Qwen3.5-0.8B-Picture-Captioner-v1.2

A lightweight vision-language model (VLM) fine-tuned to generate concise, detailed, and accurate image captions.

[!NOTE] This version is retrained exclusively on a high-resolution subset of the dataset. While output quality is similar to the first version, it features significantly improved details—particularly regarding backgrounds and OCR accuracy.

Model Details

Base Model: Huihui-Qwen3.5-0.8B-abliterated (Uncensored)
Training Dataset: agentlans/monet-caption-highres
Precision: bf16
Attention Mechanism: sdpa (Scaled Dot-Product Attention)
Max Pixels: 768 × 768 (589,824 pixels)

Training Configuration & Results

Training Configuration

Parameter	Value
Epochs	1
LoRA Rank	16 (alpha = 32)
Dropout	0.05
Learning Rate	2e-4
Weight Decay	0.01
Warmup Steps	100

Training Results

Training Loss (Smoothed): 2.7191
Validation Loss: 2.8538

Usage & Input Format

To format your inputs correctly for inference, structure your prompt as shown below. For optimal RAM efficiency, it is recommended to resize your images proportionally to 768 × 768 prior to inference.

content = [
    {"type": "image", "image": image},
    {"type": "text", "text": "Describe this image in detail."}
]

# Recommended generation settings:
# tokenize = False
# add_generation_prompt = False

👉 Click here to view example images and compare captions against the original model version.

Limitations

While efficient and capable, this model shares common failure modes typical of smaller Vision-Language Models (VLMs):

Object and Action Misidentification: May occasionally hallucinate or mislabel specific objects (e.g., confusing a red ball with a tomato) or misinterpret complex interactions between subjects and objects.
Scope & General Use: Not recommended for document scanning, statistical analysis, or medical imaging. It may also underperform on niche genres such as art criticism, provenance analysis, or anime.
Counting & Descriptions: Counts can occasionally be off by 1 and natural language descriptions can sometimes be ambiguous.
Repetitive Captions: Despite variety in the training data, the model can occasionally fall into repetitive phrasing loops.
Inference Speed: As with most VLMs, generation speeds are inherently slower than traditional text-only models or specialized tagging methods.

Licence

This model is licensed under the Apache 2.0 License.

agentlans/Qwen3.5-0.8B-Picture-Captioner-v1.2