Donut Detailer 2 (DEPRECATED)
The K/S1/S2 flavor of the Donut model patch — and why it's on the way out
- model
- MODEL
Donut Detailer 2 is the second-generation formula in the pack's model-patch family, and it exists to answer a question nobody actually asked: "what if the detailer's knobs were called Multiplier, S1, and S2 instead of Scale, Weight, and Bias?" It patches the same three SDXL regions as the original - the first input block plus output blocks 0 and 2 - but derives the weight and bias multipliers through a different little formula.
As with every "detailer" in this pack, the name does not mean what you hope it means if you came from the ADetailer world. There is no face detection, no mask, no inpaint pass. It's a per-block model patch: it clones your model, multiplies selected weight and bias tensors by a computed factor, and hands the patched model to your sampler. If you're here to fix a blurry eye, the Impact Pack's FaceDetailer is the node you actually want.
The inputs and the formula
Nine required floats, three per region:
Multiplier_in,S1_in,S2_in- input blockMultiplier_out0,S1_out0,S2_out0- first output blockMultiplier_out2,S1_out2,S2_out2- third output block
The math, from the source: weight multiplier is 1 - (Multiplier * S1 * 0.01) and bias multiplier is 1 + (Multiplier * S2 * 0.02). So Multiplier is the master gain, S1 tunes the weight side, and S2 tunes the bias side, with the 0.01/0.02 factors keeping the effect from exploding. Defaults are Multiplier 0, S1 1, S2 2 - again, a no-op until you move something.
The practical difference from the Scale/Weight/Bias node is subtle: S1 and S2 decouple the two sides a bit more directly than Scale does, and the built-in scaling constants make the node less twitchy per unit of slider travel. Some people find the "2" node easier to dial in fine detail, some find the direct one more predictable. There's no winner - it's a matter of which control scheme your muscle memory prefers.
The honest verdict
This node is DEPRECATED. It stays registered so old workflows keep loading, but it's hidden from the Add Node menu. Its formula was folded into DonutDetailerUnified, which lets you select k_s1_s2 from a dropdown and get exactly these knobs (prefixed ks_) alongside the other two formulas. If you're building a new graph, use DonutDetailerUnified with the k_s1_s2 formula - same behavior, one less node installed.
If you load someone else's workflow that references Donut Detailer 2, it will still run; the deprecation is cosmetic. Just know that the pack's authors consider this an implementation detail, not the tool of the future. Install via ComfyUI Manager (search DonutNodes) or git clone into custom_nodes/ plus pip install -r requirements.txt with the same Python as ComfyUI. No model downloads. And as with all model patches, remember to re-run the graph after touching a slider - the preview won't update itself.
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| Multiplier_in | FLOAT | 0.00-100–100 | — |
| S1_in | FLOAT | 1.00-100–100 | — |
| S2_in | FLOAT | 2.00-100–100 | — |
| Multiplier_out0 | FLOAT | 0.00-100–100 | — |
| S1_out0 | FLOAT | 1.00-100–100 | — |
| S2_out0 | FLOAT | 2.00-100–100 | — |
| Multiplier_out2 | FLOAT | 0.00-100–100 | — |
| S1_out2 | FLOAT | 1.00-100–100 | — |
| S2_out2 | FLOAT | 2.00-100–100 | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |