Zenkai IMP v1

Every inpainting or image-to-video workflow eventually hits the same wall: you have a set of reference images, each with its own description, and quite often its own mask. Zenkai IMP v1 exists to load all three at once - image, prompt, and mask - as a matched set, so you can batch a whole folder of shots through a detailer or an img2vid model without re-wiring anything between items. "IMP" is Image + Mask + Prompt, which is the entire pitch in one acronym.

The node comes from the Zenkai side of DJZ-Nodes, Drift Johnson's prompt-and-dataset tooling. If you're building a dataset or a storyboard-style reference pack, this is the loader you'll reach for; if you just need one image with a prompt, it's overkill and a normal Load Image will do.

How it works

You organize imagemaskprompts/<folder>/ inside the pack directory with a strict naming convention. For each source image, three files share a base name:

custom_nodes/DJZ-Nodes/imagemaskprompts/default/
  scene1.jpg      # the reference image
  scene1_M.jpg    # its mask (suffix "_M")
  scene1.txt      # its prompt

The node lists subfolders as a prompt_folder dropdown, finds every valid image/text/mask triple, and sorts them naturally - so scene1 comes before scene10, which matters more than you'd think. Then mode decides how you walk the list: sequential maps the seed onto the list index (looping), random samples reproducibly. If a mask file is missing it silently substitutes a solid white mask, and if an image has no matching .txt, that pair is skipped entirely. Items without a mask are also fine - you get a full mask instead of an error, which is a forgiving choice for testing.

The inputs that matter

• prompt_folder - dropdown of imagemaskprompts/ subfolders; default is auto-created
• mode - sequential (seed walks the list) or random (seeded sample)
• num_images - batch size, 1–10, loaded into a single IMAGE tensor (images are padded to the largest dimensions if they don't match)
• blacklist - optional comma-separated terms, quoted for multi-word phrases (e.g. cat, "no humans"), matched case-insensitively against the prompt text. Anything that matches is dropped from the pool before selection.

Outputs: image (IMAGE), prompt (STRING), and mask (IMAGE) - all batched the same length, so num_images = 4 gives you four images, four joined prompts (separated by |), and four masks in lockstep. Wire image and mask into an inpaint or detailer pipeline and prompt into a text encoder, and the whole folder runs as one pass.

Install

ComfyUI Manager → search DJZ-Nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/MushroomFleet/DJZ-Nodes
cd DJZ-Nodes
pip install -r requirements.txt

Restart ComfyUI. This node only needs Pillow, NumPy, and torch, which you already have - it's one of the lightest in the pack, with no ffmpeg or Video Helper Suite dependency. The imagemaskprompts/ folder (and a default/ subfolder) is created for you on first run.

Troubleshooting

Two gotchas to know about. First, the mask suffix is _M before the extension - scene1_M.jpg, not scene1_mask.jpg. Get that wrong and you'll silently get white masks, which usually look "fine" right up until your inpaint does nothing. Second, the pack's __init__.py silently skips nodes whose imports fail; if the node is missing entirely, check the console for "Unable to import Zenkai_IMPv1". The dropdown showing only default is normal - it's the auto-created placeholder folder until you add real ones.