SenseNova Reference Images (1-10)

The regular SenseNova Reference Image node is great until your edit needs three images - subject, clothing, and a background or layout reference - at which point it's one slot short and you're doing hacky batch tricks. SenseNova Reference Images (1-10) is the fix: the same reference-conditioning machinery, but with ten Image-N input slots instead of two.

Same shape as its little sibling: feed it positive and negative conditioning, plug in images, and get positive and image_condition out. Image-1 is the only required slot; the other nine are optional, so this node works as a drop-in replacement even if you only ever use two. Old workflows that pushed more than two references through the standard node auto-migrate here on import, which the author added specifically because that's where people were getting stuck.

What changes when you go multi-image

The mechanism is the same token-injection trick - each image becomes a block of image tokens with an Image-1, Image-2, ... label in the prompt, and the model attends to each as a unit. But two things scale with image count, and both are worth knowing:

• Every image gets a label, and your prompt must use them. "Image-1 supplies the person, Image-2 the garment, Image-3 the background" is exactly how the model parses it. If your prompt only says "the person and the clothes," the extra references are under-directed and the model will improvise.
• The pixel budget is shared. Each reference is resized to fit within a total budget that shrinks as you add more images. Keep individual references reasonable - a 4096² monster plus nine friends will get downscaled aggressively.

Multi-reference is also where the pack's KV cache earns its keep: the text and reference prefix is computed once per guidance branch and shared across every generated variant, so batch generation with ten references doesn't multiply the reference encoding ten times. That's the difference between "this will fit" and an OOM.

The gotcha that bites everyone

A single image per slot, always - the node errors on IMAGE batches. And remember the ordering contract: Image-1 is the main/source, and for something like clothing transfer the person goes in Image-1 and the garment in Image-2, matching how the prompt refers to them. Feed images in any order and your "wear Image-2's clothes" instruction will confidently dress the person in the background.

Beyond that, everything from the standard node applies: wire positive → KSampler positive and image_condition → KSampler negative for img_cfg=1, or route image_condition into SenseNova Edit Guider for custom sampling. It's the same pack, same install (Manager search "SenseNova U1.5 (T8)" or git clone), no extra dependencies.