Nodes/ComfyUI Extra Samplers/SamplerDPMPP_DualSDE_Momentumized
ComfyUI Node

SamplerDPMPP_DualSDE_Momentumized

DPM++ SDE with a second noise stream and momentum

By Clybius·Created 3 years ago·Updated 2 years ago· 104
SamplerDPMPP_DualSDE_Momentumized
    • SAMPLER
    noise_sampler_type
    momentum0.50
    eta1.00
    s_noise1.00
    r0.50

    SamplerDPMPP_DualSDE_Momentumized is the pack's "give DPM++ SDE a second noise engine and a memory" experiment. It's from ComfyUI Extra Samplers by Clybius, it shows up in the sampler dropdown after install, and it outputs a SAMPLER for custom sampling.

    First, the family. DPM++ SDE is the ancestral, noise-injecting member of the DPM family - non-converging, creative, a little unpredictable, and a long-time favorite for stylized SD 1.5/SDXL output. The "Dual SDE" modification, per the README, adds a second SDE on top, "akin to how 3M samples" - DPMPP 3M SDE runs an extra SDE stage for its third-order correction, and this node borrows that idea and applies it to the SDE path. So instead of one stochastic noise stream, you get two interacting ones.

    On top of that sits the "Momentumized" treatment the pack applies to several of its samplers: a momentum term (default 0.5, −1 to 1) carries the previous step's update into the current one. And there's one input the momentumized siblings don't have: r (default 0.5), which controls the balance between the two SDE streams - how much of the added SDE's character bleeds into each step.

    The inputs that matter:

    • eta - the ancestral noise strength. Default 1; lower toward 0 for a more deterministic run.
    • s_noise - scales the per-step noise injection. Default 1.
    • momentum - how much the previous step's update direction carries forward. Default 0.5.
    • r - the blend between the dual SDE streams. Default 0.5; this is the node's signature knob.
    • noise_sampler_type - the pack's usual menu (gaussian, brownian, pyramid, perlin, laplacian…). Gaussian to start.

    Output: a SAMPLER for your custom sampling chain.

    Honest framing: this is a tinkerer's sampler. The dual-SDE idea is real (3M is genuinely well-liked), and the momentum/r knobs give you a lot of texture and variety to explore on a fixed seed - but "dual stochastic streams" is precisely the kind of high-entropy setup where small changes produce big, hard-to-predict differences. The knowledge base's ancestral-vs-converging rule applies double here: you'll never fully settle on a stable image with this node, so it's for stylized experimentation, not for reproducing a result. Keep eta modest, start r at 0.5, and change one knob at a time. If you want the deterministic DPM++ experience, this isn't it - this is the SDE end of the family on purpose.

    Install is the shared pack routine - ComfyUI Manager → "ComfyUI Extra Samplers", or:

    cd ComfyUI/custom_nodes
    git clone https://github.com/Clybius/ComfyUI-Extra-Samplers
    

    then restart. No model downloads; dependency kornia.

    Categorysampling/custom_sampling/samplers

    Inputs (5)

    NameTypeDefaultDescription
    noise_sampler_typeCOMBO7 options: gaussian, uniform, brownian, highres-pyramid, pyramid, perlin, +1
    momentumFLOAT0.50-1–1
    etaFLOAT1.000–100
    s_noiseFLOAT1.000–100
    rFLOAT0.500–100

    Outputs (1)

    NameTypeDescription
    SAMPLERSAMPLER