zer0int/CLIP-Registers-Gated_MLP-ViT-L-14

CLIP Needs Registers. And Gated MLPs. And +20M params.

Fixing CLIP's modality gap via happy little accidents.

• ❤️ this CLIP? Donate if you can / want. TY! (and enjoy, either way!) 🤗

ℹ️ Update 02/June/2025:

• You can now load the model with HF 'transformers'. ✅
• Unfortunately, AutoModel produced nonsense / I couldn't get "trust_remote_code=True" to work properly - (using that was suggested in response to my pull request on GitHub).

💡 Alas, you will need to:

• Download the 'hfmodel' folder
• Use it for manually importing the correct (my custom) CLIPModel code required as per the config.json
• Minimal example code:

import torch
from PIL import Image, ImageDraw
import transformers
from hfmodel.modeling_clip import CLIPModel
from transformers import CLIPProcessor
from torchvision.transforms import ToTensor
import torch.nn.functional as F

model = CLIPModel.from_pretrained("zer0int/CLIP-Registers-Gated_MLP-ViT-L-14")
processor = CLIPProcessor.from_pretrained("zer0int/CLIP-Registers-Gated_MLP-ViT-L-14")

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

size = 224
im = Image.new("RGB", (size, size), (255, 255, 255))
draw = ImageDraw.Draw(im)

# --------- GPT-4.1's idea of a pineapple. Need an input image... ---------
body_bbox = [size*0.28, size*0.38, size*0.72, size*0.90]
draw.ellipse(body_bbox, fill=(254, 221, 72), outline=(180, 120, 0), width=5)
eye_color = (198, 134, 66)
for row in range(4):
    for col in range(3):
        ex = size*0.36 + col*size*0.09 + (row%2)*size*0.045
        ey = size*0.50 + row*size*0.085
        ew, eh = size*0.035, size*0.025
        draw.ellipse([ex-ew, ey-eh, ex+ew, ey+eh], fill=eye_color, outline=None)
leaf_color = (61, 179, 70)
leaf_base_x = size/2
leaf_base_y = size*0.38
for i, (angle, length) in enumerate([(-28, 65), (-12, 70), (0, 80), (12, 70), (28, 65)]):
    from math import radians, cos, sin
    a = radians(angle)
    tip_x = leaf_base_x + length*sin(a)
    tip_y = leaf_base_y - length*cos(a)
    left = (leaf_base_x + 13*cos(a+1.5), leaf_base_y + 13*sin(a+1.5))
    right = (leaf_base_x + 13*cos(a-1.5), leaf_base_y + 13*sin(a-1.5))
    draw.polygon([left, (tip_x, tip_y), right], fill=leaf_color)
im.save("pineapple.png")
# ---------

image = Image.open("pineapple.png").convert("RGB")
texts = ["pine", "apple", "pineapple", "orange", "pear", "person", "cat", "dog"]

inputs = processor(text=texts, images=image, return_tensors="pt", padding=True).to(device)

with torch.no_grad():
    outputs = model(**inputs)
    image_embeds = outputs.image_embeds
    text_embeds = outputs.text_embeds

image_embeds = F.normalize(image_embeds, dim=-1)
text_embeds = F.normalize(text_embeds, dim=-1)
cos_sim = image_embeds @ text_embeds.T
cos_sim = cos_sim.squeeze(0)

for text, sim in zip(texts, cos_sim):
    print(f"Similarity with '{text}': {sim.item():.4f}")

---

I just want a new Text Encoder...

• ...for my Text-to-Image (Text-to-Video) AI! \o/
• I recommend this one, the 'sweet spot' ckpt12: 👉 direct download 👈
• Even lower modality gap (text 'more alike' to image, but less accurate): direct download
• Enjoy! (You don't need to do anything else, they're just normal CLIP Text Encoders!)

⚠️ Full model (currently) not HuggingFace Transformers compatible. ⚠️

• The ViT (Vision Encoder) is basically a big mutant. Alas:
• The full model .safetensors have the 'import clip' (OpenAI) structure inside.
• It's just so you don't need to load any 'danger pickles'. :)
• Alas, currently it runs with 'import clip' code (I'm working on a HF implementation, though!).
• However, for now, I made an entire playground for the CLIP models (+ safetensors loading)! 🎉:
• 🌟 https://github.com/zer0int/CLIP-fine-tune-registers-gated ✨
• All code for fine-tuning it yourself is also included on my Git! 🤗

Wait, but what is this?!

• The Vision Transformer has +4 tokens (Register Tokens).
• ...And gated ReLU MLPs inside each layer + final Fusion MLP.
• +20M parameters (~430M -> now: ~450M)
• It's now a CLIP with an extremely low modality gap.
• See the table below for details.
• And if you want to know more about modality gaps & all details please check out the GitHub!

An image is worth 16x16 words, alas:

Attention Heatmap, pre-trained OpenAI CLIP ViT-L/14:

This model, CLIP REG-XGATED:

Text-To-Image examples, Flux.1-dev, pure CLIP (no T5) guidance:

Model Performance Overview

Task / Dataset	Metric	ViT-L/14 OpenAI (Pre-trained)	X-GATED (ckpt20 xtreme)	X-GATED (ckpt12 balanced)	X-GATED (ckpt12 balanced, ablated)
VoC-2007 (Multilabel)	mAP	0.7615	0.8140	0.8471	0.8247
MSCOCO Retrieval	Image Recall@5	0.2194	0.3565	0.3532	0.3349
	Text Recall@5	0.3034	0.5425	0.5278	0.5086
Linear Probe CIFAR-10	Acc@1	0.9535	0.9813	0.9813	0.9811
	Acc@5	0.9966	0.9997	0.9997	0.9997
	Mean Class Recall	0.9535	0.9813	0.9813	0.9811
MVT ImageNet/ObjectNet (Zero-Shot)	Accuracy	0.8453	0.8686	0.8830	0.8815
Linear Probe ILSVRC2012	Top-1	69.86%	66.43%	67.10%	68.99%
	Top-5	92.70%	91.52%	91.83%	92.64%
Modality Gap Metrics	Euclidean Gap ↓	0.8276	0.4740	0.5395	0.7486
	JSD ↓	0.5200	0.1601	0.1303	0.3310
	Wasserstein Distance ↓	0.4084	0.1742	0.2102	0.3262
	Img-Text Cos Sim (mean) ↑	0.2723	0.4926	0.4794	0.3634
	Img-Text Cos Sim (std)	0.0362	0.0814	0.0758	0.0537
	Text-Text Cos Sim (mean)	0.6807	0.6657	0.6896	0.6896
	Text-Text Cos Sim (std)	0.1344	0.1671	0.1535	0.1535

Bolded values represent the best performance for each metric.