Question: Support for Alternative Embeddings in fid_metric (e.g., CLIP for CMMD)

[Snimm]

Question: Support for Alternative Embeddings in fid_metric (e.g., CLIP for CMMD)

Hi, thanks for this great package!

I had a few questions regarding the flexibility of the fid_metric implementation:

1. Encoder:
   By default, I assume fid_metric uses Inception-v3 as the feature extractor. Is that correct?

2. Support for Alternative Encoders:
   Is it currently possible to use a different encoder (e.g., CLIP, DINO, etc.) with fid_metric?
   I’m aware of fid_dino, but wondering if there is broader support for swapping in custom embeddings.

3. CMMD Metric with CLIP:
   Specifically, I'm interested in computing the CMMD metric as described in [this paper](https://arxiv.org/pdf/2401.09603), which uses CLIP embeddings.
   I noticed you support:

   fid_metric(..., distance_type='mmd', kernel_type='rbf')

   • Does this support CLIP embeddings?
   • If not directly, is there a recommended way to plug in CLIP embeddings (or a custom feature extractor) into the FID/CMMD pipeline?

Thanks in advance!

[chaofengc]

Currently, there are no plans to support additional embeddings for FID unless there is broader demand.

That said, we would greatly appreciate contributions from the community. If you're interested in implementing this feature, feel free to modify the FID class below, it's fairly straightforward to read and extend.

IQA-PyTorch/pyiqa/archs/fid_arch.py

Line 360 in 4fd66f3

class FID(nn.Module):

Thanks again for your interest and support!