SamplerDPMPP_SDE_SUN
DPM++ SDE with the Brownian noise swapped for uniform
- SAMPLER
SamplerDPMPP_SDE_SUN was the SUNoise take on DPM++ SDE, the true stochastic solver. Where the ancestral samplers sprinkle noise on top of a deterministic walk, DPM++ SDE treats the whole trajectory as a stochastic differential equation, using a Brownian tree noise sampler to keep every step genuinely random. And if there's one sampler in this pack where the "your noise has the wrong distribution" argument bites, it's this one - this node replaced that Brownian noise with scaled uniform noise on [-sigma_down, +sigma_down], seeded deterministically from the sampling seed.
The catch, same as all six of these pages: it's gone. This node shipped only in the pack's first months (April–July 2024) and the "Complete rework, breaking changes" update deleted every per-sampler node. Current installs give you v2.0.0, which has no SamplerDPMPP_SDE_SUN - only SamplersSUNoise and SamplersSUNoiseAdvanced, both with dpmpp_sde in the dropdown.
The extra input nobody else had
Besides eta (1.0) and s_noise (1.0), this node exposed r (default 0.5). In the stock DPM++ SDE solver, r positions where the noise is sampled within each step; 0.5 is the standard midpoint choice. It's a fine-tuning knob - lower r changes how early in the step the stochasticity kicks in. The other SUN sampler nodes didn't have it, because it's specific to this solver's structure. If you were dialing r, you were deep in the weeds in a good way.
Output was the usual SAMPLER, into SamplerCustomAdvanced.
What to do with an old workflow
Swap the missing node for SamplersSUNoise → sampler_name = dpmpp_sde, and pick standard or multires noise. You lose the direct r control unless you use the Advanced selector (where r still exists, default 0.5), so nothing is actually lost. Bonus: the current code interleaves SUNoise with standard Gaussian on alternating steps, so you get the bounded-noise benefit without DPM++ SDE's tendency to get too stochastic.
Pin git checkout efc6e38 in the custom_nodes folder if you must reproduce the original behavior exactly. Most people shouldn't bother.
Install, briefly
cd ComfyUI/custom_nodes
git clone https://github.com/bvhari/ComfyUI_SUNoise
Restart. ComfyUI Manager → search "SUNoise" also works. No dependencies, no model files, GPL-3.0. The pack has no real community footprint, so the versioning story on this page is the whole troubleshooting section.
Bottom line
DPM++ SDE is a gorgeous, slow, noisy sampler - the kind that produces variety you can't get from converging solvers. SUNoise was a reasonable attempt to clean up its noise, and the modern selector node keeps the idea alive with better interleaving. One dropdown swap and you're current. And yes: 30+ steps, as the README insists.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| eta | FLOAT | 1.000–100 | — |
| s_noise | FLOAT | 1.000–100 | — |
| r | FLOAT | 0.500–100 | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| SAMPLER | SAMPLER | — |