# Requirements Specification for Future Secondary Detector Integration **Target Space**: [`https://huggingface.co/spaces/Thiru1122/Multimodel_setup`](https://huggingface.co/spaces/Thiru1122/Multimodel_setup) **Purpose**: Defines the mandatory technical specifications and artifacts required before any secondary detector (e.g. ViT, RAID, or other architecture) can be integrated into the multi-model benchmark framework. --- ## Required Specification Checklist To prevent integration failures, unverified assumptions, or runtime crashes, any candidate detector proposed for integration must provide complete, verified answers to the following 14 requirements: ### 1. Repository & Code Source - Exact public GitHub or Hugging Face repository URL. - Specific commit SHA or release tag to pin for reproducibility. ### 2. Software & Model License - Formal license type (e.g., MIT, Apache 2.0, Non-Commercial CC-BY-NC). - Confirmation of compatibility with Hugging Face Spaces hosting. ### 3. Model Architecture & Backbone - Base neural network architecture (e.g., `google/vit-base-patch16-224`, `timm/vit_large_patch14_224`, `ResNet50`). - Any custom classification neck, pooling layer, or multi-scale projection heads. ### 4. Checkpoint Location & Integrity - Accessible public download URL or Hugging Face model hub identifier (`repo_id`). - SHA256 checksum of the weight file. - Exact filename and format (e.g., `pytorch_model.bin`, `model.safetensors`, `model.pth`). - Confirmation whether the checkpoint contains the full model or classification head only. ### 5. Input Modality - Explicit confirmation that the model evaluates single static images (or frame sequences). ### 6. Input Dimensions & Aspect Ratio - Target tensor resolution (e.g., $224 \times 224$, $384 \times 384$, or variable). - Aspect ratio handling (e.g., isotropic scaling with padding vs. anisotropic bicubic resize vs. center cropping). ### 7. Preprocessing Transform Pipeline - Exact color space (RGB, BGR, or Grayscale). - Resizing interpolation mode (Bilinear, Bicubic, Lanczos). - Cropping strategy (Center crop, random crop, or no crop). - Tensor scaling (range `[0.0, 1.0]` vs. `[-1.0, 1.0]` vs. `[0, 255]`). - Exact channel mean and standard deviation vectors for normalization. ### 8. Output Format & Tensor Shape - Number of output logits: - Binary single scalar logit (`[batch, 1]`), OR - Two-class logits (`[batch, 2]` for `[real_logit, fake_logit]`), OR - Softmax probability distribution. ### 9. Label Mapping - Explicit class indexing: - Which index/sign corresponds to **REAL** (e.g., index `0` or negative logit). - Which index/sign corresponds to **FAKE** (e.g., index `1` or positive logit). ### 10. Operational Decision Threshold - Author-defined threshold value (e.g., $0.50$, $0.65$, or logit $0.00$). - Calibration details (whether Platt scaling or temperature scaling is required). ### 11. Confidence Calculation Formula - Mathematical formula used to compute user-facing confidence percentage from logits/probabilities. ### 12. Device & Memory Footprint - Peak GPU VRAM required during batch=1 inference. - Host system RAM required to load weights into memory. - Compatibility with Hugging Face ZeroGPU (`zero-a10g`) and PyTorch CPU execution. ### 13. Inference Function & Entry Point - Python function signature for single-image execution (e.g., `predict(image: PIL.Image) -> float`). - Whether model loading can be decoupled and executed lazily once at startup. ### 14. Dependency Requirements & Conflicts - Required Python library versions (`transformers`, `timm`, `torchvision`, etc.). - Confirmation of no conflicting dependencies with existing UniversalFakeDetect requirements. --- ## 15. Labeled Benchmark Evaluation Data Before performing comparative benchmarking, a curated set of test images with documentary proof of ground-truth origin must be supplied: - Verified natural photographs (diverse camera sensors, lighting, and subjects). - Verified legacy GAN synthetic images (ProGAN, StyleGAN2/3). - Verified modern diffusion synthetic images (Google Imagen, Stable Diffusion XL, Midjourney v6, Flux). - Localized image manipulations (splicing, inpainting).