File size: 8,576 Bytes
c72183f
f29517d
 
c72183f
f29517d
 
 
 
 
 
 
 
 
 
 
 
 
41b2e06
 
 
c72183f
 
f29517d
 
 
 
 
da01989
 
 
f29517d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
da01989
 
 
 
 
 
 
f29517d
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
---
library_name: cinemaclip
pipeline_tag: zero-shot-image-classification
tags:
  - clip
  - mobile-clip
  - cinema
  - film
  - movies
  - multi-task
  - hybrid
  - cinematography
  - domain-specific
  - image-classification
  - zero-shot
base_model: apple/MobileCLIP-S1-OpenCLIP
base_model_relation: finetune
license: apple-amlr
license_name: aplle-ascl
license_link: https://github.com/apple/ml-mobileclip/blob/main/LICENSE_MODELS
---

# CinemaCLIP-1.0.0

**CinemaCLIP** is a [MobileCLIP-S1](https://huggingface.co/apple/MobileCLIP-S1-OpenCLIP) fine-tune specialized for understanding the visual language of cinema at a frame level. It is a hybrid CLIP model with 23 classifier heads that represent a comprehensive taxonomy built with domain experts. For more info, see our [launch blog post](https://www.ozu.ai/cinemaclip).

This repository ships three serialized forms of the same model:
- **Torch** (`model.safetensors`): load via the `cinemaclip` Python package.
- **CoreML** (`ImageEncoder.mlmodel`, `ImageEncoder.mlpackage` and `TextEncoder.mlpackage`): on-device Apple Neural Engine inference.
- **ONNX** (`ImageEncoder.onnx`, `TextEncoder.onnx`, plus `_fp16` variants): cross-platform inference.

## Install

```bash
pip install cinemaclip            # core
pip install "cinemaclip[coreml]"  # CoreML export/inference
pip install "cinemaclip[onnx]"    # ONNX export/inference
```

## Usage (PyTorch)

```python
from PIL import Image
from cinemaclip import CinemaCLIP

model = CinemaCLIP.from_pretrained("OZU-Technology/CinemaCLIP").eval()

# End-to-end classification on a PIL image
image = Image.open("still.jpg").convert("RGB")
predictions = model.predict_image(image)
predictions["classifier_preds"]  # Classifier predictions
predictions["clip_image_embedding"]

# Just the image embedding
x = model.preprocess(image).unsqueeze(0)
image_embedding = model.encode_image(x, normalize=True)   # [1, 512]

# Just the text embedding
tokens = model.tokenizer(["a medium closeup of "])
text_embedding = model.encode_text(tokens, normalize=True)  # [1, 512]
```

The `CinemaCLIP.predict_image` method is demonstrative for how to get post-processed classifier outputs from the model. It is not super efficient or production ready, and must be treated as a reference above all else.

## Usage (CoreML)

```python
import coremltools as ct
from PIL import Image

img_encoder = ct.models.MLModel("ImageEncoder.mlpackage")
# Input must be 256x256 RGB, resized with BICUBIC for parity with the released torch outputs.
img = Image.open("still.jpg").convert("RGB").resize((256, 256), Image.Resampling.BICUBIC)
out = img_encoder.predict({"Image": img})
embedding = out["clip_image_embedding"]    # [512]
probabilities = out["probabilities"]       # [101] — concat of 23 per-category outputs

# TODO
text_encoder = ct.models.MLModel("TextEncoder.mlpackage")
```

## Usage (ONNX)

```python
from PIL import Image
from onnxruntime import InferenceSession
from torchvision import transforms as T

img = Image.open("still.jpg").convert("RGB")
preprocess = T.Compose([
    T.Resize((256, 256), interpolation=T.InterpolationMode.BICUBIC),
    T.ToTensor(),   # yields float tensor in [0, 1] — no mean/std normalization
])
x = preprocess(img).unsqueeze(0).numpy()

session = InferenceSession("ImageEncoder.onnx", providers=["CPUExecutionProvider"])
emb, probs = session.run(None, {"Image": x})
```

## Output structure

`probabilities` is a flat `[101]` vector — the concatenation of all 23 classifier heads' post-activation outputs. Label names and positions are in the shipped `CinemaNetSchema.json`:

```python
import json
schema = json.load(open("CinemaNetSchema.json"))
label_names = schema["probabilities_labels"]  # len == 101
```

The classifier heads are a mix of 3 types of classifiers:
- Single label (softmax activation)
- Multi label (sigmoid activation)
- Binary (sigmoid activation)


## Evaluation

`CinemaCLIP` outperforms not only the largest existing CLIP models (up to 28x larger), but also leading VLMs in cinematic understanding tasks (we benchmarked against the leading `4B` VLMs).

Two inference modes are reported for CinemaCLIP:
- **Classifier** — the shipped supervised heads on the CinemaCLIP image embedding.
- **0-shot** — zero-shot text/image similarity using CinemaCLIP's own text encoder.

| Category | CinemaCLIP 0-shot | CinemaCLIP Classifier | Qwen3.5-4B | Gemma4-4B | InternVL3.5-4B | Molmo2-4B | DFN ViT-H-14 | MetaCLIP PE-bigG | OpenAI ViT-L-14 | MobileCLIP-S1 | DFN ViT-L-14 | SigLIP2 SO400M | SigLIP2 ViT-gopt |
|---|--:|--:|--:|--:|--:|--:|--:|--:|--:|--:|--:|--:|--:|
| **Mean**              | **82.9** | **87.6** | **57.6** | **56.7** | **55.3** | **55.3** | **45.9** | **45.2** | **44.8** | **44.2** | **39.0** | **38.7** | **36.5** |
| Color Contrast      | 89.6 | 86.8 | 33.7 | 35.3 | 33.7 | 35.3 | 34.0 | 33.1 | 49.4 | 38.7 | 37.1 | 57.7 | 25.2 |
| Color Key           | 84.9 | 92.9 | 78.1 | 78.1 | 80.3 | 64.3 | 58.2 | 50.2 | 53.2 | 59.4 | 48.3 | 22.8 | 52.6 |
| Color Saturation    | 82.6 | 82.6 | 66.5 | 65.4 | 72.1 | 45.9 | 55.1 | 61.8 | 58.1 | 35.8 | 46.8 | 33.3 | 31.8 |
| Color Theory        | 71.3 | 72.7 | 54.0 | 51.7 | 50.7 | 48.7 | 54.7 | 51.7 | 50.7 | 47.3 | 47.7 | 31.3 | 31.7 |
| Color Tones         | 86.0 | 86.5 | 50.2 | 62.6 | 70.6 | 62.1 | 58.5 | 50.2 | 52.0 | 55.7 | 47.2 | 24.0 | 17.7 |
| Lighting Cast       | 85.9 | 90.4 | 38.3 | 53.3 | 39.8 | 35.7 | 25.4 | 29.3 | 28.8 | 35.7 | 22.8 | 37.8 | 18.2 |
| Lighting Contrast   | 93.9 | 95.3 | 29.8 | 39.1 | 38.7 | 46.1 | 35.3 | 35.5 | 32.6 | 39.0 | 39.4 | 48.4 | 37.6 |
| Lighting Edge       | 87.6 | 90.4 | 22.8 | 38.8 | 31.2 | 40.4 | 22.4 | 31.6 | 41.6 | 34.0 | 21.2 | 26.0 | 25.6 |
| Lighting Silhouette | 88.4 | 93.1 | 80.9 | 63.0 | 48.9 | 48.8 | 66.6 | 67.1 | 67.4 | 58.4 | 43.5 | 46.2 | 78.9 |
| Shot Angle          | 73.4 | 82.3 | 41.9 | 49.2 | 33.2 | 49.9 | 28.0 | 13.7 | 19.0 | 19.6 | 25.9 | 21.3 | 17.2 |
| Shot Composition    | 95.5 | 96.0 | 46.0 | 54.5 | 55.7 | 60.5 | 27.8 | 24.3 | 21.3 | 22.0 | 25.2 | 31.4 | 11.4 |
| Shot Dutch Angle    | 61.9 | 78.5 | 62.2 | 65.1 | 46.7 | 49.3 | 27.3 | 44.5 | 38.4 | 56.6 | 25.9 | 47.6 | 68.7 |
| Shot Focus          | 71.3 | 71.2 | 19.9 | 26.6 | 26.3 | 25.1 | 32.9 | 31.2 | 24.4 | 31.3 | 37.3 | 48.2 | 12.6 |
| Shot Framing        | 79.2 | 83.8 | 38.0 | 29.6 | 40.1 | 34.6 | 33.6 | 24.9 | 23.5 | 23.9 | 33.0 |  7.3 |  9.8 |
| Shot Height         | 90.5 | 91.8 | 38.1 | 37.4 | 41.2 | 53.0 | 37.6 | 33.7 | 28.9 | 24.0 | 33.6 | 29.6 | 23.9 |
| Shot Lens Size      | 67.9 | 70.6 | 49.6 | 28.0 | 43.6 | 46.6 | 32.1 | 28.0 | 34.5 | 30.1 | 25.7 | 30.1 | 17.6 |
| Shot Location       | 90.9 | 93.9 | 81.0 | 82.2 | 81.5 | 79.2 | 73.0 | 68.4 | 68.0 | 75.6 | 66.1 | 65.0 | 46.7 |
| Shot Symmetry       | 88.3 | 92.9 | 90.2 | 86.7 | 76.0 | 80.2 | 76.6 | 78.0 | 54.0 | 39.3 | 24.9 | 46.0 | 82.4 |
| Shot Time of Day    | 69.2 | 89.0 | 75.1 | 66.1 | 70.7 | 70.7 | 68.1 | 69.6 | 60.3 | 73.7 | 71.2 | 48.5 | 42.7 |
| Shot Type           | 81.8 | 90.5 | 81.3 | 61.2 | 57.0 | 57.4 | 52.8 | 40.4 | 36.5 | 35.7 | 56.7 | 46.5 | 29.7 |
| Shot Type - Crowd   | 91.5 | 99.6 | 97.2 | 88.2 | 94.3 | 94.8 | 55.9 | 69.1 | 68.6 | 77.2 | 37.3 | 52.4 | 69.3 |
| Shot Type - OTS     | 92.0 | 95.5 | 92.5 | 85.0 | 83.9 | 87.6 | 53.2 | 57.0 | 73.9 | 60.3 | 42.1 | 50.5 | 51.2 |

The `shot.lighting.direction` head ships in the classifier heads but has been excluded from the table above being a multi-label classifier.

## Files in this repo

| File | Purpose |
|---|---|
| `model.safetensors` | Blended (α=0.75) torch weights — `CinemaCLIP.from_pretrained` target |
| `config.json` | Autogenerated `__init__` kwargs for `CinemaCLIP` |
| `CinemaNetSchema.json` | Schema detailing classifier head metadata, confidence thresholds, preprocessing info |
| `ImageEncoder.mlmodel` | CoreML `"neuralnetwork"` ImageEncoder (unified embedding + probabilities) |
| `ImageEncoder.mlpackage` | CoreML ImageEncoder (unified embedding + probabilities) |
| `TextEncoder.mlpackage` | CoreML TextEncoder |
| `ImageEncoder.onnx` / `_fp16.onnx` | ONNX ImageEncoder |
| `TextEncoder.onnx` / `_fp16.onnx` | ONNX TextEncoder |


## Citation

```bibtex
@misc{cinemaclip2026,
  title        = {CinemaCLIP: A hybrid CLIP model and taxonomy for the visual language of cinema},
  author       = {Somani, Rahul and Marini, Anton and Stewart, Damian},
  year         = {2026},
  publisher    = {HuggingFace},
  doi          = {10.57967/hf/8539},
  howpublished = {\url{https://huggingface.co/OZU-Technology/CinemaCLIP}},
  note         = {Model weights and taxonomy}
}
```