Apply Advanced Flux IPAdapter

Same job as Apply Flux IPAdapter - patch your Flux model so it takes cues from a reference image - but instead of one flat ip_scale for the whole run, this version lets the influence change shape across the sampling steps. That's not a gimmick: there's a real, well-documented reason to want it.

Community testing on IP-Adapter in general (going back to the SD1.5/SDXL era, but the underlying mechanism is the same idea here) found that running an adapter's influence across every sampling step measurably softens the output compared to running it only during the early steps and backing off once composition is settled - same logic as the start/end trick on ControlNet. A single flat strength can't do that; it's either on at that level or it isn't. This node can.

How the ramp works: begin_strength and end_strength set where the influence starts and ends (both range -100 to 100, defaulting to 0 and 1 - note that's a much wider range than the plain Apply node's 0–1 ip_scale, and the negative end of the range exists if you want to invert the effect rather than just fade it). smothing_type (yes, that's the actual field name - a typo in the node itself, not this article) picks the curve connecting those two values across the run: Linear is a straight ramp, First half concentrates the change earlier in the steps, Second half concentrates it later, and Sigmoid gives you a smooth S-curve - slow to start, fast through the middle, slow to settle. If you're trying to reproduce the "strong early, fade out" pattern the community found preserves more detail, start with begin_strength higher than end_strength and a First half or Sigmoid curve, rather than the reverse.

The rest of the inputs are shared with the plain version: model is your Flux checkpoint, ip_adapter_flux comes from Load Flux IPAdapter, and image is your reference. The single output is a MODEL - same deal as the non-advanced node, this patches attention into your model rather than producing a separate conditioning object, so wire it straight into whatever sampler you're using next.

When to reach for this over the plain Apply node: if a fixed strength is giving you the classic IP-Adapter trade-off - turn it up and you get better likeness to the reference but softer, less detailed output; turn it down and detail comes back but the reference influence gets weak - that's exactly the problem begin/end/smoothing exists to solve. Front-load the strength, taper it off, and you get more of both. If you don't care and just want a working default, the plain node's single ip_scale is less to think about.

Install and prerequisites are identical to the rest of the IP-Adapter half of this pack: install via ComfyUI Manager (search "x-flux-comfyui") or git clone https://github.com/XLabs-AI/x-flux-comfyui into custom_nodes and restart. You need an OpenAI CLIP ViT-L/14 model in ComfyUI/models/clip_vision/ and the XLabs Flux IP-Adapter checkpoint in ComfyUI/models/xlabs/ipadapters/ for Load Flux IPAdapter upstream to have anything to offer this node.

One honest caveat before you spend an evening tuning curves: this whole adapter was shipped as beta by XLabs, and the early community read on it was that it's noticeably weaker at holding a reference than the SDXL-era IP-Adapters people are used to. A smoothing curve gives you more control over a weaker signal - it won't turn a soft adapter into a strong one. If you need tight face or character consistency rather than loose style transfer, that's a different job this pack doesn't really solve; XLabs' IP-Adapter is aimed at general reference-image steering, not identity locking.