feat: add SuperCLIP （Nips 2025）implementation

[SleepingWhz]

Summary

This PR integrates SuperCLIP, our NeurIPS 2025 accepted work, into the OpenCLIP framework.
SuperCLIP is a simple yet highly effective improvement over CLIP: by adding only a lightweight linear layer and introducing classification-based supervision, it enables CLIP to recover fine-grained semantic signals that contrastive learning typically overlooks.

SuperCLIP requires no additional annotated data, increases computation by only 0.077% FLOPs, and also greatly reduces CLIP’s dependence on extremely large batch sizes.

Overall, SuperCLIP delivers consistent and substantial gains across zero-shot classification, image-text retrieval, and purely visual tasks.

---

Why this matters

Despite CLIP’s strong global alignment, it struggles with fine-grained semantics such as object states, spatial relations, and actions.
As shown in Figure 1 of the paper, SuperCLIP significantly improves such distinctions with almost no architectural overhead.

Key advantages

• Large gains without extra labeled data
• +3% to +5% zero-shot classification improvement for ViT-B/L at 512M-scale training (Table 2)
• Up to +5.4% COCO/Flickr retrieval improvement (Table 2)
• Robust under long / rich captions—effectively utilizes detailed captions where CLIP degrades (Table 5)
• Enhances purely visual tasks like PASCAL/ADE20K segmentation and NYUv2 depth (Table 7)
• Alleviates small-batch degradation, maintaining accuracy where CLIP collapses (Figure 4)

Overall, SuperCLIP provides stronger fine-grained visual–text alignment at effectively zero cost.

---

What’s included

• Added superclip_model.py (complete SuperCLIP architecture)
• Added configs:
  • SuperCLIP-ViT-B-16.json
  • SuperCLIP-ViT-L-16.json
• Integrated model registry into factory.py and __init__.py
• Modified transformer.py and loss.py to support classifier-based supervision

All components remain fully optional and do not affect existing CLIP models.

---

Notes

We are happy to make any structural adjustments needed to align with OpenCLIP conventions.
Reference implementation: https://github.com/hustvl/SuperCLIP

[speedinghzl]

Hi @rwightman

Would appreciate if you could take a look at this PR when you have a chance!

The change is minimal - just adding a single linear layer on top of CLIP - but we're seeing substantial performance gains without requiring any additional labeled data. Happy to discuss the approach or provide more details if needed.

Thanks!

[rwightman]

@SleepingWhz @speedinghzl I have taken a look, but it's too much code to take on as is. I haven't taken time to figure out what the 'minimal' required addition would be that has low regression risk.

Also, there don't appear to be any pretrained weights available?