Gen2 Apply QwenImage ControlNet (outdated)

The middle of the (outdated) QwenImage ControlNet chain: this node takes your diffusion model plus a loaded QwenImage ControlNet and a control image, prepares the VideoX-style control context, and outputs a wrapped model ready for this pack's sampler. It's the node that actually does the ControlNet wiring - "Apply" in the standard ComfyUI sense, but with QwenImage's specifics baked in. You don't use it outside this pack's pipeline; its output type is GEN2_WRAPPED_MODEL, which only the Gen2 QwenImage Control Sampler consumes.

How it works

The heavy lifting is building the control context. QwenImage's ControlNet conditions on a 132-feature context, and this node assembles it the way VideoX's QwenImageControlPipeline does:

• The control image is VAE-encoded (using the GEN2_VAE from Gen2 Load QwenImage VAE) into latents, normalized with the VAE config's mean/std.
• An inpaint mask (from your mask input) is processed into a mask latent; without a mask it defaults to all-ones.
• An inpaint_image (optional) is encoded to init latents; without one, zeros.
• The three - control latents, mask latent, init latents - are concatenated (16 + 1 + 16 = 33 channels) and packed into sequence format.

That context, scaled by control_context_scale, gets attached to the wrapped transformer along with the ControlNet's control layers, so the sampler can inject the condition at the right attention points during denoising. If your image isn't a multiple of 16, it's rounded down first - feed it sizes that already divide by 16 to avoid surprises.

Inputs and the one dial that matters

• model (MODEL) - your QwenImage diffusion model from a normal Load Diffusion Model.
• controlnet (GEN2_CONTROLNET) - from Gen2 Load QwenImage ControlNet.
• vae (GEN2_VAE) - from Gen2 Load QwenImage VAE.
• control_image (IMAGE) - the conditioning image.
• control_context_scale - default 0.8, range 0–2, step 0.05. This is the control strength, analogous to ControlNet weight elsewhere: higher = tighter adherence to the condition. The KB's controlnet essay is a good reminder that on modern DiT models the published weights run lower than the old SD-era 1.0, so 0.8 as a default is a sensible starting point.
• Optional: inpaint_image (IMAGE) and mask (MASK) - together they enable the inpaint mode, where the control context also carries the init image and mask.

Output: model (GEN2_WRAPPED_MODEL).

Installation

Needs the full QwenImage stack: VideoX-Fun as a custom node, diffusers, optionally ComfyUI-GGUF. Pack install via ComfyUI Manager (search "ComfyUI-gen2"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/petmycat/ComfyUI-gen2
cd ComfyUI-gen2
pip install -r requirements.txt

Common issues

• Output won't connect to a normal KSampler. Correct - GEN2_WRAPPED_MODEL only feeds this pack's sampler. That's by design.
• ControlNet has no visible effect. Check control_context_scale (too low = near-zero conditioning) and that the ControlNet weights are actually QwenImage ones (see the loader's article). Also confirm the control image isn't being downscaled oddly by the 16-px rounding.
• Missing VAE / VideoX errors. The Apply node needs the pack's GEN2_VAE, not a stock ComfyUI VAE. Wrong VAE → wrong latents → garbage context, silently.

This is the node where the outdated pipeline earns its keep if you're chasing VideoX-exact reproducibility. Otherwise, native QwenImage ControlNet options in the wider ecosystem will be simpler to wire.