Extensions/ComfyUI-VideoDescription
ComfyUI Extension

ComfyUI-VideoDescription

Video description nodes using Qwen3-VL and NVIDIA DAM models, plus StoryBoard JSON prompt generation nodes for structured scene and character prompts.

By creativeself-kim·Created 10 months ago·Updated 5 months ago· 1
IXIWORKS-KIMJUNGHO/ComfyUI-VideoDescription
Nodes20
On cloudLocal install
CategoryIXIWORKS/ControlNet, IXIWORKS/Image
Stars1
Updated5 months ago

Nodes (20)

CN Preprocessor

Canny, depth, pose, and more — a ControlNet preprocessor without the bloat

IXIWORKS/ControlNet
CN Step Control

Step-range control with a fade

IXIWORKS/ControlNet
Crop & Resize

One node to get every image to the same aspect and size — boring, and that's the point

IXIWORKS/Image
LoRA Step Loader

A LoRA that fades in and out across the sampling steps — not the strength

IXIWORKS/LoRA
SB Concat Strings

Concatenate twenty string lists element-by-element — the storyboard glue

IXIWORKS/StoryBoard
SB JSON Parser

Your storyboard JSON, exploded into per-cut prompt lists

IXIWORKS/StoryBoard
SB Prompt Filter

Strip the keywords that fight your LoRA style or ControlNet — via Claude

IXIWORKS/StoryBoard
Bypass

The Bypass node that actually skips the upstream branch, not just itself

IXIWORKS/Utils
Concat Strings

Join two strings with any separator — the boring node you'll use constantly

IXIWORKS/Utils
Empty Latent

Pick a ratio, not pixels

IXIWORKS/Utils
Image to List

Gather several images into one list for batch processing

IXIWORKS/Utils
JSON to List

Drag a JSON file in, get a list of strings out — dot-notation key and all

IXIWORKS/Utils
Load Image List

Load a batch of images from dropdowns, straight into a list

IXIWORKS/Utils
Upload to S3

The node that ships your generated frames to a backend — presigned URLs and all

IXIWORKS/Utils
Save Text & Image

Save text and image in one node, auto-numbered the way core ComfyUI does it

IXIWORKS/Utils
String to List

Turn a handful of prompt fields into one list for batch runs

IXIWORKS/Utils
Switch

A boolean switch that only runs the branch you actually pick

IXIWORKS/Utils
Switch Case

Pick one of up to eight inputs by index — and only that one runs

IXIWORKS/Utils
Video Describe

Video Describe runs Qwen3-VL right in ComfyUI

IXIWORKS/Video
Z-Image Prompt Builder

The chat-formatted prompt trick that makes Z-Image actually listen

IXIWORKS/Prompt
Readme

IXIWORKS Tools for ComfyUI

Video description, storyboard, ControlNet, image, LoRA, and utility custom nodes for ComfyUI.

21 nodes across 6 categories.

Node List

IXIWORKS/Video (1)

| Key | Display Name | Description | |-----|-------------|-------------| | VideoDescribe | Video Describe | Full video analysis with Qwen3-VL-8B-Instruct |

IXIWORKS/StoryBoard (6)

| Key | Display Name | Description | |-----|-------------|-------------| | SBJsonParser | SB JSON Parser | Parse storyboard JSON files into scene/character data | | SBPromptBuilder | SB Prompt Builder | Combine scene data into prompt strings | | SBCharacterPrompt | SB Character Prompt | Generate natural language character descriptions | | SBSelectCut | Select Cut | Select a specific scene by index | | SBMergeStrings | SB Merge Strings | Merge two string arrays with a separator | | SBPromptFilter | SB Prompt Filter | Filter prompts for LoRA style or ControlNet type via Claude API |

IXIWORKS/ControlNet (2)

| Key | Display Name | Description | |-----|-------------|-------------| | CNPreprocessor | CN Preprocessor | Run canny/depth/lineart/openpose/mlsd preprocessors | | CNStepControl | CN Step Control | Apply step-range + linear fade to a ControlNet model patch |

IXIWORKS/Image (1)

| Key | Display Name | Description | |-----|-------------|-------------| | ImageCropResize | Crop & Resize | Crop and resize image to a target aspect ratio |

IXIWORKS/LoRA (1)

| Key | Display Name | Description | |-----|-------------|-------------| | LoRAStepLoader | LoRA Step Loader | Load LoRA with step-range scheduling via ComfyUI hooks |

IXIWORKS/Utils (10)

| Key | Display Name | Description | |-----|-------------|-------------| | UtilSwitch | Switch | Boolean switch between two inputs (lazy eval) | | UtilSwitchCase | Switch Case | Select one of N inputs by index (2–8, lazy eval) | | UtilStringToList | String to List | Convert multiple text inputs to a list for batch processing | | UtilConcatStrings | Concat Strings | Concatenate two strings with configurable separator | | UtilSaveText | Save Text | Write a string to a file in the output directory | | UtilLoadImageList | Load Image List | Load images from Input folder via dropdown selectors into a list | | UtilImageToList | Image to List | Collect multiple IMAGE inputs into a list | | UtilBypass | Bypass | Pass-through toggle — returns None when bypass=True | | UtilEmptyLatent | Empty Latent | Generate an empty latent from a named aspect ratio | | UtilJsonToList | JSON to List | Load a JSON file or string and extract items by dot-notation key |

Installation

cd ComfyUI/custom_nodes/comfyui-ixiworks-tools
pip install -r requirements.txt

Dependencies

  • torch >= 2.0.0
  • transformers >= 4.50.0
  • accelerate >= 0.20.0
  • qwen-vl-utils
  • opencv-python
  • pillow, numpy
  • anthropic (required for SBPromptFilter only)

Model Download

VideoDescribe

The Qwen3-VL model downloads automatically on first use (~16GB). To pre-download:

python download_models.py

Models are saved to ComfyUI/models/video_description/Qwen3-VL-8B-Instruct/.

CNPreprocessor

Preprocessor model weights are downloaded automatically from HuggingFace on first use and cached in the ComfyUI models directory.

Node Details

VideoDescribe

Analyzes a video file using Qwen3-VL vision-language model.

Inputs

  • video_path (STRING): Filename or absolute path. Relative paths resolve against ComfyUI/input/.
  • analysis_type (COMBO): detailed / summary / keywords
  • fps (FLOAT, default 1.0): Frames per second to sample
  • custom_prompt (STRING, optional): Overrides analysis_type preset
  • use_4bit (BOOLEAN, default False): Enable 4-bit quantization (~8GB VRAM instead of ~16GB)
  • temperature (FLOAT, default 0.7): Generation temperature (used with custom_prompt)

Outputs: description (STRING), info (STRING)

SBJsonParser

Parses a storyboard JSON file from ComfyUI/input/prompt/.

Input: JSON (COMBO) — filename selector

Outputs: zipped_prompt (ZIPPED_PROMPT list), zipped_character (ZIPPED_PROMPT list), count (INT)

JSON format:

{
  "scene": {
    "1": {
      "mainCharacter": { "koName": "", "enName": "", "description": "" },
      "subCharacter":  { "koName": "", "enName": "", "description": "" },
      "time":        { "ko": "", "en": "" },
      "weather":     { "ko": "", "en": "" },
      "cameraShot":  { "ko": "", "en": "" },
      "cameraAngle": { "ko": "", "en": "" },
      "description": { "ko": "", "en": "" },
      "composition": { "ko": "", "en": "" }
    }
  }
}

SBPromptBuilder

Combines ZIPPED_PROMPT scene data into prompt strings.

Input: zipped_prompt (ZIPPED_PROMPT list, INPUT_IS_LIST)

Output: prompt (STRING list)

SBCharacterPrompt

Generates natural language character descriptions from character data.

Input: zipped_character (ZIPPED_PROMPT list, INPUT_IS_LIST)

Output: character_prompt (STRING list)

SBSelectCut

Selects a single scene by index from a ZIPPED_PROMPT list.

Inputs: zipped_prompt (ZIPPED_PROMPT list), index (INT)

Output: selected_prompt (ZIPPED_PROMPT)

SBMergeStrings

Merges two STRING lists element-wise with a separator.

Inputs: strings_a (STRING), strings_b (STRING), separator (STRING, default " ")

Output: merged_strings (STRING list)

SBPromptFilter

Filters prompts to remove conflicting keywords for a LoRA style or ControlNet type, using the Claude API.

Inputs

  • prompt (STRING list)
  • filter_mode (COMBO): style / controlnet
  • style_type (COMBO): inksketch / ink-wash / pen-ink-illustration / ink-watercolor / watercolor-illustration
  • controlnet_type (COMBO): pose / depth / canny
  • api_key (STRING): Anthropic API key

Output: filtered_prompt (STRING list)

CNPreprocessor

Runs a ControlNet preprocessor on an image batch.

Inputs

  • image (IMAGE)
  • preprocessor (COMBO): canny / depth / lineart / openpose / mlsd
  • resolution (INT, default 512)
  • low_threshold (INT, optional, canny only)
  • high_threshold (INT, optional, canny only)

Output: image (IMAGE)

CNStepControl

Applies step-range gating and linear fade to an existing ControlNet double-block patch. Designed for Flux / Z-Image architectures.

Inputs: model (MODEL), strength_start (FLOAT), strength_end (FLOAT), start_at (FLOAT), end_at (FLOAT)

Output: model (MODEL)

ImageCropResize

Crops an image to a target aspect ratio and resizes to long_side pixels.

Inputs: image (IMAGE), aspect_ratio (COMBO), long_side (INT, default 1024)

Output: image (IMAGE)

LoRAStepLoader

Loads a LoRA with per-step strength scheduling via ComfyUI hooks API.

Inputs

  • model (MODEL), clip (CLIP)
  • lora_name (COMBO)
  • strength_start (FLOAT), strength_end (FLOAT)
  • start_at (FLOAT), end_at (FLOAT)

Outputs: model (MODEL), clip (CLIP), hooks (HOOKS)

UtilSwitch

Boolean switch with lazy evaluation — only evaluates the selected branch.

Inputs: on_false (ANY, lazy), on_true (ANY, lazy), boolean_switch (BOOLEAN)

Output: output (ANY)

UtilSwitchCase

Selects one of N inputs by index with lazy evaluation (2–8 inputs, dynamic via JS).

Inputs: count (INT), select (INT), input_0..7 (ANY, lazy, optional)

Output: output (ANY)

UtilStringToList

Converts multiple text widget inputs into a STRING list for batch processing (up to 8).

Inputs: count (INT), prompt_1 (STRING), prompt_2..8 (STRING, optional)

Output: strings (STRING list)

UtilConcatStrings

Concatenates two strings with a configurable separator.

Inputs: string_a (STRING), string_b (STRING), separator (STRING, optional)

Output: joined_string (STRING)

UtilSaveText

Writes a string to a file in the ComfyUI output directory.

Inputs: text (STRING), filename (STRING, default "output.txt")

No output (OUTPUT_NODE).

UtilLoadImageList

Loads images selected from dropdown menus (from ComfyUI/input/) into an IMAGE list. Up to 20 images, count-controlled with dynamic JS show/hide.

Inputs

  • count (INT, default 1, max 20): number of image slots to show
  • image_1 (COMBO): image file selector — Input folder contents
  • image_2..20 (COMBO, optional): additional image slots

Select [none] to skip a slot. Dropdown list auto-populates from ComfyUI/input/.

Output: images (IMAGE list)

UtilImageToList

Collects up to 8 IMAGE inputs into an IMAGE list (count-controlled, dynamic via JS).

Inputs: count (INT), image_1..8 (IMAGE, optional)

Output: images (IMAGE list)

UtilBypass

Pass-through toggle. When bypass=True, returns None (and JS sets the upstream node to bypass mode).

Inputs: bypass (BOOLEAN), input (ANY, optional)

Output: output (ANY)

UtilEmptyLatent

Generates an empty latent tensor from a named aspect ratio.

Inputs

  • ratio (COMBO): 21:9 / 1.85:1 / 16:9 / 9:16 / 1:1
  • long_side (INT, default 1024)
  • channels (COMBO): 16 (Flux/Lumina/SD3) / 4 (SD/SDXL)
  • batch_size (INT, default 1)

Outputs: latent (LATENT), width (INT), height (INT)

UtilJsonToList

Loads a JSON file or raw JSON string and extracts items as a STRING list.

Inputs

  • file_path (STRING): filename relative to ComfyUI/input/, or absolute path
  • json_string (STRING, optional): raw JSON string — overrides file_path if connected
  • key (STRING, optional): dot-notation path into JSON, e.g. "prompts.solo". Empty = root.

Dict values are always flattened into a single list.

Outputs: strings (STRING list), count (INT)

Hardware Requirements

| Configuration | GPU | VRAM | Notes | |--------------|-----|------|-------| | Minimum | RTX 3060 | 12GB | VideoDescribe with 4-bit quant | | Recommended | RTX 4090 | 24GB | Full FP16 | | Apple Silicon | M1/M2/M3 | Unified | MPS supported |

License

Apache License 2.0

Changelog

v2.1.0

  • UtilLoadImageList: replaced filenames text input with count-based dropdown selectors (Input folder)
  • UtilJsonToList: simplified to file_path + key only — removed item_field, flatten_dict, shuffle, limit

v2.0.0

  • Full node reorganization: unified IXIWORKS/* category prefix
  • Added ControlNet category: CNPreprocessor (canny/depth/lineart/openpose/mlsd), CNStepControl
  • Added Image category: ImageCropResize
  • Added LoRA category: LoRAStepLoader
  • Added Utils: UtilBypass, UtilEmptyLatent, UtilJsonToList, UtilLoadImageList, UtilImageToList
  • Added StoryBoard: SBPromptFilter (Claude API style/controlnet filtering)
  • Removed controlnet_aux dependency — direct preprocessor implementations
  • Removed Layer Separation nodes and Sampler nodes
  • JS extensions updated for all renamed nodes
  • Various bug fixes (INPUT_IS_LIST chains, IS_CHANGED, error handling)

v1.3.2

  • Added Switch Case node
  • Added Join Strings node, String to List node
  • Added frontend JS support (WEB_DIRECTORY)

v1.2.0

  • Added StoryBoard category (JSON Parser, Build Prompt, Build Character Prompt, Select Index, Merge Strings)

v1.1.0

  • Smart video path resolution, performance optimization

v1.0.0

  • Full Qwen3-VL integration with model caching