1, H, W]
  
          artifact_np_path = None
          misalign_np_path = None
  
          if output_dir:
              os.makedirs( output_dir, exist_ok=True)
  
          if mis_tokens is not None and art_tokens is not None:
              fused = torch.cat([mis_tokens, art_tokens], dim=0)
              pred = run_heatmap(fused)
              mis_pred = pred[0:1, 0]  # [1,H,W] pick first
              art_pred = pred[1:2, 0]  # [1,H,W] pick second
  
              mis_np = mis_pred[0].detach().cpu().float().numpy()
              art_np = art_pred[0].detach().cpu().float().numpy()
  
              if  output_dir:
                  misalign_np_path = os.path.join( output_dir, f"misalignment.npy")
                  artifact_np_path = os.path.join( output_dir, f"artifact.npy")
                  np.save(misalign_np_path, mis_np)
                  np.save(artifact_np_path, art_np)
  
          elif art_tokens is not None:
              pred = run_heatmap(art_tokens[:1])
              art_np = pred[0, 0].detach().cpu().float().numpy()
              if  output_dir:
                  artifact_np_path = os.path.join( output_dir, f"artifact.npy")
                  np.save(artifact_np_path, art_np)
  
          elif mis_tokens is not None:
              pred = run_heatmap(mis_tokens[:1])
              mis_np = pred[0, 0].detach().cpu().float().numpy()
              if  output_dir:
                  misalign_np_path = os.path.join( output_dir, f"misalignment.npy")
                  np.save(misalign_np_path, mis_np)


```

## 📚 Citation

If you use **ImageDoctor**, please cite:
```bibtex

@misc{guo2025imagedoctordiagnosingtexttoimagegeneration,
  author    = {Yuxiang Guo, Jiang Liu, Ze Wang, Hao Chen, Ximeng Sun, Yang Zhao, Jialian Wu, Xiaodong Yu, Zicheng Liu and Emad Barsoum},
  title     = {ImageDoctor: Diagnosing Text-to-Image Generation via Grounded Image Reasoning}, 
  eprint    = {2510.01010},
  archivePrefix={arXiv},
  year      = {2025},
  url       = {https://arxiv.org/abs/2510.01010},
'''