Apply Flux IPAdapter

This is the node that actually does something visible - Load Flux IPAdapter just opens the files, this one patches your Flux model so a reference image starts steering generation alongside your text prompt.

The mechanism is worth understanding rather than treating as a black box: this doesn't hand the sampler a separate conditioning object the way ControlNet does. It takes your model and returns a modified MODEL with the image-adapter's cross-attention patched in, so whatever you plug the output into downstream (a KSampler, XLabs Sampler, whatever) is generating with the reference image's influence already baked into the model itself. That's why the output type is MODEL rather than something ControlNet-shaped - treat this node as "model in, model out, but changed."

The four inputs, and the one that matters most: model is your Flux checkpoint, ip_adapter_flux is the loaded adapter object from Load Flux IPAdapter, image is your reference picture, and ip_scale is how strongly it leans on that reference. Here's the thing worth flagging: the default is 0.93, which is aggressive. Community guidance on IP-Adapter weight in general puts style transfer in the 0.6–0.8 range and warns that above roughly 0.7–0.8 the reference image starts overriding your prompt rather than complementing it. Shipping the default at 0.93 means out of the box this node is already tilted hard toward "just recreate the reference" - if your generations are ignoring your prompt entirely and just spitting out something close to the input photo, that's not a bug, drop ip_scale toward 0.4–0.6 and your text should start having a say again.

When to reach for this over the Advanced version: this one applies a single flat scale for the whole sampling run. If you want the adapter's influence to ramp up or down across steps instead of staying constant, that's what Apply Advanced Flux IPAdapter is for - this plain version is the simpler, faster-to-reason-about option when a flat strength is all you need.

Installing it: through ComfyUI Manager, search "x-flux-comfyui" and install - or manually, cd ComfyUI/custom_nodes && git clone https://github.com/XLabs-AI/x-flux-comfyui then restart ComfyUI (running python setup.py once from inside the folder covers the manual path). This node itself needs no extra download, but it's useless without its upstream loader being fed correctly: an OpenAI CLIP ViT-L/14 model in ComfyUI/models/clip_vision/ and the XLabs Flux IP-Adapter checkpoint (from XLabs-AI/flux-ip-adapter on HuggingFace) in ComfyUI/models/xlabs/ipadapters/.

Set expectations honestly: this is early, first-generation IP-Adapter tooling for Flux - XLabs' own docs flagged it as beta at release, and the community reaction backs that up. Reports at launch described results as noticeably weaker at holding a likeness than what people were used to from IP-Adapter-Plus on SDXL, and some users saw a contrast or sharpness shift that needed a color-match pass to clean up afterward. None of that means it's not worth using - reference-image conditioning on Flux is genuinely useful when it works - just don't expect one-shot perfect results the way a mature SDXL IP-Adapter workflow might give you. If you're chasing tighter face or character consistency specifically rather than general style, it's also worth knowing this XLabs adapter is not the same weights as Tencent's original IP-Adapter or its FaceID variants - those never shipped for Flux at all, so comparisons to "IP-Adapter on SD1.5" are comparing two different projects that happen to share a name.