Load Images(base64) ♾️Mixlab

The stock Load Image node loads exactly one file. The moment you want to feed several reference images into something at once - most commonly IP-Adapter - that one-at-a-time limit gets old fast. LoadImagesToBatch (display name "Load Images(base64)") lets you upload multiple images and hands them back as a single stacked batch, which is exactly the shape IP-Adapter wants when you're conditioning on more than one reference. The pack's README says it plainly: it's for batch input into the IP adapter.

Why that matters: IP-Adapter is image prompting - you give the model a picture and a small adapter injects that image's style or identity through its own attention path, no LoRA training required. Feed it one reference and you get that reference's look. Feed it a batch of a few shots of the same subject or the same aesthetic, and you average out the quirks of any single image for a more stable result - this is the same logic behind IP-Adapter's multi-image "Portrait"-style modes, where several photos beat one. LoadImagesToBatch is the loader that gets you that batch in the first place.

How it works

It ingests images as base64 (the "(base64)" in the name), which is how the browser hands uploaded files to the server, and stacks them into one IMAGE batch tensor. Downstream, that batch behaves like any other multi-image batch - IP-Adapter's batch/encode nodes consume it and treat the set as a group of references.

The inputs and outputs that matter

• images (IMAGEBASE64) in - your uploaded images, via the pack's base64 upload widget. This is the whole input.
• IMAGE out - a single batched IMAGE. Wire it into your IP-Adapter's image/reference input.

That's it - no per-image settings. The tuning happens on the IP-Adapter node (weight, which variant), not here.

How to install it

Via ComfyUI Manager: search comfyui-mixlab-nodes, install, restart. Or clone the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/shadowcz007/comfyui-mixlab-nodes.git

then pip install -r requirements.txt (or install.bat on the Windows portable build) and restart. You'll separately need IP-Adapter itself installed and its models downloaded to actually consume the batch - this node only prepares the images.

Common issues & troubleshooting

Batching fails or the shapes don't line up. Images of different dimensions don't stack cleanly into one tensor. If you get a size-mismatch error, resize your references to a common size first (the pack's own ResizeImageMixlab is right there) before batching, so every image in the set matches.

IP-Adapter ignores most of the batch, or one reference dominates. That's an IP-Adapter behavior, not a loader bug - at high adapter weight one strong reference can steamroll the rest, and the reference set overrides the text prompt entirely above ~0.7–0.8. Drop the weight toward 0.5–0.6 for style, or 0.3–0.5 for faces, so the batch blends instead of one image winning.

Wrong tool for the job. If you actually want to run the same workflow across several images one after another (not blend them together), that's a different pattern - you want an iterate/batch-execution loop, not a single stacked tensor. LoadImagesToBatch is specifically for "treat these N images as one grouped reference," which is the IP-Adapter case.

Uploads don't stick on a remote instance. The base64 upload is browser-driven; on a LAN or cloud box the pack's upload features generally want HTTPS to behave. If uploads silently fail, serve ComfyUI over HTTPS rather than plain http.