Skip to content
Other

CS-CLIP - An improved CLIP for precise text-image matching

2 min readupdated

The CS-CLIP model is built on OpenAI CLIP ViT-B/32 and fine-tuned on a dataset featuring so-called hard negatives (difficult negative examples based on scene graphs). Standard CLIP models often confuse word order or object relationships, mixing up phrases like "a person on a horse" and "a horse on a person," for instance. This fine-tune solves this problem, providing more reliable compositional reasoning for text-to-image retrieval and zero-shot classification. It will be useful for multimodal system developers and computer vision researchers.

Key features

  • Compositional reasoning: handles complex text descriptions and object relationships better.
  • Architecture compatibility: retains the original OpenAI CLIP ViT-B/32 structure, stored in fp32.
  • COCO training: trained on nearly 2 million caption pairs and hard negatives.
  • Dual format: weights available for both the OpenCLIP library and Transformers.

What you can do

  • Enhanced search: integrate into image search engines for precise matching of queries with complex context.
  • Classification: use for zero-shot image classification without additional training for specific classes.
  • Multimodal pipelines: apply alongside other computer vision models where understanding text and frame structure is critical.

Example

Running the model using the OpenCLIP library:

import torch, open_clip
from PIL import Image

model, _, preprocess = open_clip.create_model_and_transforms("hf-hub:juns/CS-CLIP")
tokenizer = open_clip.get_tokenizer("hf-hub:juns/CS-CLIP")
model.eval()

image = preprocess(Image.open("example.jpg")).unsqueeze(0)
text = tokenizer(["a man riding a brown horse", "a brown man riding a horse"])
with torch.no_grad():
    sims = model.encode_image(image, normalize=True) @ model.encode_text(text, normalize=True).T
Read moreDetails
Benchmarks: SugarCrepe Avg 88.59, VG-Rel 85.96, GQA two-choice 71.27
Architecture: OpenAI CLIP ViT-B/32 (QuickGELU), fp32 weights
Training: AdamW, lr 2e-6, batch size 128, 2 epochs on COCO train2014
Limitations: inherits general biases of pretrained CLIP and was fine-tuned exclusively on COCO data
Discuss Telegram 0 29

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 0.6 GB · ready files1.1 GB
  • BF16weights 0.3 GB0.8 GB
  • FP8 / INT8recommendedweights 0.1 GB0.7 GB
  • 4-bitweights 0.1 GB0.6 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · F32 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # juns/CS-CLIP
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download juns/CS-CLIP --local-dir ~/models/CS-CLIP
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to CS-CLIP
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of clip-vit-base-patch32 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 10All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face
Merges 2All on Hugging Face

About the model

juns/CS-CLIP
Author
juns
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

290

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Other

Le Dog

Le Dog