fixes: cudagc guard, rm conditioner.py, turbo depth colormap, proper normal viz, compact UI, example

README.md

@@ -21,6 +21,7 @@ tags:
 
 Estimate depth and surface normals from a single image using [FE2E](https://github.com/AMAP-ML/FE2E) (CVPR 2026).
 
-- Model: Step1X-Edit DiT + LDRN LoRA (FP8 -> dynamic INT8 on CPU)
+- Model: Step1X-Edit DiT + LDRN LoRA, pre-quantized INT8 (12 GB)
 - Single denoise step (not iterative diffusion)
 - Outputs both depth AND normal maps simultaneously
+- Takes ~29 min for 768x1024 on free CPU Basic

app.py

@@ -114,9 +114,24 @@ def generate(image):
 
     elapsed = time.time() - t0
 
-    normal_map = images[0] if images else None
-    Lpred_img = Lpred[0].clamp(0, 1).cpu()
-    depth_map = F.to_pil_image(Lpred_img)
+    import numpy as np
+    import matplotlib.cm as cm
+
+    # Normal map: normalize vectors, map [-1,1] to [0,255] RGB
+    normal_np = Rpred[0].cpu().float().numpy().transpose(1, 2, 0)  # (H, W, 3)
+    normal_norm = np.linalg.norm(normal_np, axis=-1, keepdims=True)
+    normal_norm[normal_norm < 1e-12] = 1e-12
+    normal_np = normal_np / normal_norm
+    normal_rgb = (((normal_np + 1) * 0.5) * 255).clip(0, 255).astype(np.uint8)
+    normal_map = Image.fromarray(normal_rgb)
+    normal_map = normal_map.resize(image.size)
+
+    # Depth map: apply turbo colormap for visualization
+    depth_np = Lpred[0].cpu().float().mean(dim=0).numpy()  # (H, W) average channels
+    depth_np = (depth_np - depth_np.min()) / (depth_np.max() - depth_np.min() + 1e-8)
+    depth_colored = (cm.turbo(depth_np)[:, :, :3] * 255).astype(np.uint8)
+    depth_map = Image.fromarray(depth_colored)
+    depth_map = depth_map.resize(image.size)
 
     status = f"Generated in {elapsed:.1f}s ({image.size[0]}x{image.size[1]}, single denoise, INT8)"
     print(f"[gen] {status}")
@@ -128,16 +143,15 @@ import gradio as gr
 with gr.Blocks(title="FE2E: Depth + Normal (CPU)") as demo:
     gr.Markdown(
         "**[FE2E](https://github.com/AMAP-ML/FE2E)** Depth + Normal from a single image (CVPR 2026). "
-        "Single denoise step, Step1X-Edit DiT + LDRN LoRA (pre-merged), INT8 quantized on CPU."
+        "Takes ~29 min for 768x1024, 1 step, Step1X-Edit DiT + LDRN LoRA (pre-merged), INT8 quantized on CPU."
     )
+    with gr.Row(equal_height=True):
+        input_img = gr.Image(label="Input", type="pil", height=256)
+        depth_out = gr.Image(label="Depth", type="pil", height=256)
+        normal_out = gr.Image(label="Normal", type="pil", height=256)
     with gr.Row():
-        with gr.Column():
-            input_img = gr.Image(label="Input Image", type="pil")
-            run_btn = gr.Button("Estimate Depth + Normal", variant="primary")
-        with gr.Column():
-            depth_out = gr.Image(label="Depth Map", type="pil")
-            normal_out = gr.Image(label="Normal Map", type="pil")
-            status_out = gr.Textbox(label="Status", interactive=False)
+        run_btn = gr.Button("Estimate Depth + Normal", variant="primary", size="lg")
+        status_out = gr.Textbox(label="Status", interactive=False, scale=2)
 
     run_btn.click(
         fn=generate,
@@ -147,6 +161,15 @@ with gr.Blocks(title="FE2E: Depth + Normal (CPU)") as demo:
         api_name="generate",
     )
 
+    gr.Examples(
+        examples=["assets/example.jpg"],
+        inputs=[input_img],
+        outputs=[depth_out, normal_out, status_out],
+        fn=generate,
+        cache_examples=False,
+        label="Examples",
+    )
+
 demo.queue(default_concurrency_limit=1)
 
 if __name__ == "__main__":

infer/inference.py

@@ -29,8 +29,9 @@ EMPTY_PROMPT_LATENT_PATH = REPO_ROOT / "latent" / "no_info.npz"
 
 
 def cudagc():
-    torch.cuda.empty_cache()
-    torch.cuda.ipc_collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
 
 
 def load_state_dict(model, ckpt_path, device="cuda", strict=False, assign=True):

library/lora_module.py

@@ -10,7 +10,10 @@ import math
 import os
 from contextlib import contextmanager
 from typing import Dict, List, Optional, Tuple, Type, Union
-from diffusers import AutoencoderKL
+try:
+    from diffusers import AutoencoderKL
+except ImportError:
+    AutoencoderKL = None
 import numpy as np
 import torch
 from torch import Tensor

modules/conditioner.py

@@ -1,239 +0,0 @@
-import torch
-from qwen_vl_utils import process_vision_info
-from transformers import (
-    AutoProcessor,
-    Qwen2VLForConditionalGeneration,
-    Qwen2_5_VLForConditionalGeneration,
-)
-from torchvision.transforms import ToPILImage
-
-to_pil = ToPILImage()
-
-Qwen25VL_7b_PREFIX = '''You are an expert image analyst specializing in 3D scene understanding. Your goal is to describe the 3D structure and layout of the scene in the image to aid in depth estimation tasks.
-- Focus your description on elements crucial for 3D understanding. Incorporate observations about key **objects**, their **spatial relationships** (like positions, relative distances, and relative sizes using real-world orientation), the overall **layout**, potential **camera perspective**, and visible **depth cues** (like occlusion or perspective lines) into a unified description.
-- Be concise but informative, and assume it is a real-world image. Only describe what you are very confident about.
-User Prompt:'''
-
-Qwen25VL_7b_PREFIX2 = '''Given a user prompt, generate an "Enhanced prompt" that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:
-- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.
-- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.\n
-Here are examples of how to transform or refine prompts:
-- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.
-- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.\n
-Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:
-User Prompt:'''
-
-Qwen25VL_7b_PREFIX3 = '''You are a helpful assistant. Describe the image in detail. User Prompt:'''
-
-def split_string(s):
-    # 将中文引号替换为英文引号
-    s = s.replace("“", '"').replace("”", '"')  # use english quotes
-    result = []
-    # 标记是否在引号内
-    in_quotes = False
-    temp = ""
-
-    # 遍历字符串中的每个字符及其索引
-    for idx, char in enumerate(s):
-        # 如果字符是引号且索引大于 155
-        if char == '"' and idx > 155:
-            # 将引号添加到临时字符串
-            temp += char
-            # 如果不在引号内
-            if not in_quotes:
-                # 将临时字符串添加到结果列表
-                result.append(temp)
-                # 清空临时字符串
-                temp = ""
-
-            # 切换引号状态
-            in_quotes = not in_quotes
-            continue
-        # 如果在引号内
-        if in_quotes:
-            # 如果字符是空格
-            if char.isspace():
-                pass  # have space token
-
-            # 将字符用中文引号包裹后添加到结果列表
-            result.append("“" + char + "”")
-        else:
-            # 将字符添加到临时字符串
-            temp += char
-
-    # 如果临时字符串不为空
-    if temp:
-        # 将临时字符串添加到结果列表
-        result.append(temp)
-
-    return result
-
-
-class Qwen25VL_7b_Embedder(torch.nn.Module):
-    def __init__(self, model_path, max_length=640, dtype=torch.bfloat16, device="cuda",args=None):
-        super(Qwen25VL_7b_Embedder, self).__init__()
-        self.max_length = max_length
-
-        self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
-            model_path,
-            torch_dtype=dtype,
-            attn_implementation="flash_attention_2",
-        ).to(torch.cuda.current_device())
-
-        self.model.requires_grad_(False)
-        self.processor = AutoProcessor.from_pretrained(
-            model_path, min_pixels=256 * 28 * 28, max_pixels=324 * 28 * 28
-        )
-        
-        self.prefix = Qwen25VL_7b_PREFIX if (args is not None) and (not args.old_prompt) else Qwen25VL_7b_PREFIX2
-        self.prefix = Qwen25VL_7b_PREFIX3 if (args is not None) and (args.prompt_type == "rgb" or args.prompt_type == "empty") else self.prefix
-        self.prefix_len = self.calculate_len_of_prefix()
-
-    @property
-    def device(self) -> torch.device:
-        return next(self.parameters()).device
-
-    @property
-    def dtype(self) -> torch.dtype:
-        return next(self.parameters()).dtype
-
-    def calculate_len_of_prefix(self):
-        messages_specific = [{"role": "user", "content": []}]
-        messages_specific[0]["content"].append({"type": "text", "text": f"{self.prefix}"})
-
-        text_specific = self.processor.apply_chat_template(
-            messages_specific, tokenize=False, add_generation_prompt=True, add_vision_id=True
-        )
-        
-
-        inputs_specific = self.processor(
-            text=[text_specific],
-            padding=True,
-            # return_tensors="pt",
-        )
-        len_specific = len(inputs_specific["input_ids"][0]) - 6 # 223 - 6 = self.prefix_len
-        return len_specific
-
-    def forward(self, caption, ref_images):
-        text_list = caption
-        embs = torch.zeros(
-            len(text_list),
-            self.max_length,
-            self.model.config.hidden_size,
-            dtype=torch.bfloat16,
-            device=torch.cuda.current_device(),
-        )
-        hidden_states = torch.zeros(
-            len(text_list),
-            self.max_length,
-            self.model.config.hidden_size,
-            dtype=torch.bfloat16,
-            device=torch.cuda.current_device(),
-        )
-        masks = torch.zeros(
-            len(text_list),
-            self.max_length,
-            dtype=torch.long,
-            device=torch.cuda.current_device(),
-        )
-        input_ids_list = []
-        attention_mask_list = []
-        emb_list = []
-
-        def split_string(s):
-            s = s.replace("“", '"').replace("”", '"').replace("'", '''"''')  # use english quotes
-            result = []
-            in_quotes = False
-            temp = ""
-
-            for idx,char in enumerate(s):
-                if char == '"' and idx>155:
-                    temp += char
-                    if not in_quotes:
-                        result.append(temp)
-                        temp = ""
-
-                    in_quotes = not in_quotes
-                    continue
-                if in_quotes:
-                    if char.isspace():
-                        pass  # have space token
-
-                    result.append("“" + char + "”")
-                else:
-                    temp += char
-
-            if temp:
-                result.append(temp)
-
-            return result
-
-
-        for idx, (txt, imgs) in enumerate(zip(text_list, ref_images)):
-
-            messages = [{"role": "user", "content": []}]
-            messages[0]["content"].append({"type": "text", "text": f"{self.prefix}"})
-            messages[0]["content"].append({"type": "image", "image": to_pil(imgs)})
-            # 再添加 text
-            messages[0]["content"].append({"type": "text", "text": f"{txt}"})
-            # Preparation for inference
-            text = self.processor.apply_chat_template(
-                messages, tokenize=False, add_generation_prompt=True, add_vision_id=True
-            )
-            image_inputs, video_inputs = process_vision_info(messages)
-            inputs = self.processor(
-                text=[text],
-                images=image_inputs,
-                padding=True,
-                return_tensors="pt",
-            )
-            
-            old_inputs_ids = inputs.input_ids
-            text_split_list = split_string(text)
-            token_list = []
-            for text_each in text_split_list:
-                txt_inputs = self.processor(
-                    text=text_each,
-                    images=None,
-                    videos=None,
-                    padding=True,
-                    return_tensors="pt",
-                )
-                token_each = txt_inputs.input_ids
-                if token_each[0][0] == 2073 and token_each[0][-1] == 854:
-                    token_each = token_each[:, 1:-1]
-                    token_list.append(token_each)
-                else:
-                    token_list.append(token_each)
-
-            new_txt_ids = torch.cat(token_list, dim=1).to("cuda")
-
-            new_txt_ids = new_txt_ids.to(old_inputs_ids.device)
-
-            idx1 = (old_inputs_ids == 151653).nonzero(as_tuple=True)[1][0]
-            idx2 = (new_txt_ids == 151653).nonzero(as_tuple=True)[1][0]
-            inputs.input_ids = (
-                torch.cat([old_inputs_ids[0, :idx1], new_txt_ids[0, idx2:]], dim=0)
-                .unsqueeze(0)
-                .to("cuda")
-            )
-            inputs.attention_mask = (inputs.input_ids > 0).long().to("cuda")
-            outputs = self.model(
-                input_ids=inputs.input_ids,
-                attention_mask=inputs.attention_mask,
-                pixel_values=inputs.pixel_values.to("cuda"),
-                image_grid_thw=inputs.image_grid_thw.to("cuda"),
-                output_hidden_states=True,
-            )
-
-            emb = outputs["hidden_states"][-1]
-
-            embs[idx, : min(self.max_length, emb.shape[1] - self.prefix_len)] = emb[0, self.prefix_len:][:self.max_length] #2,640,3584
-
-            masks[idx, : min(self.max_length, emb.shape[1] - self.prefix_len)] = torch.ones(
-                (min(self.max_length, emb.shape[1] - self.prefix_len)),
-                dtype=torch.long,
-                device=torch.cuda.current_device(),
-            )
-
-        return embs, masks

requirements.txt

@@ -7,5 +7,4 @@ safetensors
 einops
 numpy
 tqdm
-diffusers[torch]
-transformers
+matplotlib