Apply Flux2 Fun Control

This is the node that does the actual ControlNet work in the Flux2 Fun chain: it takes your base Flux.2 Dev model, your loaded control branch, and your prepared 260-channel context, and produces a patched model you feed to the sampler. If you've got Load and Prepare wired up but nothing is wired into this node, you have no control happening at all - just a very expensive text-to-image.

How it works, without the scary part

The classic ControlNet idea - a trainable copy of the encoder blocks whose outputs get added into the frozen model's skip connections - doesn't survive contact with a DiT architecture like Flux.2. On these models the same job gets reimplemented against attention and MLP layers, and the gen2 pack does it the clean way: it clones your Flux.2 Dev MODEL, registers the managed control branch once, and composes clone-local replacements at double blocks 0/2/4/6. The key selling point is what it doesn't do: no flux.forward_orig replacement, no global monkey-patching of ComfyUI, no heuristic token-resizing. Every change lives on the cloned model instance, so running this graph doesn't silently corrupt every other Flux workflow in your install. That's a real virtue - a lot of ControlNet ports are just global patches waiting to break the next git pull.

The inputs that matter

• model - your Flux.2 Dev diffusion model. Only Dev; Klein and Flux.1 will be rejected upstream.
• control_model - the FLUX2_FUN_MODEL output from Gen2_LoadFlux2FunControlNet.
• control_context - the FLUX2_FUN_CONTEXT output from Gen2_PrepareFlux2FunControl.
• strength - 0 to 4, default 1. How hard the condition pushes. On a union ControlNet like this, 1.0 is a sane start; drop toward 0.3–0.7 for loose guidance, and the KB's standing advice applies here more than anywhere: once the composition is formed, you don't need the condition shouting anymore.
• start_percent / end_percent - when during denoising the control applies, 0–1. This is the parameter ControlNet veterans fiddle with most. Holding the condition for the full sample can over-constrain detail; many structure-heavy workflows release it early (e.g. start 0, end 0.5).
• control_group (optional) - feed in a FLUX2_FUN_CONTROL_GROUP from the Combine node to stack multiple controls.

The output is a patched model - wire it into your sampler the same way you'd wire the unpatched one.

Install

ComfyUI Manager: search ComfyUI-gen2. Or:

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

Restart. You need ComfyUI v0.28.0+ (this section is pinned to a specific recent commit), the Flux.2 Dev stack, and the 2602 checkpoint in models/controlnet. The Flux2 Fun section itself has no extra Python deps.

Where people get burned

Two things. First, VRAM - you're loading Flux.2 Dev (the 56B stack) plus a control branch that the loader never frees. Budget for it before you blame the node. Second, temper your expectations: the author states plainly that real-weight end-to-end parity is not yet claimed and the full GPU matrix hasn't passed. The architecture is deliberate and the harness exists, but treat first results as promising rather than gospel.