🩺 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
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:
@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},
'''