Flux Ref Latent Weight ⚡

Some nodes are flagship features. Flux Ref Latent Weight is the utility version of Flux Ref Latent Controller: a single flat multiplier for one reference's K/V attention tokens, model in, model out, nothing else. No conditioning input, no spatial fade, no debug flag. It's the node you reach for when you know exactly what you want - "reference 2 is pulling too hard, turn it down to 0.7" - and you can't be bothered with the controller's extra knobs.

The pack's README calls it "the simple flat-multiplier case of the above, model-only." That's the whole pitch, and it's a good one: of the nine nodes here, this is the least likely to surprise you.

How it works

It's the same set_model_attn1_patch mechanism as Flux Ref Latent Controller, minus everything optional. The node clones your model, and every attention block scales the K and V vectors of one reference's token slice by a constant weight. The token slice is located using reference_image_num_tokens from comfy's extra_options - the same bookkeeping the controller uses - so it works on any Flux-family checkpoint (Flux.1, Kontext, Klein). Where the controller can apply a spatial gradient, this applies a plain number.

Because it needs no conditioning input, you don't have to think about conditioning plumbing at all. It only needs to know which reference (reference_index) and how hard (weight). If your model's conditioning already carries reference latents from a Flux Multi Reference Latent (or any other reference-attachment node), this patches straight on top.

The inputs that matter

There are exactly three, and you'll set all of them:

• model - your Flux-family model.
• reference_index (0–7) - which attached reference to weight, matching the order of attachment.
• weight (default 1.0, range 0–5) - the flat multiplier. 0 silences the reference entirely, 1 leaves it untouched, above 1 amplifies.

Output is a single MODEL to feed your KSampler.

Where it sits in the workflow

Anywhere between your model load and the KSampler. Chain a few of these (one per reference) for a quick-and-dirty per-reference strength setup, then graduate to Flux Ref Latent Controller if you later want spatial control. If you have a controller already doing the job, you don't need this - but if you just want the dial and none of the ceremony, this is it.

Install

Same one-line install as the rest of the pack - no dependencies beyond what ComfyUI already provides, no model files:

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-Flux-Reference-Tools

Restart ComfyUI, search "Flux Ref Latent Weight" under 🤖 CCTech/Flux Reference, or install via ComfyUI Manager ("ComfyUI-Flux-Reference-Tools").

Common issues

• Moving weight does nothing. The reference isn't actually attached to the conditioning your model is sampling with - check that a reference-attachment node ran, and that reference_index points at a reference that exists. This node can't tell you (there's no debug flag); it just patches a slice that may be empty.
• Weight above 1 washes the image out. Amplified reference K/V past ~1.5-2x tends to over-inject. Keep the range 0.5–1.5 unless you're deliberately fighting another patch.

It's the thinnest node in the pack, and that's its virtue. When a reference is bleeding through too hard and you want a clean fix with no new knobs to learn, this is the one.