Cannot run HRN on modern GPU (Blackwell sm_120 / CUDA 12.8) — outdated dependencies + Colab broken

[Moooooonk]

Hi,

I'm trying to run HRN on a server with the following setup:

• GPU: NVIDIA RTX PRO 6000 Blackwell (compute capability 12.0 / sm_120)
• Driver: 590.48.01
• CUDA: 12.8
• OS: Ubuntu 22.04.5

Problems encountered

1. Official requirements are incompatible with modern GPUs
   • torch==1.6.0 and tensorflow-gpu==2.3.0 do not support sm_120 (Blackwell) at all.
   • These versions cannot even detect the GPU.
2. numpy compatibility issue
   • np.VisibleDeprecationWarning was removed in recent numpy versions.
   • util/preprocess.py line 9 crashes on import.
3. tensorflow is required but hard-coded
   • models/hrn.py directly calls tf.version and tf.compat.v1, making it impossible to skip tensorflow.
   • tensorflow-gpu==2.3.0 does not support CUDA 12.x.
4. Official Colab notebook is broken
   • The official Colab demo no longer runs due to dependency conflicts with the current Colab runtime.
5. ModelScope demo is region-restricted
   • modelscope.cn times out from non-China networks.

What I've tried

• Creating a fresh conda env with Python 3.10
• Installing torch 2.11+cu128, tensorflow-cpu, nvdiffrast 0.4.0
• Patching util/preprocess.py (commenting out np.VisibleDeprecationWarning)
• Patching models/hrn.py (try/except for tensorflow import)
• Still working through remaining compatibility issues

Request

1. Is there a working way to run HRN on Google Colab as of 2026?
   • The official Colab notebook fails, and I haven't been able to find an updated version that works with the current Colab runtime.
   • If there is a known working Colab setup or workaround, could you please share it?
2. Could the requirements be updated to support modern CUDA (12.x) and recent PyTorch (2.x+)?
3. Or alternatively, could a Docker image or updated installation guide be provided?

Thank you for the great work on HRN. I hope this helps other users facing similar issues with newer hardware and environments.