m0a3cebee1aaac92424b0

Public inference weights for a scene-text visual-style encoder. The model takes a tightly cropped text-region image and returns normalized representations for overall style, typography/font, and appearance, plus eight local style tokens.

The repository name is intentionally opaque, but this is a public repository and the name is not an access-control mechanism.

Files

• model.safetensors: complete fine-tuned backbone and inference heads;
• config.json: architecture, input dimensions, output dimensions, and weight SHA-256;
• LICENSE: Apache License 2.0 text retained for the pretrained backbone lineage;
• NOTICE: third-party attribution and release notes.

No optimizer, scheduler, training classifier, adversarial head, manifest, local path, or pickle checkpoint is included.

Use

Install the inference code from:

https://github.com/zqpresent/aaazhouquandecangku

Then download this repository or pass its public Hugging Face ID directly:

from opaque_encoder import StyleEncoder

encoder = StyleEncoder.from_pretrained(
    "zqpresent/m0a3cebee1aaac92424b0",
    device="cpu",
)
features = encoder.encode(["text_crop.png"])
print(features["z_global"].shape)      # [1, 256]
print(features["z_font"].shape)        # [1, 128]
print(features["z_appearance"].shape)  # [1, 128]
print(features["style_tokens"].shape)  # [1, 8, 384]

The code repository documents installation, CLI usage, preprocessing, metrics, and limitations.

Output semantics

Output	Shape	Normalized	Intended use
z_global	[N, 256]	yes	overall text-style retrieval
z_font	[N, 128]	yes	typography/font-oriented retrieval
z_appearance	[N, 128]	yes	fill/stroke/shadow-oriented retrieval
style_tokens	[N, 8, 384]	no	downstream conditioning research

Scope and limitations

The encoder was trained and evaluated on counterfactual synthetic scene-text data. Held-out font-family and background-source tests support style sensitivity and low nuisance leakage in that domain, but do not establish product-level performance on real photographs. It is English-first, letter-spacing remains a weak factor, low-contrast crops and similar fonts are difficult, and the style tokens have not yet been validated with a generator.

Integrity

model.safetensors SHA-256:

edb52cbd56667a5f317271b31f329858500bb2ee20e791c631f43321b42fd544

Backbone attribution

This is a fine-tuned model using the DINOv2 ViT-S/14 architecture and pretrained backbone distributed through timm as vit_small_patch14_dinov2.lvd142m. DINOv2 was developed by Meta AI Research; the timm implementation is by Ross Wightman and contributors. See NOTICE and LICENSE. The bundled Apache text governs the applicable upstream material; public read access and that bundled text do not by themselves declare an additional license for newly produced fine-tuning contributions.