Apply Advanced Flux ControlNet

This is the same job as the plain Apply Flux ControlNet node - take a loaded XLabs ControlNet plus a reference image and turn it into a conditioning object the sampler can use - but with two extra dials that turn out to matter more than people expect: start and end.

Quick refresher on why ControlNet exists at all: your prompt decides what gets generated, ControlNet decides where it goes, by feeding the model a spatial map - edges, depth, a pose skeleton - extracted from a reference image. XLabs shipped their own Flux ControlNets (separate Canny, Depth, and HED checkpoints, not a combined "union" model) right around Flux's August 2024 launch, making them one of the first ControlNet options for Flux at all. A worthwhile expectation-setter: a head-to-head community comparison against InstantX/Shakker's and MistoLine's Flux ControlNets found depth conditioning consistently good, openpose noticeably weaker, and rated the whole 2024 Flux ControlNet generation as "still a long ways off" the quality SDXL's later union model reached. That's not a knock on this node - it's honest context for what "Flux ControlNet" meant at the time this pack shipped.

The mechanism start and end control: ControlNet doesn't have to apply for the whole denoising process. start and end (both 0.0–1.0, defaulting to 0 and 1 - i.e. the whole run) let you window when the condition is enforced. The standard trick, going back to when this control first showed up in the SD 1.5 ecosystem and got treated as a precision breakthrough: apply the ControlNet only through the early steps that decide composition, then let it go so the model can add its own detail in the later steps instead of fighting the map the whole way through. Ending around the halfway point is the common starting guess if a result looks too rigid or over-baked.

The rest of the required inputs: controlnet is the loaded model from Load Flux ControlNet, image is your reference (canny edges, a depth map, whatever matches the checkpoint you loaded), and strength (0–10, default 1) is the overall weight - the community baseline from that same comparison thread was strength ~0.6 with Euler, Normal scheduler, and guidance 3.5, so treat 1.0 as a starting point to dial down rather than a target. There's also an optional controlnet_condition input, which is how you stack multiple ControlNets: feed the output of one Apply node into the controlnet_condition input of the next, and both conditions ride into the sampler together - canny for hard edges plus depth for spatial layout is the common pairing. The single output, also called controlnet_condition, plugs into XLabs Sampler's optional controlnet_condition input - this pack's sampler is the only thing that consumes it, it won't wire into a stock KSampler.

Installing it: search "x-flux-comfyui" in ComfyUI Manager, or cd ComfyUI/custom_nodes && git clone https://github.com/XLabs-AI/x-flux-comfyui and restart. You'll also want comfyui_controlnet_aux (Fannovel16's pack) installed alongside it if you need to generate the canny/depth/pose maps yourself rather than supplying pre-processed images - XLabs' own README calls this out as a separate requirement.

Common trouble: people report the ControlNet nodes in this pack going out-of-memory even with --lowvram set, and behaving oddly under memory pressure compared to ComfyUI's native nodes - worth remembering if your workflow used to run fine and suddenly starts hanging or refusing to free VRAM after you add one of these in. If the node import fails outright with something like No module named 'comfy.ldm.flux', that's an outdated ComfyUI core, not a broken install of this pack - updating ComfyUI itself (not just the custom node) is the fix people report working.