Activity

1. JjjFangg commented on Sep 9, 2025

   

   JjjFangg

   on Sep 9, 2025

   Collaborator

   More actions

   Thanks for your feedback!

   To reproduce the reported ScreenSpot Pro score (48), please make sure the following settings are applied:

   ✅ Use the official prompt template we provided

   ✅ Set inference to greedy decoding
   temperature = 0
   top_p = 1

   ✅ Set Max pixel to 16384 during inference

   Let us know if you're still seeing discrepancies after applying them.
2. tcnguyen commented on Sep 15, 2025

   

   tcnguyen

   on Sep 15, 2025

   Author

   More actions

   @JjjFangg Hi, thanks for your reply. Did you mean max pixel = 163842828, which is the default value for Qwen2.5VL?
3. Cece1031 commented on Oct 21, 2025

   

   Cece1031

   on Oct 21, 2025

   More actions

   @JjjFangg Hi, thanks for your reply. Did you mean max pixel = 16384_28_28, which is the default value for Qwen2.5VL?

   Hi, I want to know if you can reproduce 48 in the end? I came across this problem too.
4. gaolongxi commented on Oct 23, 2025

   

   gaolongxi

   on Oct 23, 2025

   More actions

   Thanks for your feedback!

   To reproduce the reported ScreenSpot Pro score (48), please make sure the following settings are applied:

   ✅ Use the official prompt template we provided

   ✅ Set inference to greedy decoding temperature = 0 top_p = 1

   ✅ Set Max pixel to 16384 during inference

   Let us know if you're still seeing discrepancies after applying them.

   Hi, thank you for your excellent work!

   I met a similar issue when trying to reproduce the reported accuracy on ScreenSpot-v2 and ScreenSpot-Pro.
   I followed the recommended evaluation settings, but I noticed that UI-TARS-1.5-7B occasionally produces outputs that do not strictly follow the expected format click(point='(x, y)'). To handle this, I additionally support the regex pattern click\(start_box='(x, y)').

   However, even with these adjustments, my best results are 89.0 on ScreenSpot-v2 and 41.6 on ScreenSpot-Pro.

   Below is the relevant code snippet I used for evaluation. I would greatly appreciate any insights or clarifications that could help reproduce the reported results.

   prompt = """You are a GUI agent. You are given a task and your action history, with screenshots. You need to perform the next action to complete the task. \n\n## Output Format\n\nAction: ...\n\n\n## Action Space\nclick(point='<point>x1 y1</point>')\n\n## User Instruction "{instruction}" """
   
   messages = [
       {
           "role": "user",
           "content": [
               {
                   "type": "image", 
                   "image": image_path,
                   "min_pixels": self.processor.image_processor.min_pixels,
                   "max_pixels": self.processor.image_processor.max_pixels
               },
               {"type": "text", "text": full_prompt},
           ],
       }
   ]
   print("self.processor.image_processor.min_pixels:", self.processor.image_processor.min_pixels) #3136
   print("self.processor.image_processor.max_pixels:", self.processor.image_processor.max_pixels) #128450546=16384x28x28
   
   self.generation_params = {'do_sample': False, 'temperature': 0.0, 'use_cache': True, 'top_p': 1.0, 'max_new_tokens': 2048}
   generated_ids = self.model.generate(**inputs, **self.generation_params)
   generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
   
   output_text = self.processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)
   
   input_height = inputs['image_grid_thw'][0][1]*14
   input_width = inputs['image_grid_thw'][0][2]*14
   
   point = [0,0]
   try:
       match = re.search(r"click\(point='\((\d+),\s*(\d+)\)'\)", output_text[0])
       # "raw_response": "Action: click(value=
       if match:
           x, y = float(match.group(1)), float(match.group(2))
           point = [float(x/input_width),float(y/input_height)]
       else:
           match = re.search(r"click\(start_box='\((\d+),\s*(\d+)\)'\)", output_text[0])
           if match:
               x, y = float(match.group(1)), float(match.group(2))
               point = [float(x/input_width),float(y/input_height)]
   except Exception as e:
       print(e)
   
5. Mr-Bigworth commented on Dec 13, 2025

   

   Mr-Bigworth

   on Dec 13, 2025

   More actions

   You are a GUI agent.

   我也是41.7左右，请问老哥解决了不
6. gaolongxi commented on Dec 14, 2025

   

   gaolongxi

   on Dec 14, 2025

   More actions

   You are a GUI agent.

   我也是41.7左右，请问老哥解决了不

   我还没解决 :)
7. Mr-Bigworth commented on Dec 14, 2025

   

   Mr-Bigworth

   on Dec 14, 2025

   More actions

   You are a GUI agent.

   我也是41.7左右，请问老哥解决了不

   我还没解决 :)

   啊 那确实了 我现在在试试 UI-Venus 好像相对好一点