Nodes/ComfyUI-Apt_Preset/Stack_sample_data
ComfyUI Node

Stack_sample_data

Override Just the Sampling Settings You Actually Want to Change

By cardenluoĀ·Created 2 years agoĀ·Updated 21 days agoĀ· 309
Stack_sample_data
    • sample_stack
    ā—„steps0ā–ŗ
    ā—„cfg0.00ā–ŗ
    ā—„samplereulerā–ŗ
    ā—„schedulernormalā–ŗ

    The Apt_Preset pack's whole pitch is saving a full model+sampler configuration as a reusable preset, then tweaking individual pieces of it without rebuilding the graph from scratch. Stack_sample_data is the piece that lets you tweak sampling specifically - steps, CFG, sampler, and scheduler - on top of whatever a loaded preset already set, without needing to know or care what those baseline values were.

    The mechanism is worth understanding because it's a little different from a normal override node. steps and cfg both use 0 as a sentinel meaning "no change" - leave them at their default of 0 and whatever the preset (or upstream config) already had stays untouched; set either to a nonzero value and it overrides just that one field. It's a clean way to say "I like this preset, I just want fewer steps this run" without duplicating the whole preset and editing a copy. sampler and scheduler work the same conceptually, defaulting to euler and normal but pulled from a large enum of ComfyUI's full sampler and scheduler roster (45 samplers, 10 schedulers) so you can swap in dpm_2, karras, res_2s, or whatever else your workflow calls for.

    Picking values that make sense. If you're on a video model like Wan, the community's practical numbers are a reasonable starting point: 20-30 steps and CFG 3.5+ for a full-quality pass without a speed LoRA in the mix, or CFG 1.0 with 3-4 steps per pass if you are running a distillation LoRA like lightx2v or CausVid (CFG above 1.0 with those active tends to produce broken output, not just slower generation). res_2s is the quality-focused sampler in that world - noticeably slower than euler but generally regarded as getting more details right - while euler/beta is the fast, reliable default most people actually run day to day.

    Output: sample_stack, of the pack's own SAMPLE_STACK type, meant to plug into a downstream Apt_Preset sampler node alongside a loaded model preset, not into a stock KSampler directly.

    Installing it. No extra dependencies beyond the base pack: cd ComfyUI/custom_nodes && git clone https://github.com/cardenluo/ComfyUI-Apt_Preset and restart, or install through ComfyUI Manager by searching "ComfyUI-Apt_Preset." Run install.bat on Windows once after cloning - this node itself needs nothing extra, but it's the safe default across this pack.

    Where people trip up. The most common one is not realizing 0 is a special value on steps and cfg - if you actually want a sampler run at literally zero steps or zero CFG for some reason, this node can't express that; 0 always means "inherit from upstream" here, not "use zero." If your override doesn't seem to be taking effect at all, check that this node's sample_stack output is actually wired into the sampler that's reading it - like the other config nodes in this pack, it does nothing on its own if the connection downstream is missing. And if you're getting garbled or oversaturated output after dropping CFG low, double check you didn't leave a speed-distillation LoRA active at full strength alongside a full CFG value - that combination is a known source of broken generations across the Wan ecosystem, not something specific to this node.

    CategoryApt_Preset/stack/😺backup

    Inputs (4)

    NameTypeDefaultDescription
    stepsINT00–10000 0 == no change
    cfgFLOAT0.000–100 0 == no change
    samplerCOMBOeuler45 options: , euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, +39
    schedulerCOMBOnormal10 options: , simple, sgm_uniform, karras, exponential, ddim_uniform, +4

    Outputs (1)

    NameTypeDescription
    sample_stackSAMPLE_STACK—