Donut Detailer 4 (DEPRECATED)

Donut Detailer 4 is the simplest member of the pack's model-patch family, and simplicity is both its virtue and its trap. Where the original detailer wraps its knobs in a Scale/Weight/Bias formula and "Detailer 2" uses the K/S1/S2 scheme, this one just multiplies: Weight multiplies the weights of a block, Bias multiplies its biases. What you set is what the model gets. No derived multipliers, no scaling constants, no hidden math.

Which means - repeat after me - this is not an image detailer. No detection, no masking, no inpainting. It's a per-block model patch: it clones your model, multiplies the weight and bias tensors of the first input block and output blocks 0 and 2, and returns a MODEL that renders with more or less high-frequency detail until you change it. If you googled your way here hoping to fix a mushy face, the Impact Pack's FaceDetailer is the tool you actually want.

The inputs

Six required floats, two per region:

• Weight_in, Bias_in - input block
• Weight_out0, Bias_out0 - first output block
• Weight_out2, Bias_out2 - third output block

All default to 1.0, which is a perfect no-op: multiply by 1 and nothing changes. That's the honest state of this node - it does nothing until you make it do something, and it's refreshingly predictable about it. Want the model 10% stronger on the first output region's weights? Set Weight_out0 to 1.1. Done.

The _in region governs how the latent noise enters the network; the _out0/_out2 regions are where the fine detail response lives, so they're the knobs that produce visible crispness or softness. The reason this family of patches exists at all is that the denoiser's early and late blocks are disproportionately responsible for texture scale - boosting them is a poor man's version of what a block-weight LoRA does, without training anything.

Why it's deprecated (and what to use instead)

This node is marked DEPRECATED - registered so old workflows load, hidden from the Add Node menu. Its formula was merged into DonutDetailerUnified under the direct option (the dir_Weight_* / dir_Bias_* inputs). New graphs should use that node and pick direct from the dropdown; you get exactly these six knobs without installing the fossil.

Worth a note: "direct" doesn't mean "safe." Because there's no formula damping the effect, small values can be too small and 1.2 is a lot more aggressive here than on the K/S1/S2 node, where the 0.01/0.02 factors keep everything gentle. Start at 1.05 and creep up. Overshoot and you'll get crunchy, oversharpened textures that look like a bad unsharp-mask.

Install as part of the pack: ComfyUI Manager → search DonutNodes, or git clone the repo into custom_nodes/ and pip install -r requirements.txt with the same Python that runs ComfyUI. No model downloads. Re-run the graph after any slider change - patches apply at run time, not retroactively.