TS Multi Reference

Multi-reference pipelines - Qwen-Image-Edit, Flux-with-references - want you to attach one or more reference images to the conditioning stream as reference_latents, so the model has a picture (or three) to work from alongside the text. Do it by hand and it's a small wall of nodes: three VAE Encode, three Reference Latent, all wired into the right sockets. TS MultiReference collapses that into one node: up to three images in, one conditioning out, with the sizing and encoding handled for you.

How it works

For each connected image_N, the node resizes the image down to a megapixel budget (max_megapixels, 1.0 default), snaps the dimensions to a grid (divide_by, 32 default - a safe choice for Flux 2 and Qwen-Image-Edit; most VAEs only need 8 or 16), VAE-encodes it, and appends it as a reference_latent on your conditioning. If a source's aspect doesn't snap cleanly to the grid, a thin band is center-cropped off the longer side so the output aspect always matches and the picture is never stretched.

The conditioning and vae inputs are both optional - the node only runs the VAE encode when both images and conditioning are connected. Connect images alone and it just resizes them for you, which is handy on its own.

There's a genuinely thoughtful piece of plumbing with the masks: mask_1..3 are used only as a bounding-box hint (with 16 px padding) - the pixels inside the box are preserved as-is, the mask shape is not cut out. And mask_orientation auto-detects from the four corners whether your mask means "1.0 = transparent" (ComfyUI Load Image alpha) or "1.0 = subject" (SAM/BiRefNet outputs), so both conventions just work. If a full-bleed subject covers the corners and the auto-detection guesses wrong, set it explicitly.

The input you'll actually set

• max_megapixels - how big each reference gets before encoding. 1.0 is right for most pipelines; lower saves VRAM, higher keeps more detail.
• divide_by - the alignment grid. Leave 32 unless you know your model wants something else (e.g. 56/112 for Qwen).
• block_empty_slots - this one's the quiet trap. On by default, so an empty slot returns an ExecutionBlocker and downstream nodes (Save Image, PreviewImage) are skipped rather than run with a blank. Turn it off when you want empty slots to pass None so downstream nodes with optional IMAGE inputs keep running.

The outputs mirror the inputs: image_1..3 (the resized references, blocker-or-None when a slot is empty) and conditioning (your input with one reference_latent appended per connected image).

Install

No extra dependencies - it's a pure-Python conditioning utility inside comfyui-timesaver.

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-timesaver
cd comfyui-timesaver
python -m pip install -r requirements.txt

Common issues

• "Nothing generates." Check block_empty_slots - if an empty slot is blocking the branch, that's it doing exactly what it says.
• "Reference looks wrong/cropped." If a thin band was cropped to fit the grid, that's by design (never stretched). And if the mask bbox landed on the wrong region, fix mask_orientation.
• "Doesn't work with my model." It only does two things - resize and VAE-encode - so it works with whatever model accepts reference_latents in its conditioning. If your model doesn't, this node won't make it.

The one honest caveat: this is scaffolding, not magic. It removes the tedious encode wiring from multi-reference workflows - the model still has to know what to do with the references, which is a Qwen-Edit / Flux-reference capability, not something this node provides.