Apply Ultimate ControlNet (FLUX)

If you came up in A1111, you know the three-way toggle on every ControlNet unit: Balanced, "My prompt is more important", and "ControlNet is more important". ComfyUI's built-in apply nodes never bothered to bring it over. This node - "Apply Ultimate ControlNet (FLUX)" from the colorfix-v3 pack - is that dropdown, rebuilt for Flux.

It does the same wiring job as the core ControlNetApplyAdvanced: take a CONTROL_NET, an IMAGE, and positive/negative conditioning, and hand back conditioning with the controlnet baked in. What it adds is control_mode, and that's the whole reason to use it over the core node.

How the modes actually work

The A1111 trio comes from sd-webui-controlnet, and two mechanisms power it. "My prompt" applies soft injection - an exponential decay across the controlnet's layers (0.825 per layer in A1111) so the deep structural layers keep full weight while the shallow detail layers fade out. "ControlNet" adds cfg injection - zeroing the control signal on the uncond batch so CFG amplification pushes the control harder.

Both of those assume a UNet running real CFG. Flux runs at cfg_scale=1.0 with guidance baked into the model, and it's a DiT with ~57 control injection points. There's no uncond path to zero. So this node swaps cfg injection for symmetric layer scaling: "My prompt" decays the shallow layers toward a ~6% minimum weight, "ControlNet" boosts them up to ~1.5x. Same idea, no CFG dependency.

The inputs that matter

• strength - the one you'll actually tune. The range goes to 10, but on Flux you'll live around 0.3–0.9. Modern Flux controlnets are calibrated for lower weights than the SD-era 1.0 default, and overcooking it is the most common failure mode.
• start_percent / end_percent - denoising fractions for when the control applies. Running the whole sample is fine for tile; for pose or depth you'll often release the condition around 0.6 once composition has formed.
• control_mode - Balanced is a safe default. Reach for the other two when the prompt and the control are fighting over the image.

positive and negative come from CLIP Text Encode, image is your condition (the raw image for tile, or a preprocessed map), and the optional vae only matters if your controlnet expects latent input. Outputs are the same two CONDITIONING wires, straight into your KSampler.

Install

The pack is pure Python with no requirements.txt and no model downloads - you're installing some blur kernels and a callback, not a dependency tree.

cd ComfyUI/custom_nodes
git clone https://github.com/ihorpankin/comfyui-colorfix-v3.git

Restart ComfyUI and it shows up under colorfix-v3/flux. ComfyUI Manager works too - search "comfyui-colorfix-v3".

Where people get burned

The controlnet you load has to be a Flux controlnet; no SD-era file loads on Flux, ever. And if you came for tile upscaling, this node pairs naturally with the pack's Flux Tile ColorFix Patcher - and note the pack's own Flux example workflow actually wires up the plain SD/SDXL apply node instead of this one. The FLUX variant is a drop-in swap, so don't overthink which one the workflow calls for.