Reference ControlNet (Finetune) ππ π π
Tune attention and AdaIN separately
- CONTROL_NET
This is the power-user version of Reference ControlNet. The regular node gives you one style_fidelity and one ref_weight that apply to whichever reference mode you picked. This one splits every knob in two - one set for the attention path, one set for the AdaIN path - so you can, say, borrow a reference's color aggressively while only lightly nudging its structure, or the reverse. If you've ever found the plain Reference node either too heavy-handed or not selective enough, this is where you go.
It's the same underlying Reference mechanism, just with the two halves of it pulled apart into independent controls.
How it works
Reference works by sharing a reference image's features with your generation during sampling. There are two channels for that. The attention channel shares self-attention, which carries more of the reference's spatial/style character. The AdaIN channel matches feature statistics (mean and variance), which reads mostly as color and tone. The standard Reference node runs one of these, or both, with a single strength. This node exposes attention and AdaIN as two fully separate sub-controls, so you can dial each independently.
Same caveat as all Reference work: it's an SD 1.5 / SDXL technique and has no equivalent on the DiT bases. On a Flux, Qwen, Z-Image, or Anima workflow there's nothing for it to hook into.
The inputs and outputs that matter
Six required inputs, three per channel - and they mirror each other exactly:
attn_style_fidelity,attn_ref_weight,attn_strength- style fidelity, reference weight, and overall strength for the attention path (all default to 0.5 / 1.0 / 1.0).adain_style_fidelity,adain_ref_weight,adain_strength- the same three for the AdaIN path.
Setting a channel's *_strength to 0 effectively turns that channel off, which is how you isolate just attention or just AdaIN. Output is a CONTROL_NET for your apply node; as with all Reference nodes, the reference image itself goes into the apply node's image input.
How to install it
ComfyUI Manager: search ComfyUI-Advanced-ControlNet, install, restart. Or:
cd ComfyUI/custom_nodes
git clone https://github.com/Kosinkadink/ComfyUI-Advanced-ControlNet
then restart. No dependencies and no model to download - Reference is built into the pack.
Common issues & troubleshooting
This is overkill for what you're doing. If you just want "make it look like this image," start with the plain Reference ControlNet. Only reach for the Finetune node when the single-strength version can't separate what you want to keep from what you want to drop.
One channel is fighting the other. That's exactly what the split is for - turn one down. If color is right but the composition is bleeding through, cut attn_strength; if the structure is fine but the palette is off, work the adain_* set.
Nothing happens on a modern base. Reference, in every form, is SD 1.5 / SDXL only. There's no post-Flux version of it.
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| attn_style_fidelity | FLOAT | 0.500β1 | β |
| attn_ref_weight | FLOAT | 1.000β1 | β |
| attn_strength | FLOAT | 1.000β1 | β |
| adain_style_fidelity | FLOAT | 0.500β1 | β |
| adain_ref_weight | FLOAT | 1.000β1 | β |
| adain_strength | FLOAT | 1.000β1 | β |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| CONTROL_NET | CONTROL_NET | β |