Batch_API_caption☀

If you've ever labeled a dataset by hand, you know why this node exists. Batch_API_caption takes a folder of images, sends each one to a multimodal API in parallel, and writes a .txt caption next to it - same basename, one file per image. It's the batch version of API_caption, and it's aimed squarely at the people whose bottleneck is "a thousand images, no captions."

For LoRA training specifically, captions are the highest-leverage part of dataset prep for the current LLM-encoder models: describe what should stay changeable, leave what should stay fixed undescribed. A tool that does this at folder scale is the difference between a weekend project and a real training run.

How it works

Give it an input directory, and it lists every .jpg/.jpeg/.png/.bmp/.webp, then fans them out across a thread pool (default 6 workers, up to 32). Each worker base64-embeds the file's raw bytes (no re-encoding - it ships the original file, so no quality loss from a resize pass), POSTs to the same OpenAI-style endpoint the other API nodes use, and writes the returned text to <output_dir>/<basename>.txt. It only writes a file on success, so failures don't leave empty or half-written captions. It retries once on transient errors, strips <think> blocks, and at the end returns a JSON log with total, succeeded, failed, and the first error if any.

Inputs

• input_dir - required, must exist.
• output_dir - where captions go; if empty, it defaults to input_dir.
• api_type / api_url / API_Key / model_name - same provider map as the other API nodes (Siliconflow default, T8zhenzhen, OpenRouter, Other). Default model Qwen/Qwen3-VL-32B-Instruct.
• prompt - the system instruction (default: detailed Chinese description, 800-word cap).
• output_language - Chinese or English.
• thinking_mode, temperature, max_tokens - shared knobs.
• concurrency - 1 to 32, default 6. Be careful: this multiplies your token spend and the provider's rate limits.

Outputs: log (the JSON summary string) and save_path (where the .txt files went).

Gotchas

• Mind your concurrency and rate limits. 32 parallel calls to a free-tier key is a fast way to get 429s. Start at 6, raise only if the provider is happy.
• It's a blocking node. The workflow sits in "running" while the folder processes. A thousand images at ~10s each with 6 workers is a long Queue run - budget accordingly, and don't kill ComfyUI mid-run without checking the log.
• Your key must be real. It validates that API_Key isn't the <your_key> placeholder and returns a clear error if it is.
• A failure is survivable. Failed files just get logged and skipped; the run continues and you see the count in the JSON.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hasasasa/ComfyUI_DashuaiTools
cd ComfyUI_DashuaiTools
pip install -r requirements.txt

then restart ComfyUI, or install via ComfyUI Manager. Needs requests.

For dataset labeling at any real scale, this is the pack's workhorse - and if you pair it with LoadTextList on the output folder, you've closed the loop between "images" and "prompts with matching filenames."