[BUG] Reproduction gap between Hugging Face checkpoint model and paper demo results

Describe the bug

The official checkpoint (diffusion_harmonizer.pkl from Hugging Face)
does not reproduce the qualitative results shown in the paper/demo.
Specifically, the color/tone harmonization between inserted foreground
objects and the background scene is nearly absent, and shadow synthesis
is significantly weaker than demonstrated.

Using the first demo example from the paper(the scene with the Lightning McQueen-like
cartoon car inserted into a real street intersection):

• Expected (paper/demo): The foreground cartoon car blends
  naturally into the scene — color tone matches the environment lighting,
  and a realistic shadow is synthesized beneath/around the car.
• Actual (official checkpoint): The foreground McQueen car shows
  almost no color/tone harmonization with the background. The shadow
  effect around the foreground car is unsatisfactory. The two background
  cars look relatively reasonable, but the prominent foreground car does
  not blend into the scene as showcased.

Steps/Code to reproduce bug

1. Download the official checkpoint diffusion_harmonizer.pkl from
   Hugging Face.
2. Run inference using the provided inference script with the demo
   input (Lightning McQueen scene from 1st page on paper):

python inference_pix2pix_turbo_harmonizer.py \
--input_image examples/ \
--model_path diffusion_harmonizer.pkl \
--timestep 250 \
--nontemporal

3. Compare the output against the paper/demo qualitative results.
4. I also tried different inference timesteps (e.g., 100, 200, 250, 500)
   — the overall result still does not match the demo quality.

Expected behavior

The output should match the qualitative results shown in the paper and
demo materials:

• Foreground objects should have their color/tone harmonized with the
  background scene.
• Realistic shadows should be synthesized for inserted objects.
• The overall result should look natural, as demonstrated in the paper
  figures.

Environment overview (please complete the following information)

• Environment location: Docker
• Method of Deployment: Docker
• Built harmonizer environment on top of the base image per the
  provided Dockerfile.

Environment details

• Hardware: NVIDIA RTX 4090
• Operating System: Ubuntu 22.04 (Linux 5.15.0)
• Number of physical nodes: 1 (inference only)
• Network connectivity between the nodes: N/A (single-node inference)

Additional context

I would like to ask:

1. Was the paper/demo result generated using the same checkpoint
   (diffusion_harmonizer.pkl) that is currently released on Hugging Face?
2. Were there any additional training datasets, fine-tuning stages, or
   internal checkpoints used for generating the demo examples that differ
   from the released checkpoint?
3. The official GitHub repository does not appear to implement the
   multi-scale VGG perceptual loss described in the paper (Section 7.2). Is
   the released checkpoint trained with the loss recipe described in
   the paper?

Thank you for this excellent work — any clarification would be greatly
appreciated.