Reference Latent+ for ComfyUI

A drop-in replacement for ComfyUI's stock ReferenceLatent node, with the controls I always wished it had.
Per-image strength, per-image timestep gating, per-image MediaPipe auto-masks, megapixel cap on attention cost, and 1–4 image inputs in a single node.

Works with Flux, Flux2/Klein, Lumina2, Z-Image, Wan, Hunyuan, Qwen-Image — anything that consumes reference_latents.

---

Why I built this

The stock ReferenceLatent is a tiny wrapper — it just attaches a latent to the conditioning's reference_latents key and walks away. That's fine for the obvious case (drop in a ref, get an output), but the moment I tried to push it I hit walls:

• No way to dial strength up or down — it's binary on/off.
• No timestep gating — the ref is active for the whole denoising schedule whether I want it or not.
• The ref's whole image leaks into output: pose, framing, background, even crop. There's no "use the face only, ignore the rest" knob.
• Multi-ref means wiring multiple LoadImage → VAE Encode → ReferenceLatent chains by hand.
• Bigger refs = quadratic attention cost, with no easy way to cap it.

I built Reference Latent+ as the node I wished existed. Each image gets its own strength, its own timestep range, and its own auto-mask config. Slots reveal progressively as you connect them. MediaPipe-driven masks for face / body / clothes / background are built in. There's a global cap on ref size to keep attention compute under control.

I tested all of this against Flux2/Klein 9B (the model I'm building Fizgig for), but the underlying mechanism is generic — it works on every model in ComfyUI that consumes reference_latents.

---

What's new vs stock ReferenceLatent

• Per-image signed strength (-5.0 … 50.0, default 0.85). 1.0 = stock strength. 0 = bypass that image entirely. <0 = experimental "anti-reference" — model pulled away from this ref's features.
• Per-image timestep range. Restrict any ref to a specific portion of the denoising schedule. Different refs can have different ranges; the node partitions the schedule so each timestep has exactly one set of active refs (no double-counting via overlapping conditioning).
• Per-image compositional masks with MediaPipe auto-detection. Tick face / hair / body / clothes / background to isolate just that part of the ref before encoding. Or wire your own MASK and it overrides.
• Three mask-fill modes — pixel_grey (in-distribution soft neutral), latent_zero (cleanest "ignore these tokens"), latent_noise (randomised attention shape). All three are novel; no existing ComfyUI node masks reference latents.
• Megapixel cap with aspect-preserving downscale. Default 1 MP — keeps attention compute proportional to a 1024² generation regardless of how big your source image is. Never upscales.
• 1–4 image inputs in one node. No manual VAE Encode wiring. Slots reveal progressively as previous ones get connected; collapsible per-image config groups appear only for connected images.
• VAE input on the node — handles VAE encode internally so I don't have to chain it.

---

Quick start

1. Drop the comfyui-ReferenceLatentPlus folder into ComfyUI/custom_nodes/.
2. (Optional, for the auto-mask feature) pip install mediapipe into your ComfyUI venv. The selfie_multiclass model auto-downloads on first use.
3. Restart ComfyUI. The node appears under advanced/conditioning > Reference Latent+.

There's a ready-made example workflow in examples/klein-9b-reference-latent-plus.json — drag it into ComfyUI to see the wiring (Flux2/Klein 9B Distilled, 4-step). Adjust the model filenames in the loaders to match what you have installed.

Wire it up:

LoadImage  ───► image_1     ┐
VAE        ───► vae         │
Prompt CLIP ──► conditioning ├──► Reference Latent+ ───► KSampler [positive]
                            │
(optional) LoadMask ──► mask_1

That's it. As you connect image_1, an image_2 slot appears. Connect that and image_3 shows up. And so on.

---

Inputs at a glance

Required

• conditioning (CONDITIONING) — your prompt's CONDITIONING.
• vae (VAE) — for encoding the input images.
• image_1 (IMAGE) — first reference (always required).
• max_megapixels (FLOAT, 0.1–4.0, default 1.0). Caps each ref by area. Aspect ratio preserved, never upscales. Lower = faster.
• mask_fill_mode (DROPDOWN). How masked-out regions are neutralised:
  • pixel_grey (default) — replace with mid-grey before VAE encode. In-distribution.
  • latent_zero — encode full image, zero the latent at masked positions. Cleanest "ignore" but off-distribution.
  • latent_noise — encode full image, fill masked latent positions with deterministic Gaussian noise.

Optional reference inputs (progressive disclosure)

• image_2 / image_3 / image_4 — additional refs. Slots reveal as you connect the previous one.
• mask_1 / mask_2 / mask_3 / mask_4 — explicit per-image masks. Override the auto-mask config if wired.

Per-image collapsible groups (one per connected image)

• strength (-5.0 to 50.0, default 0.85) — signed scaling of this image's ref latent.
• face / hair / body / clothes / background — MediaPipe-driven auto-mask regions.
• ignore_area (none / left / right / top / bottom) — exclude a side strip from the auto-mask.
• feather (0–200 px) — soften mask edges.
• grow (-200 to 200 px) — grow (>0) or shrink (<0) the mask area.
• start_percent / end_percent (0.0–1.0) — per-image timestep range. ComfyUI's percent: 0.0 = start of denoising (noisy), 1.0 = end (clean).

---

Mask precedence

For each image:

1. If mask_N is wired → use it, ignore that image's auto-mask config.
2. Else if any region boolean is ticked → run MediaPipe, generate mask.
3. Else → no mask, image used as-is (= what stock ReferenceLatent would do).

---

Empirical notes (Flux2/Klein 9B)

I built this against Klein and tested most of the defaults there:

• strength = 0.85 felt like the sweet spot for typical prompts. 1.0 was slightly too dominant; <0.5 too subtle.
• For style/identity, gate the ref to the late timesteps — try start_percent=0.6, end_percent=1.0. Klein's clean-end timesteps carry style/identity signal; the noisy-early timesteps drive composition. Restricting refs to the clean end lets your prompt's composition come through unconstrained.

On other models the mechanism still works but the sweet-spot values may differ — try strength=1.0 first, then dial.

---

Compatibility

The node attaches reference_latents to the conditioning dict. Models that read this key (and therefore work with this node):

• Flux family — Flux1, Flux2/Klein, LongCatImage
• Lumina2 / Z-Image
• Wan family — Wan21, Wan22, Wan21_HuMo, Wan22_Animate
• Hunyuan
• Qwen-Image

The optional reference_latents_method key is read by Flux-family models only. Others ignore it harmlessly.

---

Limitations / known gotchas

• MediaPipe is optional. If you don't have it installed, the auto-mask feature is silently a no-op — ticking a region does nothing. Explicit mask_N wiring still works.
• latent_noise and latent_zero are off-distribution. No model was trained with masked-out reference latents; behaviour at high mask coverage is exploratory. pixel_grey stays in-distribution and is the safe default.
• Negative strength is experimental. Models weren't trained on negative refs; results vary. Useful for "generate something that does NOT look like this" but unpredictable.
• The masking-of-ref-latents idea is novel. No other ComfyUI node does it (I checked). That means there's no community-validated baseline for what works — test on your own use cases.

---

Support the project

If this node saves you wiring time or unlocks something you couldn't do with the stock node, consider supporting development:

I also build Fizgig — a focused trainer, profiler, repair workbench and explorer for Flux 2 Klein 9B LoRAs. If you train LoRAs on Klein, take a look.