>>18476
1. Manually divide images into folders based on the large categories like beheading, choking, etc.

2. Manually further divide images into subcategories. In the beheading category for example, throat-slit, separating, separated, no-head, and head-only, are the subcategories.

3. For each subcategory folders, manually place "_captioning.ini" which describes what kind of scene it is.

4. The first run of an automatic captioning. I have used this VLM on llama-server for the entire process. It is an AI censorship removed.
https://huggingface.co/mradermacher/Huihui-Qwen3.5-27B-abliterated-GGUF
In the first run, I asked VLM to fill a formatted YAML file based on its observation. The "_captioning.ini" was also included in the instruction.

5. Further runs of auto-captioning asking specific things to VLM. For example, the number of people in the image, their clothings, actions, and censorship. Each run updates the YAML file. I also run a couple of validation scripts.

6. Manually inspect and correct the YAMLs. For example, the VLM tend to interpret throat slit scene as choking. The another important correction needed was to describe a uniform with a same wording.

7. YAML to TXT. Using the same Huihui-Qwen3.5, I asked it to generate a fluent caption based on the YAML file. Some important attributes (artist for example) were preserved as tags.

That is how I spent $2000 on GPU. The Python scripts are kindly made by Claude Opus.