Nodes/SP-Nodes/GodnessMerger_TimeEmbed
ComfyUI Node

GodnessMerger_TimeEmbed

Merging the block that tells the model 'how far along' it is

By bananasss00·Created 2 years ago·Updated 6 days ago· 20
GodnessMerger_TimeEmbed
    • GM_TIME_EMBED
    seed0
    min0.0
    max1.0
    type
    custom_value1.00

    Last of the five plain ratio-generator nodes in SP-Nodes' "Godness Merger" system, alongside LabelEmb, MiddleBlock, OutputBlocks, and Out. Like its siblings, GodnessMerger_TimeEmbed takes no model and produces no model - it outputs a single ratio, tagged GM_TIME_EMBED, for a merge combiner elsewhere in the pack to consume. That combiner isn't part of this batch of node briefs, so I can only tell you what feeds into it, not exactly how it gets used once it arrives.

    What time_embed actually is

    The name matches the real sub-module in the classic Stable Diffusion U-Net class: time_embed is the small MLP that turns the scalar diffusion timestep - "how many denoising steps are left" - into a conditioning vector that gets injected throughout the rest of the network. Every ResBlock in the U-Net reads from it. It's a tiny piece of the model in parameter terms, but it's the mechanism the network uses to behave differently at step 1 versus step 20 of the same sampling run - early steps laying down rough structure, later steps refining detail. Merging this block changes how a blended checkpoint interprets "where it is" in the schedule, which is a subtler effect than merging middle_block or output_blocks, but a real one if two source checkpoints were trained with meaningfully different noise schedules or step counts in mind.

    The mechanism

    Same shape as every other plain Godness Merger node: seed, min, max, type (custom or random), custom_value. custom mode hands back custom_value exactly. random mode draws a number between min and max, deterministic given seed - same seed, same draw, every run. No weights are touched here; it's a labeled number, nothing more.

    Fields to actually set

    • type - custom for a fixed ratio, random for a reproducible roll.
    • min / max - the 0–1 range random mode samples from.
    • custom_value - only relevant in custom mode.
    • seed - what makes "random" reproducible.

    Installing SP-Nodes

    Through Manager: search "SP-Nodes," install, restart. By hand:

    cd ComfyUI/custom_nodes
    git clone https://github.com/bananasss00/ComfyUI-SP-Nodes
    

    Restart once it's cloned. No model download here, no extra dependency - it's plain arithmetic.

    Troubleshooting

    Missing-node red box if the pack isn't installed when you load a workflow that references it - install via Manager, restart, resolved.

    The thing worth internalizing about this whole ratio-generator family: none of them do anything you can see on their own. If you've wired one up, run the graph, and nothing seems different, the first thing to check isn't the ratio value - it's whether the downstream combiner node is actually connected to this node's output at all. A disconnected GM_TIME_EMBED output fails silently rather than throwing an error, which is the most common way to lose time on this pack's merge tools. And as always with the random mode across this family: note the seed the moment you land on a result you like, because there's no way to recover it after the fact.

    CategorySP-Nodes/model_merging

    Inputs (5)

    NameTypeDefaultDescription
    seedINT00–18446744073709550000
    minFLOAT0.00–1
    maxFLOAT1.00–1
    typeCOMBO2 options: custom, random
    custom_valueFLOAT1.000–1

    Outputs (1)

    NameTypeDescription
    GM_TIME_EMBEDGM_TIME_EMBED