Josia多图加载

JosiaMultiImageLoader (Josia多图加载) is the "caption a folder of images without clicking Queue forty times" node. Drop a pile of images in - upload, drag-and-drop, or paste - and it either hands them to downstream nodes as a proper batch or, in its smartest mode, makes ComfyUI run the entire rest of the workflow once per image automatically. For image-to-text pipelines (captioning, tagging for training data), that single feature is the whole reason to install it.

List mode vs batch mode

The output_mode switch chooses the personality:

• List mode (default off → on): images_out returns a Python list and the node marks that output with ComfyUI's OUTPUT_IS_LIST flag. Downstream nodes that support list expansion get executed N times - once per image - in a single queue. Feed it 40 photos and a llama.cpp captioning node runs 40 times, no manual batching, no seed-increment hacks. Each image keeps its own correct aspect and size, so nothing gets letterboxed.
• Batch mode: all images get merged into one torch batch. This is the mode for pixel-pipeline stuff like img2img over a folder. Caveat, from the README, is an honest math limitation: if the images have mixed aspect ratios, a tensor batch has to be one canvas size, so ComfyUI letterboxes the odd ones out (black bars). If you need each image at its true size, use list mode or the per-image ports.

The per-image ports and the scaling panel

There are 50 individual image_1 … image_50 outputs plus a total_count (INT), so you can also wire specific slots into specific downstream nodes - deterministic positions instead of a fold. And the resize panel handles the "VAE wants ~1MP" problem: enable_resize master switch, megapixels (0 = no scaling), edge_value for longest/shortest-edge scaling, resolution_steps for progressive downscaling (the README recommends 2–16 steps for big shrink jobs to cut aliasing), interpolation (lanczos default), and multiple_of alignment (default 16, matching VAE tiles). The images optional input lets an upstream image slot in before the loaded ones.

Install

Zero extra Python dependencies:

cd ComfyUI/custom_nodes
git clone https://github.com/Josia-doit/ComfyUI_JosiaNodes

or via ComfyUI Manager (search ComfyUI_JosiaNodes), restart. The whole gallery UI (thumbnails, drag-drop, sort) is frontend JS - the README is explicit that this node depends on web/js, so don't prune that folder.

Gotchas

First: the mixed-ratio letterboxing isn't a bug, it's PyTorch tensors being rectangular. When someone's caption file has a 4:3 photo labeled with a 9:16 letterbox, that's batch mode doing its job - use list mode or per-image ports. Second: list mode explicitly doesn't support upstream list chaining; if you need to feed a list from another node in, set the upstream to batch mode. And finally, list mode + heavy VLM models multiplies your VRAM budget by N - queueing 50 image caption runs on a 6GB card is how you learn patience.

One thing to be clear about: if your goal is a manual gallery where you pick images by hand from a dropdown, that's the stock LoadImage's job. This node's superpower is automation - it wants to feed the whole folder through, not stand there while you choose one.