Merge local demo into new Space history

.gitattributes

@@ -33,3 +33,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+stimuli/figure_ground.png filter=lfs diff=lfs merge=lfs -text
+stimuli/NeonColorSaeedi.jpg filter=lfs diff=lfs merge=lfs -text
+stimuli/Neon_Color_Circle.jpg filter=lfs diff=lfs merge=lfs -text
+stimuli/Confetti_illusion.png filter=lfs diff=lfs merge=lfs -text
+stimuli/CornsweetBlock.png filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,41 @@
+# Python cache files
+__pycache__/
+*.pyc
+*.pyo
+*.pyd
+.Python
+
+# Log files
+logs/
+*.log
+
+# Model files (will be downloaded at runtime)
+models/
+*.pt
+*.pth
+*.ckpt
+
+# Large image files (uploaded via web interface)
+stimuli/*.jpg
+stimuli/*.png
+stimuli/*.jpeg
+
+# Environment files
+.env
+.venv
+env/
+venv/
+
+# IDE files
+.vscode/
+.idea/
+*.swp
+*.swo
+
+# OS files
+.DS_Store
+Thumbs.db
+
+# Temporary files
+*.tmp
+*.temp

Dockerfile

@@ -0,0 +1,31 @@
+FROM python:3.9-slim
+
+WORKDIR /code
+
+# Install system dependencies
+RUN apt-get update && apt-get install -y \
+    build-essential \
+    git \
+    libgl1-mesa-glx \
+    libglib2.0-0 \
+    && rm -rf /var/lib/apt/lists/*
+
+# Copy only requirements to leverage Docker caching
+COPY ./requirements.txt /code/requirements.txt
+
+# Install Python dependencies
+RUN pip install --no-cache-dir --upgrade -r /code/requirements.txt
+
+# Copy all code and data
+COPY . /code/
+
+# Create necessary directories
+RUN mkdir -p /code/models
+RUN mkdir -p /code/stimuli
+
+# Make sure stimuli and models are writable
+RUN chmod -R 777 /code/models
+RUN chmod -R 777 /code/stimuli
+
+# Set up the command to run the app
+CMD ["python", "app.py"] 

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 GenerativeInferenceDemo
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE. 

README.md

@@ -1,14 +1,131 @@
 ---
-title: Hallucination Prediction Simple
-emoji: 👀
-colorFrom: green
-colorTo: yellow
+title: Human Hallucination Prediction
+emoji: 👁️
+colorFrom: blue
+colorTo: purple
 sdk: gradio
-sdk_version: 6.11.0
-python_version: '3.12'
+sdk_version: 5.23.1
 app_file: app.py
 pinned: false
-short_description: Simplified interface for generative inference by TToosi
+license: mit
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Human Hallucination Prediction
+
+This Gradio demo predicts whether humans will experience visual hallucinations or illusions when viewing specific images. Using adversarially robust neural networks, this tool can forecast perceptual phenomena like illusory contours, figure-ground reversals, and other Gestalt effects before humans report them.
+
+## How It Works
+
+This tool uses **generative inference** with adversarially robust neural networks to predict human visual hallucinations. Robust models trained with adversarial examples develop more human-like perceptual biases, allowing them to predict when humans will perceive:
+
+- **Illusory contours** (Kanizsa shapes, Ehrenstein illusion)
+- **Figure-ground ambiguity** (Rubin's vase, bistable images)
+- **Color spreading effects** (Neon color illusion)
+- **Gestalt grouping** (Continuity, proximity)
+- **Brightness illusions** (Cornsweet effect)
+
+## Features
+
+- **Predict hallucinations** from uploaded images or example illusions
+- **Visualize the prediction process** step-by-step
+- **Compare different models** (robust vs. standard)
+- **Adjust prediction parameters** for different perceptual phenomena
+- **Pre-configured examples** of classic visual illusions
+
+## Usage
+
+1. **Choose an input**: Pick a pre-configured example illusion from the dropdown, or upload your own image.
+2. **Load Parameters**: Click **"Load Parameters"** to fill in optimal prediction settings for that example (or adjust them manually).
+3. **Select the affected part of the visual field**: Set where the model should focus by clicking on the input image or the mask preview to define the mask center, then adjust **Mask center X/Y**, **Mask radius**, and **Mask sigma** in the Adaptive Gaussian mask section if needed. The preview circle shows the region that will receive stronger constraint during inference.
+4. **Run inference**: Click **"Run Generative Inference"** to start the prediction. Progress and intermediate steps are shown in the interface.
+5. **View results**: Inspect the predicted perceptual effects, visualizations, and any generated outputs in the result panels.
+
+## Scientific Background
+
+This demo is based on research showing that adversarially robust neural networks develop perceptual representations similar to human vision. By using generative inference (optimizing images to maximize model confidence), we can reveal what perceptual structures the network expects to see—which often matches what humans hallucinate or perceive in ambiguous images.
+
+## Prerequisites
+
+- **Python** 3.8 or higher
+- **pip** (Python package manager)
+
+## Installation
+
+1. **Clone the repository**
+
+   ```bash
+   git clone https://huggingface.co/spaces/ttoosi/Human_Hallucination_Prediction
+   cd Human_Hallucination_Prediction
+   ```
+
+2. **Create a virtual environment** (recommended)
+
+   ```bash
+   python -m venv venv
+   source venv/bin/activate   # On Windows: venv\Scripts\activate
+   ```
+
+3. **Install dependencies**
+
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+4. **Run the app**
+
+   ```bash
+   python app.py
+   ```
+
+   Optional: specify a port with `--port` (default is 7860):
+
+   ```bash
+   python app.py --port 8861
+   ```
+
+The web app will be available at **http://localhost:7860** (or the port you specified).
+
+**Note:** Model weights (e.g. robust ResNet50) are downloaded automatically from Hugging Face on first run and cached in the `models/` directory. The app also creates a `stimuli/` directory for example images.
+
+### Running with Docker
+
+```bash
+docker build -t human-hallucination-prediction .
+docker run -p 7860:7860 human-hallucination-prediction
+```
+
+Then open http://localhost:7860 in your browser.
+
+## The Prediction Process
+
+1. **Input**: You provide an ambiguous or illusion-inducing image (or use a built-in example).
+2. **Generative inference**: The adversarially robust network iteratively updates the image to maximize its confidence, guided by your chosen parameters (model, layer, noise, step size, etc.).
+3. **Prediction**: The resulting changes reveal the perceptual structures the network expects—which correspond to what humans tend to hallucinate or perceive in such images.
+4. **Visualization**: The interface shows the predicted hallucination and intermediate steps as the optimization runs.
+
+## Models
+
+- **Robust ResNet50**: Trained with adversarial examples (ε=3.0), develops human-like perceptual biases
+- **Standard ResNet50**: Standard ImageNet training without adversarial robustness
+
+## Citation
+
+If you use this work in your research, please cite:
+
+```bibtex
+@article{toosi2024hallucination,
+  title={Predicting Human Visual Hallucinations with Robust Neural Networks},
+  author={Toosi, Tahereh},
+  year={2024}
+}
+```
+
+## About
+
+**Developed by [Tahereh Toosi](https://toosi.github.io)**
+
+This demo demonstrates how adversarially robust neural networks can predict human perceptual hallucinations before they occur.
+
+## License
+
+This project is licensed under the MIT License - see the LICENSE file for details.

app.py

@@ -0,0 +1,905 @@
+import gradio as gr
+import torch
+import numpy as np
+from PIL import Image, ImageDraw
+try:
+    from spaces import GPU
+except ImportError:
+    # Define a no-op decorator if running locally
+    def GPU(func):
+        return func
+    
+import os
+import re
+import json
+import argparse
+from datetime import datetime
+from inference import GenerativeInferenceModel, get_inference_configs, get_imagenet_labels
+
+# Parse command line arguments
+parser = argparse.ArgumentParser(description='Run Generative Inference Demo')
+parser.add_argument('--port', type=int, default=7860, help='Port to run the server on')
+args = parser.parse_args()
+
+# Create model directories if they don't exist
+os.makedirs("models", exist_ok=True)
+os.makedirs("stimuli", exist_ok=True)
+SAVED_RUNS_DIR = "saved_runs"
+os.makedirs(SAVED_RUNS_DIR, exist_ok=True)
+
+# Load ImageNet labels for biased-inference dropdown (1000 classes)
+IMAGENET_LABELS = get_imagenet_labels()
+
+# Check if running on Hugging Face Spaces
+if "SPACE_ID" in os.environ:
+    default_port = int(os.environ.get("PORT", 7860))
+else:
+    default_port = 8861  # Local default port
+
+# Initialize model
+model = GenerativeInferenceModel()
+
+# Define example images and their parameters with updated values from the research
+examples = [
+    {
+        "image": os.path.join("stimuli", "farm1.jpg"),
+        "name": "farm1",
+        "wiki": "https://en.wikipedia.org/wiki/Visual_perception",
+        "papers": [
+            "[Adversarially Robust Vision](https://github.com/MadryLab/robustness)",
+            "[Generative Inference](https://doi.org/10.1016/j.tics.2003.08.003)"
+        ],
+        "method": "Prior-Guided Drift Diffusion",
+        "reverse_diff": {
+            "model": "resnet50_robust",
+            "layer": "all",
+            "initial_noise": 0.0,
+            "diffusion_noise": 0.02,
+            "step_size": 1.0,
+            "iterations": 501,
+            "epsilon": 40.0
+        },
+        "inference_normalization": "off",
+        "use_adaptive_eps": False,
+        "use_adaptive_step": False,
+        "mask_center_x": 0.0,
+        "mask_center_y": 0.0,
+        "mask_radius": 0.2,
+        "mask_sigma": 0.3,
+        "eps_max_mult": 300.0,
+        "eps_min_mult": 1.0,
+        "step_max_mult": 10.0,
+        "step_min_mult": 1.0,
+    },
+    {
+        "image": os.path.join("stimuli", "ArtGallery1.jpg"),
+        "name": "ArtGallery1",
+        "wiki": "https://en.wikipedia.org/wiki/Visual_perception",
+        "papers": [
+            "[Adversarially Robust Vision](https://github.com/MadryLab/robustness)",
+            "[Generative Inference](https://doi.org/10.1016/j.tics.2003.08.003)"
+        ],
+        "method": "Prior-Guided Drift Diffusion",
+        "reverse_diff": {
+            "model": "resnet50_robust",
+            "layer": "layer4",
+            "initial_noise": 0.5,
+            "diffusion_noise": 0.002,
+            "step_size": 0.1,
+            "iterations": 501,
+            "epsilon": 40.0
+        },
+        "inference_normalization": "off",
+        "use_adaptive_eps": False,
+        "use_adaptive_step": True,
+        "mask_center_x": 0.0,
+        "mask_center_y": -1.0,
+        "mask_radius": 0.1,
+        "mask_sigma": 0.2,
+        "eps_max_mult": 30.0,
+        "eps_min_mult": 1.0,
+        "step_max_mult": 100.0,
+        "step_min_mult": 1.0,
+    },
+    {
+        "image": os.path.join("stimuli", "urbanoffice1.jpg"),
+        "name": "UrbanOffice1",
+        "wiki": "https://en.wikipedia.org/wiki/Visual_perception",
+        "papers": [
+            "[Adversarially Robust Vision](https://github.com/MadryLab/robustness)",
+            "[Generative Inference](https://doi.org/10.1016/j.tics.2003.08.003)"
+        ],
+        "method": "Prior-Guided Drift Diffusion",
+        "reverse_diff": {
+            "model": "resnet50_robust",
+            "layer": "all",
+            "initial_noise": 1.0,
+            "diffusion_noise": 0.002,
+            "step_size": 1.0,
+            "iterations": 500,
+            "epsilon": 40.0
+        },
+        "inference_normalization": "off",
+        "use_adaptive_eps": False,
+        "use_adaptive_step": True,
+        "mask_center_x": 0.5,
+        "mask_center_y": 0.0,
+        "mask_radius": 0.2,
+        "mask_sigma": 0.2,
+        "eps_max_mult": 20.0,
+        "eps_min_mult": 1.0,
+        "step_max_mult": 50.0,
+        "step_min_mult": 0.2,
+    },
+    {
+        "image": os.path.join("stimuli", "Neon_Color_Circle.jpg"),
+        "name": "Neon Color Spreading",
+        "wiki": "https://en.wikipedia.org/wiki/Neon_color_spreading",
+        "papers": [
+            "[Color Assimilation](https://doi.org/10.1016/j.visres.2000.200.1)",
+            "[Perceptual Filling-in](https://doi.org/10.1016/j.tics.2003.08.003)"
+        ],
+        "method": "Prior-Guided Drift Diffusion",
+        "reverse_diff": {
+            "model": "resnet50_robust",
+            "layer": "layer3",
+            "initial_noise": 0.8,
+            "diffusion_noise": 0.003,
+            "step_size": 1.0,
+            "iterations": 101,
+            "epsilon": 20.0
+        },
+        "use_adaptive_eps": False,
+        "use_adaptive_step": False,
+        "mask_center_x": 0.0,
+        "mask_center_y": 0.0,
+        "mask_radius": 0.2,
+        "mask_sigma": 1.0,
+        "eps_max_mult": 1.0,
+        "eps_min_mult": 1.0,
+        "step_max_mult": 1.0,
+        "step_min_mult": 1.0,
+    },
+    {
+        "image": os.path.join("stimuli", "Kanizsa_square.jpg"),
+        "name": "Kanizsa Square",
+        "wiki": "https://en.wikipedia.org/wiki/Kanizsa_triangle",
+        "papers": [
+            "[Gestalt Psychology](https://en.wikipedia.org/wiki/Gestalt_psychology)",
+            "[Neural Mechanisms](https://doi.org/10.1016/j.tics.2003.08.003)"
+        ],
+        "method": "Prior-Guided Drift Diffusion",
+        "reverse_diff": {
+            "model": "resnet50_robust",
+            "layer": "all",
+            "initial_noise": 0.0,
+            "diffusion_noise": 0.005,
+            "step_size": 0.64,
+            "iterations": 100,
+            "epsilon": 5.0
+        },
+        "use_adaptive_eps": False,
+        "use_adaptive_step": False,
+        "mask_center_x": 0.0,
+        "mask_center_y": 0.0,
+        "mask_radius": 0.2,
+        "mask_sigma": 1.0,
+        "eps_max_mult": 1.0,
+        "eps_min_mult": 1.0,
+        "step_max_mult": 1.0,
+        "step_min_mult": 1.0,
+    },
+    {
+        "image": os.path.join("stimuli", "CornsweetBlock.png"),
+        "name": "Cornsweet Illusion",
+        "wiki": "https://en.wikipedia.org/wiki/Cornsweet_illusion",
+        "papers": [
+            "[Brightness Perception](https://doi.org/10.1016/j.visres.2000.200.1)",
+            "[Edge Effects](https://doi.org/10.1016/j.tics.2003.08.003)"
+        ],
+        "instructions": "Both blocks are gray in color (the same), use your finger to cover the middle line. Hit 'Load Parameters' and then hit 'Run Generative Inference' to see how the model sees the blocks.",
+        "method": "Prior-Guided Drift Diffusion",
+        "reverse_diff": {
+            "model": "resnet50_robust",
+            "layer": "layer3",
+            "initial_noise": 0.5,
+            "diffusion_noise": 0.005,
+            "step_size": 0.8,
+            "iterations": 51,
+            "epsilon": 20.0
+        },
+        "use_adaptive_eps": False,
+        "use_adaptive_step": False,
+        "mask_center_x": 0.0,
+        "mask_center_y": 0.0,
+        "mask_radius": 0.2,
+        "mask_sigma": 1.0,
+        "eps_max_mult": 1.0,
+        "eps_min_mult": 1.0,
+        "step_max_mult": 1.0,
+        "step_min_mult": 1.0,
+    },
+    {
+        "image": os.path.join("stimuli", "face_vase.png"),
+        "name": "Rubin's Face-Vase (Object Prior)",
+        "wiki": "https://en.wikipedia.org/wiki/Rubin_vase",
+        "papers": [
+            "[Figure-Ground Perception](https://en.wikipedia.org/wiki/Figure-ground_(perception))",
+            "[Bistable Perception](https://doi.org/10.1016/j.tics.2003.08.003)"
+        ],
+        "method": "Prior-Guided Drift Diffusion",
+        "reverse_diff": {
+            "model": "resnet50_robust",
+            "layer": "avgpool",
+            "initial_noise": 0.9,
+            "diffusion_noise": 0.003,
+            "step_size": 0.58,
+            "iterations": 100,
+            "epsilon": 0.81
+        },
+        "use_adaptive_eps": False,
+        "use_adaptive_step": False,
+        "mask_center_x": 0.0,
+        "mask_center_y": 0.0,
+        "mask_radius": 0.2,
+        "mask_sigma": 1.0,
+        "eps_max_mult": 1.0,
+        "eps_min_mult": 1.0,
+        "step_max_mult": 1.0,
+        "step_min_mult": 1.0,
+    },
+    {
+        "image": os.path.join("stimuli", "Confetti_illusion.png"),
+        "name": "Confetti Illusion",
+        "wiki": "https://www.youtube.com/watch?v=SvEiEi8O7QE",
+        "papers": [
+            "[Color Perception](https://doi.org/10.1016/j.visres.2000.200.1)",
+            "[Context Effects](https://doi.org/10.1016/j.tics.2003.08.003)"
+        ],
+        "method": "Prior-Guided Drift Diffusion",
+        "reverse_diff": {
+            "model": "resnet50_robust",
+            "layer": "layer3",
+            "initial_noise": 0.1,
+            "diffusion_noise": 0.003,
+            "step_size": 0.5,
+            "iterations": 101,
+            "epsilon": 20.0
+        },
+        "use_adaptive_eps": False,
+        "use_adaptive_step": False,
+        "mask_center_x": 0.0,
+        "mask_center_y": 0.0,
+        "mask_radius": 0.2,
+        "mask_sigma": 1.0,
+        "eps_max_mult": 1.0,
+        "eps_min_mult": 1.0,
+        "step_max_mult": 1.0,
+        "step_min_mult": 1.0,
+    },
+    {
+        "image": os.path.join("stimuli", "EhresteinSingleColor.png"),
+        "name": "Ehrenstein Illusion",
+        "wiki": "https://en.wikipedia.org/wiki/Ehrenstein_illusion",
+        "papers": [
+            "[Subjective Contours](https://doi.org/10.1016/j.visres.2000.200.1)",
+            "[Neural Processing](https://doi.org/10.1016/j.tics.2003.08.003)"
+        ],
+        "method": "Prior-Guided Drift Diffusion",
+        "reverse_diff": {
+            "model": "resnet50_robust",
+            "layer": "layer3",
+            "initial_noise": 0.5,
+            "diffusion_noise": 0.005,
+            "step_size": 0.8,
+            "iterations": 101,
+            "epsilon": 20.0
+        },
+        "use_adaptive_eps": False,
+        "use_adaptive_step": False,
+        "mask_center_x": 0.0,
+        "mask_center_y": 0.0,
+        "mask_radius": 0.2,
+        "mask_sigma": 1.0,
+        "eps_max_mult": 1.0,
+        "eps_min_mult": 1.0,
+        "step_max_mult": 1.0,
+        "step_min_mult": 1.0,
+    },
+    {
+        "image": os.path.join("stimuli", "GroupingByContinuity.png"),
+        "name": "Grouping by Continuity",
+        "wiki": "https://en.wikipedia.org/wiki/Principles_of_grouping",
+        "papers": [
+            "[Gestalt Principles](https://en.wikipedia.org/wiki/Gestalt_psychology)",
+            "[Visual Organization](https://doi.org/10.1016/j.tics.2003.08.003)"
+        ],
+        "method": "Prior-Guided Drift Diffusion",
+        "reverse_diff": {
+            "model": "resnet50_robust",
+            "layer": "layer3",
+            "initial_noise": 0.0,
+            "diffusion_noise": 0.005,
+            "step_size": 0.4,
+            "iterations": 101,
+            "epsilon": 4.0
+        },
+        "use_adaptive_eps": False,
+        "use_adaptive_step": False,
+        "mask_center_x": 0.0,
+        "mask_center_y": 0.0,
+        "mask_radius": 0.2,
+        "mask_sigma": 1.0,
+        "eps_max_mult": 1.0,
+        "eps_min_mult": 1.0,
+        "step_max_mult": 1.0,
+        "step_min_mult": 1.0,
+    },
+    {
+        "image": os.path.join("stimuli", "figure_ground.png"),
+        "name": "Figure-Ground Illusion",
+        "wiki": "https://en.wikipedia.org/wiki/Figure-ground_(perception)",
+        "papers": [
+            "[Gestalt Principles](https://en.wikipedia.org/wiki/Gestalt_psychology)",
+            "[Perceptual Organization](https://doi.org/10.1016/j.tics.2003.08.003)"
+        ],
+        "method": "Prior-Guided Drift Diffusion",
+        "reverse_diff": {
+            "model": "resnet50_robust",
+            "layer": "layer3",
+            "initial_noise": 0.1,
+            "diffusion_noise": 0.003,
+            "step_size": 0.5,
+            "iterations": 101,
+            "epsilon": 3.0
+        },
+        "use_adaptive_eps": False,
+        "use_adaptive_step": False,
+        "mask_center_x": 0.0,
+        "mask_center_y": 0.0,
+        "mask_radius": 0.2,
+        "mask_sigma": 1.0,
+        "eps_max_mult": 1.0,
+        "eps_min_mult": 1.0,
+        "step_max_mult": 1.0,
+        "step_min_mult": 1.0,
+    }
+]
+
+def _input_image_stem(image):
+    """Return a safe filename stem from the input image: known name or 'user_img'."""
+    if image is None:
+        return "user_img"
+    path = None
+    if isinstance(image, str) and (os.path.isfile(image) or os.path.exists(image)):
+        path = image
+    if isinstance(image, dict) and image.get("path") and os.path.exists(image.get("path", "")):
+        path = image["path"]
+    if path:
+        name = os.path.splitext(os.path.basename(path))[0]
+        # Safe for filenames: alphanumeric, underscore, hyphen only; max length
+        safe = re.sub(r"[^\w\-]", "_", name).strip("_") or "user_img"
+        return safe[:80] if len(safe) > 80 else safe
+    return "user_img"
+
+
+def _get_image_path_for_stem(img):
+    """Extract file path from Gradio image value (path string, dict with path, or PIL) for stem tracking."""
+    if img is None:
+        return ""
+    if isinstance(img, str) and (os.path.isfile(img) or os.path.exists(img)):
+        return img
+    if isinstance(img, dict) and img.get("path"):
+        p = img["path"]
+        if isinstance(p, str) and os.path.exists(p):
+            return p
+    return ""
+
+
+def _update_tracked_image_path(img):
+    """Keep path only when it's a known stimulus (e.g. from stimuli/); else '' so stem is 'user_img'."""
+    path = _get_image_path_for_stem(img)
+    if path and "stimuli" in path:
+        return path
+    return ""
+
+
+def _config_to_json_serializable(c):
+    """Return a copy of config with only JSON-serializable values."""
+    if isinstance(c, dict):
+        return {k: _config_to_json_serializable(v) for k, v in c.items()}
+    if isinstance(c, (list, tuple)):
+        return [_config_to_json_serializable(x) for x in c]
+    if isinstance(c, (bool, int, float, str, type(None))):
+        return c
+    if hasattr(c, "item"):  # e.g. numpy scalar
+        return c.item()
+    return str(c)
+
+
+@GPU 
+def run_inference(image, model_type, inference_type, eps_value, num_iterations, 
+                 initial_noise=0.05, diffusion_noise=0.3, step_size=0.8, model_layer="layer3",
+                 use_adaptive_eps=False, use_adaptive_step=False,
+                 mask_center_x=0.0, mask_center_y=0.0, mask_radius=0.3, mask_sigma=0.2,
+                 eps_max_mult=4.0, eps_min_mult=1.0, step_max_mult=4.0, step_min_mult=1.0,
+                 use_biased_inference=False, biased_class_name="",
+                 current_image_path=""):
+    # Check if image is provided
+    if image is None:
+        return None, [], "Please upload an image before running inference.", None
+    
+    # Convert eps to float
+    eps = float(eps_value)
+    step_size_f = float(step_size)
+    # Coerce checkbox values (Gradio can send bool, None, or other)
+    use_adaptive_eps = bool(use_adaptive_eps) if use_adaptive_eps is not None else False
+    use_adaptive_step = bool(use_adaptive_step) if use_adaptive_step is not None else False
+    
+    # Load inference configuration based on the selected type
+    config = get_inference_configs(inference_type=inference_type, eps=eps, n_itr=int(num_iterations))
+    
+    # Handle Prior-Guided Drift Diffusion specific parameters
+    if inference_type == "Prior-Guided Drift Diffusion":
+        config['initial_inference_noise_ratio'] = float(initial_noise)
+        config['diffusion_noise_ratio'] = float(diffusion_noise)
+        config['step_size'] = step_size_f
+        config['top_layer'] = model_layer
+    
+    # Inference normalization off (option removed from UI)
+    config['inference_normalization'] = 'off'
+    config['recognition_normalization'] = 'off'
+    
+    # Adaptive epsilon (Gaussian mask)
+    if use_adaptive_eps:
+        config['adaptive_epsilon'] = {
+            'enabled': True,
+            'base_epsilon': eps,
+            'center_x': float(mask_center_x),
+            'center_y': float(mask_center_y),
+            'flat_radius': float(mask_radius),
+            'sigma': float(mask_sigma),
+            'max_multiplier': float(eps_max_mult),
+            'min_multiplier': float(eps_min_mult),
+        }
+    else:
+        config['adaptive_epsilon'] = None
+    
+    # Adaptive step size (same Gaussian mask location/size, different multipliers)
+    if use_adaptive_step:
+        config['adaptive_step_size'] = {
+            'enabled': True,
+            'base_step_size': step_size_f,
+            'center_x': float(mask_center_x),
+            'center_y': float(mask_center_y),
+            'flat_radius': float(mask_radius),
+            'sigma': float(mask_sigma),
+            'max_multiplier': float(step_max_mult),
+            'min_multiplier': float(step_min_mult),
+        }
+    else:
+        config['adaptive_step_size'] = None
+    
+    # Biased inference: bias perception toward a target ImageNet class
+    use_biased_inference = bool(use_biased_inference) if use_biased_inference is not None else False
+    biased_class_name = (biased_class_name or "").strip() if biased_class_name else ""
+    if use_biased_inference and biased_class_name:
+        config['biased_inference'] = {'enable': True, 'class': biased_class_name}
+    else:
+        config['biased_inference'] = config.get('biased_inference') or {'enable': False, 'class': None}
+    
+    # Run generative inference
+    result = model.inference(image, model_type, config)
+    
+    # Extract results based on return type
+    if isinstance(result, tuple):
+        # Old format returning (output_image, all_steps)
+        output_image, all_steps = result
+    else:
+        # New format returning dictionary
+        output_image = result['final_image']
+        all_steps = result['steps']
+    
+    # Create animation frames
+    frames = []
+    for i, step_image in enumerate(all_steps):
+        # Convert tensor to PIL image
+        step_pil = Image.fromarray((step_image.permute(1, 2, 0).cpu().numpy() * 255).astype(np.uint8))
+        frames.append(step_pil)
+    
+    # Convert the final output image to PIL
+    final_image = Image.fromarray((output_image.permute(1, 2, 0).cpu().numpy() * 255).astype(np.uint8))
+    
+    # Always save GIF and config and offer as downloads (browser will ask where to save)
+    save_status = ""
+    files_for_download = None
+    if frames:
+        # Use tracked path when available (e.g. from Load Parameters); else derive from image (PIL loses path)
+        stem = _input_image_stem(current_image_path if (current_image_path and current_image_path.strip()) else image)
+        unique_id = f"{datetime.now().strftime('%Y%m%d_%H%M%S')}_{stem}"
+        gif_path = os.path.join(SAVED_RUNS_DIR, f"{unique_id}.gif")
+        config_path = os.path.join(SAVED_RUNS_DIR, f"{unique_id}_config.json")
+        try:
+            frames[0].save(
+                gif_path,
+                save_all=True,
+                append_images=frames[1:],
+                loop=0,
+                duration=200,
+            )
+            save_config = {
+                "model_type": model_type,
+                "input_image_name": stem,
+                **_config_to_json_serializable(config),
+            }
+            with open(config_path, "w") as f:
+                json.dump(save_config, f, indent=2)
+            files_for_download = [gif_path, config_path]
+            save_status = "**Download results** — Use the links below to save the GIF and config to your device (your browser may ask where to save)."
+        except Exception as e:
+            save_status = f"Save failed: {e}"
+    
+    return final_image, frames, save_status, files_for_download
+
+def _image_to_pil(img):
+    """Convert Gradio image value (PIL, numpy, path, or dict) to PIL Image; return None if invalid."""
+    if img is None:
+        return None
+    if isinstance(img, Image.Image):
+        return img
+    if isinstance(img, np.ndarray):
+        return Image.fromarray(img.astype(np.uint8))
+    if isinstance(img, dict) and "path" in img:
+        return Image.open(img["path"]).convert("RGB")
+    if isinstance(img, str) and os.path.exists(img):
+        return Image.open(img).convert("RGB")
+    return None
+
+
+def _image_size(img):
+    """Return (width, height) from Gradio image value, or (1, 1) if unknown."""
+    pil = _image_to_pil(img)
+    if pil is not None:
+        return pil.size
+    return (1, 1)
+
+
+def image_click_to_center(evt: gr.SelectData):
+    """Convert click (x, y) on image to normalized mask center (-1 to 1). Returns (center_x, center_y)."""
+    if evt.index is None or (isinstance(evt.index, (list, tuple)) and len(evt.index) < 2):
+        return 0.0, 0.0
+    x, y = float(evt.index[0]), float(evt.index[1])
+    w, h = _image_size(evt.value)
+    if w <= 0 or h <= 0:
+        return 0.0, 0.0
+    center_x = (x / w) * 2.0 - 1.0
+    center_y = (y / h) * 2.0 - 1.0
+    center_x = max(-1.0, min(1.0, center_x))
+    center_y = max(-1.0, min(1.0, center_y))
+    return center_x, center_y
+
+
+def draw_mask_overlay(image, center_x, center_y, radius):
+    """Draw the Gaussian mask center and radius on a copy of the image. Returns PIL or None."""
+    pil = _image_to_pil(image)
+    if pil is None:
+        return None
+    img = pil.convert("RGB").copy()
+    w, h = img.size
+    cx_px = (float(center_x) + 1.0) / 2.0 * w
+    cy_px = (float(center_y) + 1.0) / 2.0 * h
+    radius_px = float(radius) * min(w, h) / 2.0
+    draw = ImageDraw.Draw(img)
+    # Circle for radius
+    draw.ellipse(
+        [cx_px - radius_px, cy_px - radius_px, cx_px + radius_px, cy_px + radius_px],
+        outline="#E11D48",
+        width=2 * max(2, min(w, h) // 150),
+    )
+    # Center dot
+    r = max(2, min(w, h) // 80)
+    draw.ellipse([cx_px - r, cy_px - r, cx_px + r, cy_px + r], fill="#E11D48", outline="#FFF")
+    return img
+
+
+# Helper function to apply example parameters (adaptive mask off by default unless example defines it)
+def apply_example(example):
+    rd = example["reverse_diff"]
+    mcx = example.get("mask_center_x", 0.0)
+    mcy = example.get("mask_center_y", 0.0)
+    mrad = example.get("mask_radius", 0.3)
+    mask_img = draw_mask_overlay(example["image"], mcx, mcy, mrad)
+    return [
+        example["image"],
+        rd.get("model", "resnet50_robust"),
+        example["method"],
+        rd["epsilon"],
+        rd["iterations"],
+        rd["initial_noise"],
+        rd["diffusion_noise"],
+        rd["step_size"],
+        rd["layer"],
+        example.get("use_adaptive_eps", False),
+        example.get("use_adaptive_step", False),
+        mcx,
+        mcy,
+        example.get("mask_radius", 0.3),
+        example.get("mask_sigma", 0.2),
+        example.get("eps_max_mult", 4.0),
+        example.get("eps_min_mult", 1.0),
+        example.get("step_max_mult", 4.0),
+        example.get("step_min_mult", 1.0),
+        example.get("use_biased_inference", False),
+        example.get("biased_class_name", ""),
+        example["image"],  # keep path for save filename (e.g. UrbanOffice1 -> urbanoffice1)
+        mask_img,
+        gr.Group(visible=True),
+    ]
+
+# Define the interface
+with gr.Blocks(title="Human Hallucination Prediction", css="""
+.purple-button {
+    background-color: #8B5CF6 !important;
+    color: white !important;
+    border: none !important;
+}
+.purple-button:hover {
+    background-color: #7C3AED !important;
+}
+""") as demo:
+    gr.Markdown("# Human Hallucination Prediction")
+    gr.Markdown("**Predict what visual hallucinations humans may experience** using neural networks.")
+    
+    gr.Markdown("""
+    **How to predict hallucinations:**
+    1. **Select an example image** below and click "Load Parameters" to set the prediction settings
+    2. **Click "Run Generative Inference"** to predict what hallucination humans may perceive
+    3. **View the prediction**: Watch as the model reveals the perceptual structures it expects—matching what humans typically hallucinate
+    4. **You can upload your own images**
+    5. **You can download the results** as a .gif file together with the configs.json
+    """)
+    with gr.Row():
+        with gr.Column(scale=1):
+            # Inputs (track path so save filenames use stimulus name when from example)
+            default_image_path = os.path.join("stimuli", "urbanoffice1.jpg")
+            image_input = gr.Image(label="Input Image (click to set mask center)", type="pil", value=default_image_path)
+            current_image_path_state = gr.State(value=default_image_path)
+            mask_preview = gr.Image(
+                label="Mask center preview (click to set center — circle shows mask)",
+                type="pil",
+                interactive=False,
+            )
+            # Run Inference button right below the image
+            run_button = gr.Button("🔮 Predict Hallucination", variant="primary", elem_classes="purple-button")
+            
+            # Parameters toggle button
+            params_button = gr.Button("⚙️ Play with the parameters", variant="secondary")
+            
+            # Parameters section (initially hidden)
+            with gr.Group(visible=False) as params_section:
+                with gr.Row():
+                    model_choice = gr.Dropdown(
+                        choices=["resnet50_robust", "standard_resnet50"], # "resnet50_robust_face" - hidden for deployment
+                        value="resnet50_robust", 
+                        label="Model"
+                    )
+                    
+                    inference_type = gr.Dropdown(
+                        choices=["Prior-Guided Drift Diffusion", "IncreaseConfidence"], 
+                        value="Prior-Guided Drift Diffusion", 
+                        label="Inference Method"
+                    )
+                
+                with gr.Row():
+                    eps_slider = gr.Slider(minimum=0.0, maximum=40.0, value=40.0, step=0.01, label="Epsilon (Stimulus Fidelity)")
+                    iterations_slider = gr.Slider(minimum=1, maximum=600, value=500, step=1, label="Number of Iterations")
+                
+                with gr.Row():
+                    initial_noise_slider = gr.Slider(minimum=0.0, maximum=1.0, value=1.0, step=0.01, 
+                                                   label="Drift Noise")
+                    diffusion_noise_slider = gr.Slider(minimum=0.0, maximum=0.05, value=0.002, step=0.001, 
+                                                    label="Diffusion Noise")
+                    
+                with gr.Row():
+                    step_size_slider = gr.Slider(minimum=0.01, maximum=2.0, value=1.0, step=0.01, 
+                                               label="Update Rate")
+                    layer_choice = gr.Dropdown(
+                        choices=["all", "conv1", "bn1", "relu", "maxpool", "layer1", "layer2", "layer3", "layer4", "avgpool"],
+                        value="all",
+                        label="Model Layer"
+                    )
+                
+                gr.Markdown("### 🎯 Adaptive Gaussian mask (spatially varying constraint)")
+                gr.Markdown("Define where on the image the mask is centered and how large its radius is. Coordinates: **-1** = left/top, **1** = right/bottom, **0** = center.")
+                with gr.Row():
+                    use_adaptive_eps_check = gr.Checkbox(value=False, label="Use adaptive epsilon (stronger/weaker constraint by region)")
+                    use_adaptive_step_check = gr.Checkbox(value=True, label="Use adaptive step size (stronger/weaker updates by region)")
+                with gr.Row():
+                    mask_center_x_slider = gr.Slider(minimum=-1.0, maximum=1.0, value=0.5, step=0.05, label="Mask center X")
+                    mask_center_y_slider = gr.Slider(minimum=-1.0, maximum=1.0, value=0.0, step=0.05, label="Mask center Y")
+                with gr.Row():
+                    mask_radius_slider = gr.Slider(minimum=0.01, maximum=1.0, value=0.2, step=0.01, label="Mask radius (flat region size)")
+                    mask_sigma_slider = gr.Slider(minimum=0.05, maximum=1.0, value=0.2, step=0.01, label="Mask sigma (fall-off outside radius)")
+                with gr.Row():
+                    eps_max_mult_slider = gr.Slider(minimum=0.1, maximum=350.0, value=20.0, step=0.1, label="Epsilon: multiplier at center")
+                    eps_min_mult_slider = gr.Slider(minimum=0.1, maximum=10.0, value=1.0, step=0.1, label="Epsilon: multiplier at periphery")
+                with gr.Row():
+                    step_max_mult_slider = gr.Slider(minimum=0.1, maximum=150.0, value=50.0, step=0.1, label="Step size: multiplier at center")
+                    step_min_mult_slider = gr.Slider(minimum=0.1, maximum=10.0, value=0.2, step=0.1, label="Step size: multiplier at periphery")
+                gr.Markdown("### 🎯 Biased inference")
+                gr.Markdown("Bias the prediction toward a specific ImageNet category (1000 classes).")
+                with gr.Row():
+                    use_biased_inference_check = gr.Checkbox(value=False, label="Use biased inference (bias toward a target class)")
+                    biased_class_dropdown = gr.Dropdown(
+                        choices=[("— No bias —", "")] + [(label, label) for label in sorted(IMAGENET_LABELS)],
+                        value="",
+                        label="Biased toward category",
+                        allow_custom_value=False,
+                        filterable=True,
+                    )
+        with gr.Column(scale=2):
+            # Outputs
+            output_image = gr.Image(label="Predicted Hallucination")
+            output_frames = gr.Gallery(label="Hallucination Prediction Process", columns=5, rows=2)
+            save_status_md = gr.Markdown(value="")
+            download_files = gr.File(label="Download results (GIF + config)", file_count="multiple")
+    
+    # Examples section with integrated explanations
+    gr.Markdown("## Examples")
+    gr.Markdown("Select an example and click Load Parameters to apply its settings")
+    
+    # For each example, create a row with the image and explanation side by side
+    for i, ex in enumerate(examples):
+        with gr.Row():
+            # Left column for the image
+            with gr.Column(scale=1):
+                # Display the example image
+                example_img = gr.Image(value=ex["image"], type="filepath", label=f"{ex['name']}")
+                load_btn = gr.Button(f"Load Parameters", variant="primary")
+                
+                # Set up the load button to apply this example's parameters
+                load_btn.click(
+                    fn=lambda ex=ex: apply_example(ex),
+                    outputs=[
+                        image_input, model_choice, inference_type,
+                        eps_slider, iterations_slider,
+                        initial_noise_slider, diffusion_noise_slider,
+                        step_size_slider, layer_choice,
+                        use_adaptive_eps_check, use_adaptive_step_check,
+                        mask_center_x_slider, mask_center_y_slider,
+                        mask_radius_slider, mask_sigma_slider,
+                        eps_max_mult_slider, eps_min_mult_slider,
+                        step_max_mult_slider, step_min_mult_slider,
+                        use_biased_inference_check, biased_class_dropdown,
+                        current_image_path_state,
+                        mask_preview,
+                        params_section,
+                    ],
+                )
+            
+            # Right column for the explanation
+            with gr.Column(scale=2):
+                gr.Markdown(f"### {ex['name']}")
+                if ex["name"] not in ("farm1", "ArtGallery1", "UrbanOffice1"):
+                    gr.Markdown(f"[Read more on Wikipedia]({ex['wiki']})")
+                
+                # Show instructions if they exist
+                if "instructions" in ex:
+                    gr.Markdown(f"**Instructions:** {ex['instructions']}")
+                
+        
+        if i < len(examples) - 1:  # Don't add separator after the last example
+            gr.Markdown("---")
+    
+    # Set up event handler for the main inference
+    run_button.click(
+        fn=run_inference,
+        inputs=[
+            image_input, model_choice, inference_type,
+            eps_slider, iterations_slider,
+            initial_noise_slider, diffusion_noise_slider,
+            step_size_slider, layer_choice,
+            use_adaptive_eps_check, use_adaptive_step_check,
+            mask_center_x_slider, mask_center_y_slider,
+            mask_radius_slider, mask_sigma_slider,
+            eps_max_mult_slider, eps_min_mult_slider,
+            step_max_mult_slider, step_min_mult_slider,
+            use_biased_inference_check, biased_class_dropdown,
+            current_image_path_state,
+        ],
+        outputs=[output_image, output_frames, save_status_md, download_files]
+    )
+    
+    # Toggle parameters visibility
+    def toggle_params():
+        return gr.Group(visible=True)
+
+    params_button.click(
+        fn=toggle_params,
+        outputs=[params_section],
+    )
+
+    # Click on input image or mask preview to set Gaussian mask center
+    def _mask_preview_inputs():
+        return [image_input, mask_center_x_slider, mask_center_y_slider, mask_radius_slider]
+
+    image_input.select(
+        fn=image_click_to_center,
+        outputs=[mask_center_x_slider, mask_center_y_slider],
+    )
+    mask_preview.select(
+        fn=image_click_to_center,
+        outputs=[mask_center_x_slider, mask_center_y_slider],
+    )
+    # Update mask preview when image or center/radius change
+    image_input.change(
+        fn=draw_mask_overlay,
+        inputs=_mask_preview_inputs(),
+        outputs=[mask_preview],
+    )
+    # Keep tracked path for save filename: known stimulus name or clear so stem becomes 'user_img'
+    image_input.change(
+        fn=_update_tracked_image_path,
+        inputs=[image_input],
+        outputs=[current_image_path_state],
+    )
+    mask_center_x_slider.change(
+        fn=draw_mask_overlay,
+        inputs=_mask_preview_inputs(),
+        outputs=[mask_preview],
+    )
+    mask_center_y_slider.change(
+        fn=draw_mask_overlay,
+        inputs=_mask_preview_inputs(),
+        outputs=[mask_preview],
+    )
+    mask_radius_slider.change(
+        fn=draw_mask_overlay,
+        inputs=_mask_preview_inputs(),
+        outputs=[mask_preview],
+    )
+    # Populate mask preview on load
+    demo.load(
+        fn=draw_mask_overlay,
+        inputs=_mask_preview_inputs(),
+        outputs=[mask_preview],
+    )
+    
+    # About section
+    gr.Markdown("""
+    ## 🧠 About Hallucination Prediction
+    
+    This tool predicts human visual hallucinations using **generative inference** with adversarially robust neural networks. Robust models develop human-like perceptual biases, allowing them to forecast what perceptual structures humans will experience.
+    
+    ### Prediction Methods:
+    
+    **Prior-Guided Drift Diffusion** (Primary Method)  
+    Starting from a noisy representation, the model converges toward what it expects to perceive—revealing predicted hallucinations
+    
+    **IncreaseConfidence**  
+    Moving away from unlikely interpretations to reveal the most probable perceptual experience
+    
+    ### Parameters:
+    - **Drift Noise**: Initial uncertainty in the prediction process
+    - **Diffusion Noise**: Stochastic exploration during prediction
+    - **Update Rate**: Speed of convergence to the predicted hallucination
+    - **Number of Iterations**: How many prediction steps to perform
+    - **Model Layer**: Which perceptual level to predict from (early edges vs. high-level objects)
+    - **Epsilon (Stimulus Fidelity)**: How closely the prediction must match the input stimulus
+    
+    ### Why Does This Work?
+    
+    Adversarially robust neural networks develop perceptual representations similar to human vision. When we use generative inference to reveal what these networks "expect" to see, it matches what humans hallucinate in ambiguous images—allowing us to predict human perception.
+    
+    **Developed by [Tahereh Toosi](https://toosi.github.io)**
+    """)
+
+# Launch the demo
+if __name__ == "__main__":
+    print(f"Starting server on port {args.port}")
+    demo.launch(
+        server_name="0.0.0.0",
+        server_port=args.port,
+        share=False,
+        debug=True
+    )

huggingface-metadata.json

@@ -0,0 +1,11 @@
+{
+  "title": "Human Hallucination Prediction",
+  "emoji": "👁️",
+  "colorFrom": "blue",
+  "colorTo": "purple",
+  "sdk": "gradio",
+  "sdk_version": "5.23.1",
+  "app_file": "app.py",
+  "pinned": false,
+  "license": "mit"
+} 

inference.py

@@ -0,0 +1,1049 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision.models as models
+import torchvision.transforms as transforms
+from torchvision.models.resnet import ResNet50_Weights
+from PIL import Image
+import numpy as np
+import os
+import requests
+import time
+import copy
+from collections import OrderedDict
+from pathlib import Path
+
+# Check for available hardware acceleration
+if torch.cuda.is_available():
+    device = torch.device("cuda")
+elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+    device = torch.device("mps")  # Use Apple Metal Performance Shaders for M-series Macs
+else:
+    device = torch.device("cpu")
+print(f"Using device: {device}")
+
+# Constants
+MODEL_URLS = {
+    # Use /resolve/main/ for raw file download; /blob/main/ returns HTML and causes "invalid load key, '<'"
+    'resnet50_robust': 'https://huggingface.co/ttoosi/robust_resnet50_eps_3.00_epoch_best/resolve/main/resnet50_robust.pt',
+    'resnet50_standard': 'https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps0.ckpt',
+    'resnet50_robust_face': 'https://huggingface.co/ttoosi/resnet50_robust_face/resolve/main/100_checkpoint.pt'
+}
+
+IMAGENET_MEAN = [0.485, 0.456, 0.406]
+IMAGENET_STD = [0.229, 0.224, 0.225]
+
+# Define the transforms based on whether normalization is on or off
+def get_transform(input_size=224, normalize=False, norm_mean=IMAGENET_MEAN, norm_std=IMAGENET_STD):
+    if normalize:
+        return transforms.Compose([
+            transforms.Resize(input_size),
+            transforms.CenterCrop(input_size),
+            transforms.ToTensor(),
+            transforms.Normalize(norm_mean, norm_std),
+        ])
+    else:
+        return transforms.Compose([
+            transforms.Resize(input_size),
+            transforms.CenterCrop(input_size),
+            transforms.ToTensor(),
+        ])
+
+# Default transform without normalization
+transform = transforms.Compose([
+    transforms.Resize(224),
+    transforms.CenterCrop(224),
+    transforms.ToTensor(),
+])
+
+normalize_transform = transforms.Normalize(IMAGENET_MEAN, IMAGENET_STD)
+
+
+def create_flat_top_gaussian_mask(image_shape, center_x, center_y, flat_radius, sigma, max_multiplier=4.0, min_multiplier=1.0, device=None):
+    """
+    Create a flat-top Gaussian epsilon/step-size mask for adaptive constraint application.
+    Coordinates are normalized: -1 = left/top edge, 1 = right/bottom edge, 0 = center.
+
+    Args:
+        image_shape: (batch, channels, height, width)
+        center_x: Center x (normalized -1 to 1)
+        center_y: Center y (normalized -1 to 1)
+        flat_radius: Radius of flat-top region (normalized)
+        sigma: Std for Gaussian fall-off outside flat region
+        max_multiplier: Multiplier inside flat region
+        min_multiplier: Multiplier far from center
+        device: torch device for the output tensor
+    Returns:
+        Tensor (1, 1, H, W)
+    """
+    _, _, H, W = image_shape
+    if device is None:
+        device = torch.device("cpu")
+    y_coords = torch.linspace(-1, 1, H, device=device)
+    x_coords = torch.linspace(-1, 1, W, device=device)
+    y_coords, x_coords = torch.meshgrid(y_coords, x_coords, indexing='ij')
+    dist = torch.sqrt((x_coords - center_x) ** 2 + (y_coords - center_y) ** 2)
+    multiplier = torch.where(
+        dist <= flat_radius,
+        torch.full_like(dist, max_multiplier),
+        min_multiplier + (max_multiplier - min_multiplier) * torch.exp(-((dist - flat_radius) ** 2) / (2 * sigma ** 2))
+    )
+    return multiplier.unsqueeze(0).unsqueeze(0)  # (1, 1, H, W)
+
+
+def extract_middle_layers(model, layer_index):
+    """
+    Extract a subset of the model up to a specific layer.
+    
+    Args:
+        model: The neural network model
+        layer_index: String 'all' for the full model, or a layer identifier (string or int)
+                    For ResNet: integers 0-8 representing specific layers
+                    For ViT: strings like 'encoder.layers.encoder_layer_3'
+    
+    Returns:
+        A modified model that outputs features from the specified layer
+    """
+    if isinstance(layer_index, str) and layer_index == 'all':
+        return model
+    
+    # Special case for ViT's encoder layers with DataParallel wrapper
+    if isinstance(layer_index, str) and layer_index.startswith('encoder.layers.encoder_layer_'):
+        try:
+            target_layer_idx = int(layer_index.split('_')[-1])
+            
+            # Create a deep copy of the model to avoid modifying the original
+            new_model = copy.deepcopy(model)
+            
+            # For models wrapped in DataParallel
+            if hasattr(new_model, 'module'):
+                # Create a subset of encoder layers up to the specified index
+                encoder_layers = nn.Sequential()
+                for i in range(target_layer_idx + 1):
+                    layer_name = f"encoder_layer_{i}"
+                    if hasattr(new_model.module.encoder.layers, layer_name):
+                        encoder_layers.add_module(layer_name, 
+                                               getattr(new_model.module.encoder.layers, layer_name))
+                
+                # Replace the encoder layers with our truncated version
+                new_model.module.encoder.layers = encoder_layers
+                
+                # Remove the heads since we're stopping at the encoder layer
+                new_model.module.heads = nn.Identity()
+                
+                return new_model
+            else:
+                # Direct model access (not DataParallel)
+                encoder_layers = nn.Sequential()
+                for i in range(target_layer_idx + 1):
+                    layer_name = f"encoder_layer_{i}"
+                    if hasattr(new_model.encoder.layers, layer_name):
+                        encoder_layers.add_module(layer_name, 
+                                               getattr(new_model.encoder.layers, layer_name))
+                
+                # Replace the encoder layers with our truncated version
+                new_model.encoder.layers = encoder_layers
+                
+                # Remove the heads since we're stopping at the encoder layer
+                new_model.heads = nn.Identity()
+                
+                return new_model
+                
+        except (ValueError, IndexError) as e:
+            raise ValueError(f"Invalid ViT layer specification: {layer_index}. Error: {e}")
+    
+    # Handling for ViT whole blocks
+    elif hasattr(model, 'blocks') or (hasattr(model, 'module') and hasattr(model.module, 'blocks')):
+        # Check for DataParallel wrapper
+        base_model = model.module if hasattr(model, 'module') else model
+        
+        # Create a deep copy to avoid modifying the original
+        new_model = copy.deepcopy(model)
+        base_new_model = new_model.module if hasattr(new_model, 'module') else new_model
+        
+        # Add the desired number of transformer blocks
+        if isinstance(layer_index, int):
+            # Truncate the blocks
+            base_new_model.blocks = base_new_model.blocks[:layer_index+1]
+            
+        return new_model
+    
+    else:
+        # Original ResNet/VGG handling
+        modules = list(model.named_children())
+        print(f"DEBUG - extract_middle_layers - Looking for '{layer_index}' in {[name for name, _ in modules]}")
+        
+        cutoff_idx = next((i for i, (name, _) in enumerate(modules) 
+                          if name == str(layer_index)), None)
+        
+        if cutoff_idx is not None:
+            # Keep modules up to and including the target
+            new_model = nn.Sequential(OrderedDict(modules[:cutoff_idx+1]))
+            return new_model
+        else:
+            raise ValueError(f"Module {layer_index} not found in model")
+
+# Get ImageNet labels
+def get_imagenet_labels():
+    url = "https://raw.githubusercontent.com/anishathalye/imagenet-simple-labels/master/imagenet-simple-labels.json"
+    response = requests.get(url)
+    if response.status_code == 200:
+        return response.json()
+    else:
+        raise RuntimeError("Failed to fetch ImageNet labels")
+
+# Download model if needed
+def download_model(model_type):
+    if model_type not in MODEL_URLS or MODEL_URLS[model_type] is None:
+        return None  # Use PyTorch's pretrained model
+    
+    # Handle special case for face model
+    if model_type == 'resnet50_robust_face':
+        model_path = Path("models/resnet50_robust_face_100_checkpoint.pt")
+    else:
+        model_path = Path(f"models/{model_type}.pt")
+    
+    if not model_path.exists():
+        print(f"Downloading {model_type} model...")
+        url = MODEL_URLS[model_type]
+        response = requests.get(url, stream=True)
+        if response.status_code == 200:
+            with open(model_path, 'wb') as f:
+                for chunk in response.iter_content(chunk_size=8192):
+                    f.write(chunk)
+            print(f"Model downloaded and saved to {model_path}")
+        else:
+            raise RuntimeError(f"Failed to download model: {response.status_code}")
+    return model_path
+
+class NormalizeByChannelMeanStd(nn.Module):
+    def __init__(self, mean, std):
+        super(NormalizeByChannelMeanStd, self).__init__()
+        if not isinstance(mean, torch.Tensor):
+            mean = torch.tensor(mean)
+        if not isinstance(std, torch.Tensor):
+            std = torch.tensor(std)
+        self.register_buffer("mean", mean)
+        self.register_buffer("std", std)
+        
+    def forward(self, tensor):
+        return self.normalize_fn(tensor, self.mean, self.std)
+    
+    def normalize_fn(self, tensor, mean, std):
+        """Differentiable version of torchvision.functional.normalize"""
+        # here we assume the color channel is at dim=1
+        mean = mean[None, :, None, None]
+        std = std[None, :, None, None]
+        return tensor.sub(mean).div(std)
+
+class InferStep:
+    def __init__(self, orig_image, eps, step_size, adaptive_eps_config=None, adaptive_step_config=None):
+        self.orig_image = orig_image
+        dev = orig_image.device
+
+        # Adaptive epsilon (spatially varying constraint)
+        if adaptive_eps_config and adaptive_eps_config.get("enabled", False):
+            self.use_adaptive_eps = True
+            base_eps = float(adaptive_eps_config.get("base_epsilon", eps))
+            self.eps_map = create_flat_top_gaussian_mask(
+                orig_image.shape,
+                float(adaptive_eps_config.get("center_x", 0.0)),
+                float(adaptive_eps_config.get("center_y", 0.0)),
+                float(adaptive_eps_config.get("flat_radius", 0.3)),
+                float(adaptive_eps_config.get("sigma", 0.2)),
+                float(adaptive_eps_config.get("max_multiplier", 4.0)),
+                float(adaptive_eps_config.get("min_multiplier", 1.0)),
+                device=dev,
+            ) * base_eps
+            self.eps_map = self.eps_map.to(dev)
+        else:
+            self.use_adaptive_eps = False
+            self.eps = eps
+
+        # Adaptive step size (spatially varying update strength)
+        if adaptive_step_config and adaptive_step_config.get("enabled", False):
+            self.use_adaptive_step_size = True
+            base_step = float(adaptive_step_config.get("base_step_size", step_size))
+            self.step_size_map = create_flat_top_gaussian_mask(
+                orig_image.shape,
+                float(adaptive_step_config.get("center_x", 0.0)),
+                float(adaptive_step_config.get("center_y", 0.0)),
+                float(adaptive_step_config.get("flat_radius", 0.3)),
+                float(adaptive_step_config.get("sigma", 0.2)),
+                float(adaptive_step_config.get("max_multiplier", 4.0)),
+                float(adaptive_step_config.get("min_multiplier", 1.0)),
+                device=dev,
+            ) * base_step
+            self.step_size_map = self.step_size_map.to(dev)
+        else:
+            self.use_adaptive_step_size = False
+            self.step_size = step_size
+
+    def project(self, x):
+        # Per-pixel L2 (color) constraint: scale diff so color_norm <= eps (or eps_map).
+        # Constraint is per-pixel L2 on RGB (spherical), not L∞ per channel.
+        diff = x - self.orig_image  # (B, C, H, W)
+        color_norm = torch.norm(diff, dim=1, keepdim=True)  # (B, 1, H, W)
+        if self.use_adaptive_eps:
+            # eps_map has shape (1, 1, H, W), broadcasts to (B, 1, H, W)
+            scale = torch.clamp(self.eps_map / (color_norm + 1e-10), max=1.0)
+        else:
+            scale = torch.clamp(self.eps / (color_norm + 1e-10), max=1.0)
+        diff = diff * scale
+        return torch.clamp(self.orig_image + diff, 0, 1)
+
+    def step(self, x, grad):
+        # Same gradient everywhere; scale magnitude by step_size (or step_size_map). Multiply, not add/subtract.
+        l = len(x.shape) - 1
+        grad_norm = torch.norm(grad.view(grad.shape[0], -1), dim=1).view(-1, *([1] * l))
+        scaled_grad = grad / (grad_norm + 1e-10)
+        if self.use_adaptive_step_size:
+            return scaled_grad * self.step_size_map
+        return scaled_grad * self.step_size
+
+def get_iterations_to_show(n_itr):
+    """Generate a dynamic list of iterations to show based on total iterations."""
+    if n_itr <= 50:
+        return [1, 5, 10, 20, 30, 40, 50, n_itr]
+    elif n_itr <= 100:
+        return [1, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, n_itr]
+    elif n_itr <= 200:
+        return [1, 5, 10, 20, 30, 40, 50, 75, 100, 125, 150, 175, 200, n_itr]
+    elif n_itr <= 500:
+        return [1, 5, 10, 20, 30, 40, 50, 75, 100, 150, 200, 250, 300, 350, 400, 450, 500, n_itr]
+    else:
+        # For very large iterations, show more evenly distributed points
+        return [1, 5, 10, 20, 30, 40, 50, 75, 100, 150, 200, 250, 300, 350, 400, 450, 500, 
+                int(n_itr*0.6), int(n_itr*0.7), int(n_itr*0.8), int(n_itr*0.9), n_itr]
+
+def get_inference_configs(inference_type='IncreaseConfidence', eps=0.5, n_itr=50, step_size=1.0):
+    """Generate inference configuration with customizable parameters.
+    
+    Args:
+        inference_type (str): Type of inference ('IncreaseConfidence' or 'Prior-Guided Drift Diffusion')
+        eps (float): Maximum perturbation size
+        n_itr (int): Number of iterations
+        step_size (float): Step size for each iteration
+    """
+    
+    # Base configuration common to all inference types
+    config = {
+        'loss_infer': inference_type,  # How to guide the optimization
+        'n_itr': n_itr,  # Number of iterations
+        'eps': eps,  # Maximum perturbation size
+        'step_size': step_size,  # Step size for each iteration
+        'diffusion_noise_ratio': 0.0,  # No diffusion noise
+        'initial_inference_noise_ratio': 0.0,  # No initial noise
+        'top_layer': 'all',  # Use all layers of the model
+        'inference_normalization': 'off',  # 'on' or 'off' (match psychiatry implementation)
+        'recognition_normalization': 'off',
+        'iterations_to_show': get_iterations_to_show(n_itr),  # Dynamic iterations to visualize
+        'misc_info': {'keep_grads': False},  # Additional configuration
+        'biased_inference': {'enable': False, 'class': None}  # Bias perception toward a target class (ImageNet label)
+    }
+    
+    # Customize based on inference type
+    if inference_type == 'IncreaseConfidence':
+        config['loss_function'] = 'CE'  # Cross Entropy
+    
+    elif inference_type == 'Prior-Guided Drift Diffusion':
+        config['loss_function'] = 'MSE'  # Mean Square Error
+        config['initial_inference_noise_ratio'] = 0.05  # Initial noise for diffusion
+        config['diffusion_noise_ratio'] = 0.01  # Add noise during diffusion
+    
+    elif inference_type == 'GradModulation':
+        config['loss_function'] = 'CE'  # Cross Entropy
+        config['misc_info']['grad_modulation'] = 0.5  # Gradient modulation strength
+    
+    elif inference_type == 'CompositionalFusion':
+        config['loss_function'] = 'CE'  # Cross Entropy
+        config['misc_info']['positive_classes'] = []  # Classes to maximize
+        config['misc_info']['negative_classes'] = []  # Classes to minimize
+    
+    return config
+
+class GenerativeInferenceModel:
+    def __init__(self):
+        self.models = {}
+        self.normalizer = NormalizeByChannelMeanStd(IMAGENET_MEAN, IMAGENET_STD).to(device)
+        self.labels = get_imagenet_labels()
+        
+    def verify_model_integrity(self, model, model_type):
+        """
+        Verify model integrity by running a test input through it.
+        Returns whether the model passes basic integrity check.
+        """
+        try:
+            print(f"\n=== Running model integrity check for {model_type} ===")
+            # Create a deterministic test input directly on the correct device
+            test_input = torch.zeros(1, 3, 224, 224, device=device)
+            test_input[0, 0, 100:124, 100:124] = 0.5  # Red square
+            
+            # Run forward pass
+            with torch.no_grad():
+                output = model(test_input)
+            
+            # Check output shape
+            if output.shape != (1, 1000):
+                print(f"❌ Unexpected output shape: {output.shape}, expected (1, 1000)")
+                return False
+                
+            # Get top prediction
+            probs = torch.nn.functional.softmax(output, dim=1)
+            confidence, prediction = torch.max(probs, 1)
+            
+            # Calculate basic statistics on output
+            mean = output.mean().item()
+            std = output.std().item()
+            min_val = output.min().item()
+            max_val = output.max().item()
+            
+            print(f"Model integrity check results:")
+            print(f"- Output shape: {output.shape}")
+            print(f"- Top prediction: Class {prediction.item()} with {confidence.item()*100:.2f}% confidence")
+            print(f"- Output statistics: mean={mean:.3f}, std={std:.3f}, min={min_val:.3f}, max={max_val:.3f}")
+            
+            # Basic sanity checks
+            if torch.isnan(output).any():
+                print("❌ Model produced NaN outputs")
+                return False
+                
+            if output.std().item() < 0.1:
+                print("⚠️ Low output variance, model may not be discriminative")
+                
+            print("✅ Model passes basic integrity check")
+            return True
+            
+        except Exception as e:
+            print(f"❌ Model integrity check failed with error: {e}")
+            # Rather than failing completely, we'll continue
+            return True
+    
+    def load_model(self, model_type):
+        """Load model from checkpoint or use pretrained model."""
+        if model_type in self.models:
+            print(f"Using cached {model_type} model")
+            return self.models[model_type]
+        
+        # Record loading time for performance analysis
+        start_time = time.time()
+        model_path = download_model(model_type)
+        
+        # Create a sequential model with normalizer and ResNet50
+        resnet = models.resnet50()
+        model = nn.Sequential(
+            self.normalizer,  # Normalizer is part of the model sequence
+            resnet
+        )
+        
+        # Load the model checkpoint
+        if model_path:
+            print(f"Loading {model_type} model from {model_path}...")
+            try:
+                checkpoint = torch.load(model_path, map_location=device, weights_only=False)
+                
+                # Print checkpoint structure for better understanding
+                print("\n=== Analyzing checkpoint structure ===")
+                if isinstance(checkpoint, dict):
+                    print(f"Checkpoint contains keys: {list(checkpoint.keys())}")
+                    
+                    # Examine 'model' structure if it exists
+                    if 'model' in checkpoint and isinstance(checkpoint['model'], dict):
+                        model_dict = checkpoint['model']
+                        # Get sample of keys to understand structure
+                        first_keys = list(model_dict.keys())[:5]
+                        print(f"'model' contains keys like: {first_keys}")
+                        
+                        # Check for common prefixes in the model dict
+                        prefixes = set()
+                        for key in list(model_dict.keys())[:100]:  # Check first 100 keys
+                            parts = key.split('.')
+                            if len(parts) > 1:
+                                prefixes.add(parts[0])
+                        if prefixes:
+                            print(f"Common prefixes in model dict: {prefixes}")
+                else:
+                    print(f"Checkpoint is not a dictionary, but a {type(checkpoint)}")
+                
+                # Handle different checkpoint formats
+                if 'model' in checkpoint:
+                    # Format from madrylab robust models
+                    state_dict = checkpoint['model']
+                    print("Using 'model' key from checkpoint")
+                elif 'state_dict' in checkpoint:
+                    state_dict = checkpoint['state_dict']
+                    print("Using 'state_dict' key from checkpoint")
+                else:
+                    # Direct state dict
+                    state_dict = checkpoint
+                    print("Using checkpoint directly as state_dict")
+                
+                # Handle prefix in state dict keys for ResNet part
+                resnet_state_dict = {}
+                prefixes_to_try = ['', 'module.', 'model.', 'attacker.model.']
+                resnet_keys = set(resnet.state_dict().keys())
+                
+                # First check if we can find keys directly in the attacker.model path
+                print("\n=== Phase 1: Checking for specific model structures ===")
+                
+                # Check for 'module.model' structure (seen in actual checkpoint)
+                module_model_keys = [key for key in state_dict.keys() if key.startswith('module.model.')]
+                if module_model_keys:
+                    print(f"Found 'module.model' structure with {len(module_model_keys)} parameters")
+                    # Extract all parameters from module.model
+                    for source_key, value in state_dict.items():
+                        if source_key.startswith('module.model.'):
+                            target_key = source_key[len('module.model.'):]
+                            resnet_state_dict[target_key] = value
+                            
+                    print(f"Extracted {len(resnet_state_dict)} parameters from module.model")
+                
+                # Check for 'attacker.model' structure
+                attacker_model_keys = [key for key in state_dict.keys() if key.startswith('attacker.model.')]
+                if attacker_model_keys:
+                    print(f"Found 'attacker.model' structure with {len(attacker_model_keys)} parameters")
+                    # Extract all parameters from attacker.model
+                    for source_key, value in state_dict.items():
+                        if source_key.startswith('attacker.model.'):
+                            target_key = source_key[len('attacker.model.'):]
+                            resnet_state_dict[target_key] = value
+                            
+                    print(f"Extracted {len(resnet_state_dict)} parameters from attacker.model")
+                    
+                    # Check if 'model' (not attacker.model) exists as a fallback
+                    model_keys = [key for key in state_dict.keys() if key.startswith('model.') and not key.startswith('attacker.model.')]
+                    if model_keys and len(resnet_state_dict) < len(resnet_keys):
+                        print(f"Found additional 'model.' structure with {len(model_keys)} parameters")
+                        # Try to complete missing parameters
+                        for source_key, value in state_dict.items():
+                            if source_key.startswith('model.'):
+                                target_key = source_key[len('model.'):]
+                                if target_key in resnet_keys and target_key not in resnet_state_dict:
+                                    resnet_state_dict[target_key] = value
+                                    
+                else:
+                    # Check for other known structures
+                    structure_found = False
+                    
+                    # Check for 'model.' prefix
+                    model_keys = [key for key in state_dict.keys() if key.startswith('model.')]
+                    if model_keys:
+                        print(f"Found 'model.' structure with {len(model_keys)} parameters")
+                        for source_key, value in state_dict.items():
+                            if source_key.startswith('model.'):
+                                target_key = source_key[len('model.'):]
+                                resnet_state_dict[target_key] = value
+                        structure_found = True
+                    
+                    # Check for ResNet parameters at the top level
+                    top_level_resnet_keys = 0
+                    for key in resnet_keys:
+                        if key in state_dict:
+                            top_level_resnet_keys += 1
+                            
+                    if top_level_resnet_keys > 0:
+                        print(f"Found {top_level_resnet_keys} ResNet parameters at top level")
+                        for target_key in resnet_keys:
+                            if target_key in state_dict:
+                                resnet_state_dict[target_key] = state_dict[target_key]
+                        structure_found = True
+                    
+                    # If no structure was recognized, try the prefix mapping approach
+                    if not structure_found:
+                        print("No standard model structure found, trying prefix mappings...")
+                        for target_key in resnet_keys:
+                            for prefix in prefixes_to_try:
+                                source_key = prefix + target_key
+                                if source_key in state_dict:
+                                    resnet_state_dict[target_key] = state_dict[source_key]
+                                    break
+                
+                # If we still can't find enough keys, try a final approach of removing prefixes
+                if len(resnet_state_dict) < len(resnet_keys):
+                    print(f"Found only {len(resnet_state_dict)}/{len(resnet_keys)} parameters, trying prefix removal...")
+                    
+                    # Track matches found through prefix removal
+                    prefix_matches = {prefix: 0 for prefix in ['module.', 'model.', 'attacker.model.', 'attacker.']}
+                    layer_matches = {}  # Track matches by layer type
+                    
+                    # Count parameter keys by layer type for analysis
+                    for key in resnet_keys:
+                        layer_name = key.split('.')[0] if '.' in key else key
+                        if layer_name not in layer_matches:
+                            layer_matches[layer_name] = {'total': 0, 'matched': 0}
+                        layer_matches[layer_name]['total'] += 1
+                    
+                    # Try keys with common prefixes
+                    for source_key, value in state_dict.items():
+                        # Skip if already found
+                        target_key = source_key
+                        matched_prefix = None
+                        
+                        # Try removing various prefixes
+                        for prefix in ['module.', 'model.', 'attacker.model.', 'attacker.']:
+                            if source_key.startswith(prefix):
+                                target_key = source_key[len(prefix):]
+                                matched_prefix = prefix
+                                break
+                        
+                        # If the target key is in the ResNet keys, add it to the state dict
+                        if target_key in resnet_keys and target_key not in resnet_state_dict:
+                            resnet_state_dict[target_key] = value
+                            
+                            # Update match statistics
+                            if matched_prefix:
+                                prefix_matches[matched_prefix] += 1
+                            
+                            # Update layer matches
+                            layer_name = target_key.split('.')[0] if '.' in target_key else target_key
+                            if layer_name in layer_matches:
+                                layer_matches[layer_name]['matched'] += 1
+                    
+                    # Print detailed prefix removal statistics
+                    print("\n=== Prefix Removal Statistics ===")
+                    total_matches = sum(prefix_matches.values())
+                    print(f"Total parameters matched through prefix removal: {total_matches}/{len(resnet_keys)} ({(total_matches/len(resnet_keys))*100:.1f}%)")
+                    
+                    # Show matches by prefix
+                    print("\nMatches by prefix:")
+                    for prefix, count in sorted(prefix_matches.items(), key=lambda x: x[1], reverse=True):
+                        if count > 0:
+                            print(f"  {prefix}: {count} parameters")
+                    
+                    # Show matches by layer type
+                    print("\nMatches by layer type:")
+                    for layer, stats in sorted(layer_matches.items(), key=lambda x: x[1]['total'], reverse=True):
+                        match_percent = (stats['matched'] / stats['total']) * 100 if stats['total'] > 0 else 0
+                        print(f"  {layer}: {stats['matched']}/{stats['total']} ({match_percent:.1f}%)")
+                    
+                    # Check for specific important layers (conv1, layer1, etc.)
+                    critical_layers = ['conv1', 'bn1', 'layer1', 'layer2', 'layer3', 'layer4', 'fc']
+                    print("\nStatus of critical layers:")
+                    for layer in critical_layers:
+                        if layer in layer_matches:
+                            match_percent = (layer_matches[layer]['matched'] / layer_matches[layer]['total']) * 100
+                            status = "✅ COMPLETE" if layer_matches[layer]['matched'] == layer_matches[layer]['total'] else "⚠️ INCOMPLETE"
+                            print(f"  {layer}: {layer_matches[layer]['matched']}/{layer_matches[layer]['total']} ({match_percent:.1f}%) - {status}")
+                        else:
+                            print(f"  {layer}: Not found in model")
+                
+                # Load the ResNet state dict
+                if resnet_state_dict:
+                    try:
+                        # Use strict=False to allow missing keys
+                        result = resnet.load_state_dict(resnet_state_dict, strict=False)
+                        missing_keys, unexpected_keys = result
+                        
+                        # Generate detailed information with better formatting
+                        loading_report = []
+                        loading_report.append(f"\n===== MODEL LOADING REPORT: {model_type} =====")
+                        loading_report.append(f"Total parameters in checkpoint: {len(resnet_state_dict):,}")
+                        loading_report.append(f"Total parameters in model: {len(resnet.state_dict()):,}")
+                        loading_report.append(f"Missing keys: {len(missing_keys):,} parameters")
+                        loading_report.append(f"Unexpected keys: {len(unexpected_keys):,} parameters")
+
+                        # Calculate percentage of parameters loaded
+                        loaded_keys = set(resnet_state_dict.keys()) - set(unexpected_keys)
+                        loaded_percent = (len(loaded_keys) / len(resnet.state_dict())) * 100
+                        
+                        # Determine loading success status
+                        if loaded_percent >= 99.5:
+                            status = "✅ COMPLETE - All important parameters loaded"
+                        elif loaded_percent >= 90:
+                            status = "🟡 PARTIAL - Most parameters loaded, should still function"
+                        elif loaded_percent >= 50:
+                            status = "⚠️ INCOMPLETE - Many parameters missing, may not function properly"
+                        else:
+                            status = "❌ FAILED - Critical parameters missing, will not function properly"
+                            
+                        loading_report.append(f"Successfully loaded: {len(loaded_keys):,} parameters ({loaded_percent:.1f}%)")
+                        loading_report.append(f"Loading status: {status}")
+                        
+                        # If loading is severely incomplete, fall back to PyTorch's pretrained model
+                        if loaded_percent < 50:
+                            loading_report.append("\n⚠️ WARNING: Loading from checkpoint is too incomplete.")
+                            loading_report.append("⚠️ Falling back to PyTorch's pretrained model to avoid broken inference.")
+                            
+                            # Create a new ResNet model with pretrained weights
+                            resnet = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
+                            model = nn.Sequential(self.normalizer, resnet)
+                            loading_report.append("✅ Successfully loaded PyTorch's pretrained ResNet50 model")
+                        
+                        # Show missing keys by layer type
+                        if missing_keys:
+                            loading_report.append("\nMissing keys by layer type:")
+                            layer_types = {}
+                            for key in missing_keys:
+                                # Extract layer type (e.g., 'conv', 'bn', 'layer1', etc.)
+                                parts = key.split('.')
+                                if len(parts) > 0:
+                                    layer_type = parts[0]
+                                    if layer_type not in layer_types:
+                                        layer_types[layer_type] = 0
+                                    layer_types[layer_type] += 1
+                            
+                            # Add counts by layer type
+                            for layer_type, count in sorted(layer_types.items(), key=lambda x: x[1], reverse=True):
+                                loading_report.append(f"  {layer_type}: {count:,} parameters")
+                                
+                            loading_report.append("\nFirst 10 missing keys:")
+                            for i, key in enumerate(sorted(missing_keys)[:10]):
+                                loading_report.append(f"  {i+1}. {key}")
+                        
+                        # Show unexpected keys if any
+                        if unexpected_keys:
+                            loading_report.append("\nFirst 10 unexpected keys:")
+                            for i, key in enumerate(sorted(unexpected_keys)[:10]):
+                                loading_report.append(f"  {i+1}. {key}")
+                                
+                        loading_report.append("========================================")
+                        
+                        # Convert report to string and print it
+                        report_text = "\n".join(loading_report)
+                        print(report_text)
+                        
+                        # Also save to a file for reference
+                        os.makedirs("logs", exist_ok=True)
+                        with open(f"logs/model_loading_{model_type}.log", "w") as f:
+                            f.write(report_text)
+                        
+                        # Look for normalizer parameters as well
+                        if any(key.startswith('attacker.normalize.') for key in state_dict.keys()):
+                            norm_state_dict = {}
+                            for key, value in state_dict.items():
+                                if key.startswith('attacker.normalize.'):
+                                    norm_key = key[len('attacker.normalize.'):]
+                                    norm_state_dict[norm_key] = value
+                            
+                            if norm_state_dict:
+                                try:
+                                    self.normalizer.load_state_dict(norm_state_dict, strict=False)
+                                    print("Successfully loaded normalizer parameters")
+                                except Exception as e:
+                                    print(f"Warning: Could not load normalizer parameters: {e}")
+                    except Exception as e:
+                        print(f"Warning: Error loading ResNet parameters: {e}")
+                        # Fall back to loading without normalizer
+                        model = resnet  # Use just the ResNet model without normalizer
+            except Exception as e:
+                print(f"Error loading model checkpoint: {e}")
+                # Fallback to PyTorch's pretrained model
+                print("Falling back to PyTorch's pretrained model")
+                resnet = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
+                model = nn.Sequential(self.normalizer, resnet)
+        else:
+            # Fallback to PyTorch's pretrained model
+            print("No checkpoint available, using PyTorch's pretrained model")
+            resnet = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
+            model = nn.Sequential(self.normalizer, resnet)
+        
+        model = model.to(device)
+        model.eval()  # Set to evaluation mode
+        
+        # Verify model integrity
+        self.verify_model_integrity(model, model_type)
+        
+        # Store the model for future use
+        self.models[model_type] = model
+        end_time = time.time()
+        load_time = end_time - start_time
+        print(f"Model {model_type} loaded in {load_time:.2f} seconds")
+        return model
+    
+    def inference(self, image, model_type, config):
+        """Run generative inference on the image."""
+        # Time the entire inference process
+        inference_start = time.time()
+        
+        # Load model if not already loaded
+        model = self.load_model(model_type)
+        
+        # Check if image is a file path
+        if isinstance(image, str):
+            if os.path.exists(image):
+                image = Image.open(image).convert('RGB')
+            else:
+                raise ValueError(f"Image path does not exist: {image}")
+        elif isinstance(image, torch.Tensor):
+            raise ValueError(f"Image type {type(image)}, looks like already a transformed tensor")
+        
+        # Prepare image tensor - match original code's conditional transform
+        load_start = time.time()
+        use_norm = config['inference_normalization'] == 'on'
+        custom_transform = get_transform(
+            input_size=224, 
+            normalize=use_norm, 
+            norm_mean=IMAGENET_MEAN, 
+            norm_std=IMAGENET_STD
+        )
+        
+        # Special handling for GradModulation as in original
+        if config['loss_infer'] == 'GradModulation' and 'misc_info' in config and 'grad_modulation' in config['misc_info']:
+            grad_modulation = config['misc_info']['grad_modulation']
+            image_tensor = custom_transform(image).unsqueeze(0).to(device)
+            image_tensor = image_tensor * (1-grad_modulation) + grad_modulation * torch.randn_like(image_tensor).to(device)
+        else:
+            image_tensor = custom_transform(image).unsqueeze(0).to(device)
+            
+        image_tensor.requires_grad = True
+        print(f"Image loaded and processed in {time.time() - load_start:.2f} seconds")
+        
+        # Check model structure
+        is_sequential = isinstance(model, nn.Sequential)
+        
+        # Get original predictions
+        with torch.no_grad():
+            # If the model is sequential with a normalizer, skip the normalization step
+            if is_sequential and isinstance(model[0], NormalizeByChannelMeanStd):
+                print("Model is sequential with normalization")
+                # Get the core model part (typically at index 1 in Sequential)
+                core_model = model[1]
+                if config['inference_normalization'] == 'on':
+                    output_original = model(image_tensor)  # Model includes normalization
+                else:
+                    output_original = core_model(image_tensor)  # Model includes normalization
+                
+            else:
+                print("Model is not sequential with normalization")
+                # Use manual normalization for non-sequential models
+                if config['inference_normalization'] == 'on':
+                    normalized_tensor = normalize_transform(image_tensor)
+                    output_original = model(normalized_tensor)
+                else:
+                    output_original = model(image_tensor)
+                core_model = model
+            
+            probs_orig = F.softmax(output_original, dim=1)
+            conf_orig, classes_orig = torch.max(probs_orig, 1)
+            
+            # Get least confident classes
+            _, least_confident_classes = torch.topk(probs_orig, k=100, largest=False)
+        
+        # Initialize inference step (adaptive Gaussian mask from config)
+        adaptive_eps = config.get("adaptive_epsilon")
+        adaptive_step = config.get("adaptive_step_size")
+        if adaptive_eps and adaptive_eps.get("enabled"):
+            print(f"Applying adaptive epsilon mask: center=({adaptive_eps.get('center_x'):.2f}, {adaptive_eps.get('center_y'):.2f}), radius={adaptive_eps.get('flat_radius'):.2f}")
+        if adaptive_step and adaptive_step.get("enabled"):
+            print(f"Applying adaptive step-size mask: center=({adaptive_step.get('center_x'):.2f}, {adaptive_step.get('center_y'):.2f}), radius={adaptive_step.get('flat_radius'):.2f}")
+        infer_step = InferStep(
+            image_tensor,
+            config["eps"],
+            config["step_size"],
+            adaptive_eps_config=adaptive_eps,
+            adaptive_step_config=adaptive_step,
+        )
+        
+        # Biased inference: resolve class name to index (ImageNet simple labels, case-insensitive)
+        biased_inference_config = config.get('biased_inference', {'enable': False, 'class': None})
+        biased_class_index = None
+        biased_class_tensor = None
+        if biased_inference_config.get('enable', False):
+            class_name = biased_inference_config.get('class') or biased_inference_config.get('class_name')
+            if class_name:
+                try:
+                    biased_class_index = next(
+                        i for i, label in enumerate(self.labels)
+                        if label.lower() == class_name.lower()
+                    )
+                    biased_class_tensor = torch.tensor(
+                        [biased_class_index], device=device, dtype=torch.long
+                    )
+                    print(f"Biased inference: biasing toward class '{self.labels[biased_class_index]}' (index {biased_class_index})")
+                except StopIteration:
+                    raise ValueError(
+                        f"biased_inference class '{class_name}' not found in ImageNet simple labels. "
+                        "Use a label from imagenet-simple-labels (e.g. 'goldfish', 'tabby cat')."
+                    )
+            else:
+                print("Biased inference enabled but no class specified; ignoring.")
+        
+        # Storage for inference steps
+        # Create a new tensor that requires gradients
+        x = image_tensor.clone().detach().requires_grad_(True)
+        all_steps = [image_tensor[0].detach().cpu()]
+        
+        # For Prior-Guided Drift Diffusion, extract selected layer and initialize with noisy features
+        noisy_features = None
+        layer_model = None
+        if config['loss_infer'] == 'Prior-Guided Drift Diffusion':
+            print(f"Setting up Prior-Guided Drift Diffusion with layer {config['top_layer']} and noise {config['initial_inference_noise_ratio']}...")
+            
+            # Extract model up to the specified layer
+            try:
+                # Start by finding the actual model to use
+                base_model = model
+                
+                # Handle DataParallel wrapper if present
+                if hasattr(base_model, 'module'):
+                    base_model = base_model.module
+                
+                # Log the initial model structure
+                print(f"DEBUG - Initial model structure: {type(base_model)}")
+                
+                # If we have a Sequential model (which is likely our normalizer + model structure)
+                if isinstance(base_model, nn.Sequential):
+                    print(f"DEBUG - Sequential model with {len(list(base_model.children()))} children")
+                    
+                    # If this is our NormalizeByChannelMeanStd + ResNet pattern
+                    if len(list(base_model.children())) >= 2:
+                        # The actual ResNet model is the second component (index 1)
+                        actual_model = list(base_model.children())[1]
+                        print(f"DEBUG - Using ResNet component: {type(actual_model)}")
+                        print(f"DEBUG - Available layers: {[name for name, _ in actual_model.named_children()]}")
+                        
+                        # Extract from the actual ResNet
+                        layer_model = extract_middle_layers(actual_model, config['top_layer'])
+                    else:
+                        # Just a single component Sequential
+                        layer_model = extract_middle_layers(base_model, config['top_layer'])
+                else:
+                    # Not Sequential, might be direct model
+                    print(f"DEBUG - Available layers: {[name for name, _ in base_model.named_children()]}")
+                    layer_model = extract_middle_layers(base_model, config['top_layer'])
+                
+                print(f"Successfully extracted model up to layer: {config['top_layer']}")
+            except ValueError as e:
+                print(f"Layer extraction failed: {e}. Using full model.")
+                layer_model = model
+            
+            # Add noise to the image - exactly match original code
+            added_noise = config['initial_inference_noise_ratio'] * torch.randn_like(image_tensor).to(device)
+            noisy_image_tensor = image_tensor + added_noise
+            
+            # Compute noisy features - simplified to match original code
+            noisy_features = layer_model(noisy_image_tensor)
+            
+            print(f"Noisy features computed for Prior-Guided Drift Diffusion target with shape: {noisy_features.shape if hasattr(noisy_features, 'shape') else 'unknown'}")
+        
+        # Main inference loop
+        print(f"Starting inference loop with {config['n_itr']} iterations for {config['loss_infer']}...")
+        loop_start = time.time()
+        for i in range(config['n_itr']):
+            # Reset gradients
+            x.grad = None
+            
+            # Forward pass - use layer_model for Prior-Guided Drift Diffusion, full model otherwise
+            if config['loss_infer'] == 'Prior-Guided Drift Diffusion' and layer_model is not None:
+                # Use the extracted layer model for Prior-Guided Drift Diffusion
+                # In original code, normalization is handled at transform time, not during forward pass
+                output = layer_model(x)
+            else:
+                # Standard forward pass with full model
+                # Simplified to match original code's approach
+                output = model(x)
+            
+            # Calculate loss and gradients based on inference type
+            try:
+                if config['loss_infer'] == 'Prior-Guided Drift Diffusion':
+                    # Use MSE loss to match the noisy features
+                    assert config['loss_function'] == 'MSE', "Reverse Diffusion loss function must be MSE"
+                    if noisy_features is not None:
+                        loss = F.mse_loss(output, noisy_features)
+                        grad = torch.autograd.grad(loss, x)[0]  # Removed retain_graph=True to match original
+                    else:
+                        raise ValueError("Noisy features not computed for Prior-Guided Drift Diffusion")
+                
+                else:  # Default 'IncreaseConfidence' approach
+                    # Get the least confident classes
+                    num_classes = min(10, least_confident_classes.size(1))
+                    target_classes = least_confident_classes[0, :num_classes]
+                    
+                    # Create targets for least confident classes
+                    targets = torch.tensor([idx.item() for idx in target_classes], device=device)
+                    
+                    # Use a combined loss to increase confidence
+                    loss = 0
+                    for target in targets:
+                        # Create one-hot target
+                        one_hot = torch.zeros_like(output)
+                        one_hot[0, target] = 1
+                        # Use loss to maximize confidence
+                        loss = loss + F.mse_loss(F.softmax(output, dim=1), one_hot)
+                    
+                    grad = torch.autograd.grad(loss, x, retain_graph=True)[0]
+                
+                # Biased inference: subtract gradient that pushes toward target class
+                if biased_inference_config.get('enable', False) and biased_class_tensor is not None:
+                    output_full = model(x)
+                    loss_biased = F.cross_entropy(output_full, biased_class_tensor)
+                    grad_biased = torch.autograd.grad(loss_biased, x)[0]
+                    grad = grad - grad_biased
+                
+                if grad is None:
+                    print("Warning: Direct gradient calculation failed")
+                    # Fall back to random perturbation
+                    random_noise = (torch.rand_like(x) - 0.5) * 2 * config['step_size']
+                    x = infer_step.project(x + random_noise)
+                else:
+                    # Update image with gradient - do this exactly as in original code
+                    adjusted_grad = infer_step.step(x, grad)
+                    
+                    # Add diffusion noise if specified
+                    diffusion_noise = config['diffusion_noise_ratio'] * torch.randn_like(x).to(device)
+                    
+                    # Apply gradient and noise in one operation before projecting, exactly as in original
+                    x = infer_step.project(x.clone() + adjusted_grad + diffusion_noise)
+                    
+            except Exception as e:
+                print(f"Error in gradient calculation: {e}")
+                # Fall back to random perturbation - match original code
+                random_noise = (torch.rand_like(x) - 0.5) * 2 * config['step_size']
+                x = infer_step.project(x.clone() + random_noise)
+            
+            # Store step if in iterations_to_show
+            if i+1 in config['iterations_to_show'] or i+1 == config['n_itr']:
+                all_steps.append(x[0].detach().cpu())
+        
+        # Print some info about the inference
+        with torch.no_grad():
+            if is_sequential and isinstance(model[0], NormalizeByChannelMeanStd):
+                if config['inference_normalization'] == 'on':
+                    final_output = model(x)
+                else:
+                    final_output = core_model(x)
+            else:
+                if config['inference_normalization'] == 'on':
+                    normalized_x = normalize_transform(x)
+                    final_output = model(normalized_x)
+                else:
+                    final_output = model(x)
+                
+            final_probs = F.softmax(final_output, dim=1)
+            final_conf, final_classes = torch.max(final_probs, 1)
+            
+            # Calculate timing information
+            loop_time = time.time() - loop_start
+            total_time = time.time() - inference_start
+            avg_iter_time = loop_time / config['n_itr'] if config['n_itr'] > 0 else 0
+            
+            print(f"Original top class: {classes_orig.item()} ({conf_orig.item():.4f})")
+            print(f"Final top class: {final_classes.item()} ({final_conf.item():.4f})")
+            print(f"Inference loop completed in {loop_time:.2f} seconds ({avg_iter_time:.4f} sec/iteration)")
+            print(f"Total inference time: {total_time:.2f} seconds")
+            
+        # Return results in format compatible with both old and new code
+        return {
+            'final_image': x[0].detach().cpu(),
+            'steps': all_steps,
+            'original_class': classes_orig.item(),
+            'original_confidence': conf_orig.item(),
+            'final_class': final_classes.item(),
+            'final_confidence': final_conf.item()
+        }
+
+# Utility function to show inference steps
+def show_inference_steps(steps, figsize=(15, 10)):
+    import matplotlib.pyplot as plt
+    
+    n_steps = len(steps)
+    fig, axes = plt.subplots(1, n_steps, figsize=figsize)
+    
+    for i, step_img in enumerate(steps):
+        img = step_img.permute(1, 2, 0).numpy()
+        axes[i].imshow(img)
+        axes[i].set_title(f"Step {i}")
+        axes[i].axis('off')
+    
+    plt.tight_layout()
+    return fig 

requirements.txt

@@ -0,0 +1,9 @@
+torch
+torchvision
+numpy
+pillow
+gradio
+matplotlib
+requests
+tqdm
+huggingface_hub