FIX.md

FIX.md — How to Stop All Startup Errors

The logs show the OLD engine code is still running. The files from the previous session were not copied into the project. Do these steps in order.

---

Step 1 — Replace fingerprint engine

Copy fingerprint_engine.py (from outputs) to:

src/engines/fingerprint/engine.py

This removes all broken legacy models and fixes the trust_remote_code duplicate kwarg bug.

Replaces with 3 working models:

• Organika/sdxl-detector
• haywoodsloan/ai-image-detector-deploy
• dima806/deepfake_vs_real_image_detection

---

Step 2 — Replace coherence engine

Copy coherence_engine.py (from outputs) to:

src/engines/coherence/engine.py

This removes the broken wav2vec-based audio model path with incompatible weights and random output behavior. Coherence now runs visual-only (FaceNet + MediaPipe).

---

Step 3 — Replace SSTGNN engine

Copy sstgnn_engine.py (from outputs) to:

src/engines/sstgnn/engine.py

Removes legacy ViT .pt checkpoint usage. Uses dima806 + prithivMLmods only.

---

Step 4 — Fix the Dockerfile (libGLESv2 error)

Replace your Dockerfile with this exactly:

FROM python:3.11-slim

RUN apt-get update && apt-get install -y --no-install-recommends \
    ffmpeg \
    libgl1 \
    libglib2.0-0 \
    libsm6 \
    libxext6 \
    libxrender1 \
    libgles2 \
    libegl1 \
    libgbm1 \
    libgomp1 \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /app

COPY requirements.txt .
RUN pip install --no-cache-dir --upgrade pip && \
    pip install --no-cache-dir -r requirements.txt

COPY . .

ENV MODEL_CACHE_DIR=/data/models
ENV TOKENIZERS_PARALLELISM=false
ENV MESA_GL_VERSION_OVERRIDE=3.3
ENV PYOPENGL_PLATFORM=egl
ENV PYTHONUNBUFFERED=1

EXPOSE 7860
CMD ["python", "spaces/app.py"]

The key additions are libgles2 libegl1 libgbm1 — MediaPipe requires OpenGL ES even for CPU-only inference. Without these packages it always throws libGLESv2.so.2: cannot open shared object file.

---

Step 5 — Fix requirements.txt (torch CVE block)

Replace the torch lines in requirements.txt:

torch>=2.6.0
torchvision>=0.21.0
torchaudio>=2.6.0

Torch < 2.6 blocks loading .pt files due to CVE-2025-32434. Since legacy .pt checkpoints are removed, this is still a safety measure for other dependencies that may call torch.load.

---

Step 6 — Rebuild and redeploy

# If running locally / Docker:
docker build --no-cache -t genai-deepdetect .
docker run -p 7860:7860 genai-deepdetect

# If on HuggingFace Spaces:
git add src/engines/fingerprint/engine.py
git add src/engines/coherence/engine.py
git add src/engines/sstgnn/engine.py
git add Dockerfile
git add requirements.txt
git commit -m "fix: remove broken models, add libgles2 for mediapipe"
git push

HF Spaces will rebuild the Docker image automatically on push. Watch the build logs — the apt-get install should now include libgles2.

---

What the fixed startup should look like

Fingerprint engine: loading models...
  ✓ detector: Organika/sdxl-detector
  ✓ detector: haywoodsloan/ai-image-detector-deploy
  ✓ detector: dima806/deepfake_vs_real_image_detection
  ✓ CLIP ViT-L/14 loaded for generator attribution
Fingerprint engine ready: 3 detectors, CLIP=ok

Coherence engine: loading models...
  ✓ FaceNet MTCNN + InceptionResnetV1 (VGGFace2) loaded
  ✓ MediaPipe FaceMesh loaded          ← only works after Dockerfile fix
Coherence engine ready: facenet=ok, mediapipe=ok

SSTGNN engine: loading models...
  ✓ SSTGNN detector: dima806/deepfake_vs_real_image_detection
  ✓ SSTGNN detector: prithivMLmods/Deep-Fake-Detector-Model
  ✓ MediaPipe FaceMesh loaded for SSTGNN graph
SSTGNN engine ready: 2 detectors, mediapipe=ok

---

Summary

Error	Cause	Fix
Legacy GenD warnings	custom architecture	removed from engine
Legacy ViT torch CVE error	.pt file + torch < 2.6	removed from engine
trust_remote_code TypeError	duplicate kwarg in _build_image_classifier	removed from all pipeline() calls
wav2vec MISSING/UNEXPECTED keys	custom m_ssl.* namespace, incompatible	removed from engine
libGLESv2.so.2 not found	missing apt packages in Docker	add libgles2 libegl1 libgbm1