Game SD Reference Batch

IP-Adapter's whole trick is image prompting: hand it a reference picture and it encodes that image into the generation, so your chibi actually looks like the chibi you drew before. But what if you have more than one reference? Your best SD/chibi render, plus the original standing design, plus a style reference for how the lineart should feel? Game SD Reference Batch exists for that exact situation - it stacks your references into a single batched IMAGE that goes into one IP-Adapter input.

Why you'd reach for it

In a "same character, every emotion" workflow the reference image is doing a lot of work: it's what keeps the face, hair, and outfit from drifting between moods. IP-Adapter doesn't care whether you give it one image or five - it turns whatever images it receives into embeddings, so feeding it a batch means more reference signal per generation. The pack's philosophy is "same sd_reference image every run," and this node is how you turn that plus a couple of supporting references into one clean wire instead of juggling three separate adapter instances.

How it works

Simple, and readable in the source. It builds a list starting with your sd_reference, then appends original_reference and style_reference - but only if the matching include_original / include_style toggle is on and the image input is actually connected. Any optional reference gets resized (bilinear) to match your sd_reference's height and width, then everything is concatenated along the batch dimension into one tensor. That's the entire mechanism: batch, resize, concat.

The inputs that matter:

• sd_reference - required. Your main character reference; it defines the resolution the whole batch gets resized to.
• original_reference / style_reference - optional extra images. Per the README, original_reference is things like your original standing design, style_reference is the style guide.
• include_original / include_style - both default to false. This is the trap: wire an image in and forget to flip the toggle and it's silently dropped. Nothing errors; your reference just never shows up.

Output is a single reference_images (IMAGE batch) that wires straight into an IP-Adapter node's image input.

Installing it

Identical to the rest of the pack - no extra dependencies, no model downloads. It's pure torch/numpy/PIL.

Via ComfyUI Manager: search "comfyui-game-sd-character-pack", install, restart.

Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/hyyyein/comfyui-game-sd-character-pack
# restart ComfyUI

Nodes appear under "Game SD Character".

Common issues

• Only one reference seems to be applied. Check the include_original / include_style toggles - they default off, and unlike a missing connection this fails silently. Also make sure the images are actually wired in, not just loaded.
• Everything comes out at the sd_reference resolution. That's deliberate - a batch needs matching dimensions to concatenate. If your style reference is a weird aspect ratio it'll get squished to fit, so pick an sd_reference resolution that works for all of them.
• "IP-Adapter is old news, why bother?" Fair. IP-Adapter is an SD 1.5/SDXL-era tool - it loads on nothing newer, and the 2026 community has largely moved character consistency onto instruction-editing models. But this pack is explicitly a chibi/SD-character workflow, and inside that stack IP-Adapter is still the zero-training reference tool. If you're on an Illustrious or SDXL base, this node makes total sense; if you're on Flux-era bases, the whole pack is aimed elsewhere.

One honest note: this is a brand-new pack with zero community footprint yet, so there's no pile of shared war stories to lean on. The node is a thin, correct wrapper over ComfyUI's batching - the only real failure mode is the one everyone hits, and it's the toggle.