ComfyUI Node
💬 Ray's LLM: Folder Captioner
Caption images alphabetically with a connected CLIP/VLM on GPU. Return aligned lists of answers, original images, and absolute paths. Connect a full vision-language CLIP (for example Qwen3-VL from CLIPLoader, type qwen_image), not a text-only encoder. No sidecars are written. Caption one image at a time. Enable return_images only when needed: output images are decoded after captioning and remain in RAM until outputs are released; use start_index and limit to bound that output.
💬 Ray's LLM: Folder Captioner
- clip
- text
- image
- image_path
◄folder►
◄recurse_subfoldersfalse►
◄start_index0►
◄limit-1►
◄user_promptDescribe the visible contents of this image in detail.►
◄max_length1024►
◄temperature0.25►
◄top_k40►
◄top_p0.90►
◄min_p0.05►
◄repetition_penalty1.05►
◄seed-1►
◄thinkingfalse►
◄return_imagesfalse►
◄cleanup_every16►
Category👑 Ray/💬 LLM
Inputs (16)
| Name | Type | Default | Description |
|---|---|---|---|
| clip | CLIP | Full generative vision-language CLIP, such as Qwen3-VL from CLIPLoader (type qwen_image). Text-only encoders cannot caption. | |
| folder | STRING | Enter the folder containing images to caption. | |
| recurse_subfolders | BOOLEAN | false | Include images in subfolders, sorted by relative path. |
| start_index | INT | 0 | Skip this many images in alphabetical relative-path order. |
| limit | INT | -1 | Caption at most this many images; use -1 for all remaining images. |
| user_prompt | STRING | Describe the visible contents of this image in detail. | Enter the instruction to apply independently to every image. |
| max_length | INT | 10241–32768 | Set the maximum number of generated tokens per image. |
| temperature | FLOAT | 0.250–2 | Set sampling randomness; use 0 for greedy decoding. |
| top_k | INT | 400–1000 | Keep this many candidate tokens; use 0 to disable top-k filtering. |
| top_p | FLOAT | 0.900–1 | Set the cumulative probability threshold for nucleus sampling. |
| min_p | FLOAT | 0.050–1 | Filter tokens relative to the most probable token; use 0 to disable. |
| repetition_penalty | FLOAT | 1.050–5 | Set the repetition penalty; use 1 for no penalty. |
| seed | INT | -1-1–2147483647 | -1 for random; any >=0 value is reproducible. |
| thinking | BOOLEAN | false | Enable thinking when supported by the selected model; return its decoded response unchanged. |
| return_imagesopt | BOOLEAN | false | Return original-size images after all captions finish. Off keeps the image output empty and avoids retaining the folder's decoded images in RAM. |
| cleanup_everyopt | INT | 160–1024 | Collect transient objects and release unused ComfyUI VRAM caches every N captions. 0 disables periodic cleanup; low-memory cleanup still applies. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| text | STRING | Decoded VLM answers, one per image, in scan order. |
| image | IMAGE | Original RGB images, decoded after captioning when return_images is enabled; otherwise an empty list. |
| image_path | STRING | Absolute source image paths, paired with the answers and images. |