Nodes/SP-Nodes/GodnessMerger_OutputBlocks
ComfyUI Node

GodnessMerger_OutputBlocks

Merging the decoder half of the U-Net

By bananasss00·Created 2 years ago·Updated 10 days ago· 20
GodnessMerger_OutputBlocks
    • GM_OUTPUT_BLOCKS
    seed0
    min0.0
    max1.0
    type
    custom_value1.00

    Another ingredient in SP-Nodes' modular "Godness Merger" system: GodnessMerger_OutputBlocks takes no model, produces no model, and exists purely to decide what merge ratio to apply to one specific chunk of a U-Net. The result comes out as GM_OUTPUT_BLOCKS, a typed value that only means something once it reaches a combiner node elsewhere in the pack - this batch of briefs doesn't cover that node, so consider the wiring beyond this point a gap, not a secret I'm withholding.

    Why output_blocks is worth its own control

    The name matches the real sub-module in the classic Stable Diffusion U-Net: after the encoder (input_blocks) squeezes the image down and the bottleneck (middle_block) sits at the lowest resolution, output_blocks is the decoder - the upsampling path that reconstructs spatial resolution and detail on the way back out. It's a bigger chunk of the network than time_embed, label_emb, or out, and it's where a lot of fine texture and detail-level character gets decided, as opposed to middle_block's outsized influence on composition and overall structure. A common merge instinct - take one checkpoint's overall "shape" and another's finer rendering style - leans on exactly this kind of split: keep more of one model's earlier blocks, lean toward the other's output_blocks.

    The mechanics

    Same recipe as every non-experimental Godness Merger ratio node: seed, min, max, type (custom or random), custom_value. custom mode returns custom_value as-is. random mode draws a number between min and max, deterministic from seed. Nothing here touches actual weights - it's a ratio generator, full stop.

    Inputs worth knowing

    • type - custom for a fixed ratio, random for a reproducible roll.
    • min / max - the range random mode draws from (0–1).
    • custom_value - only read in custom mode.
    • seed - pins the random draw so "random" is actually reproducible.

    Installing SP-Nodes

    Via ComfyUI Manager: search "SP-Nodes," install, restart. Manual:

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

    Restart afterward. No dependency, no model download - this node is pure ratio bookkeeping.

    What to watch for

    The usual missing-node red box appears if you load a workflow that uses this before the pack is installed - install through Manager, restart, and it resolves.

    Because output_blocks is a fairly large chunk of the network to be governed by a single scalar, small changes to min/max here can move results more than the equivalent range would on GodnessMerger_Out or GodnessMerger_TimeEmbed. If a merge is coming out closer to one parent than you expected, this is one of the first ratios worth checking - right alongside whatever your combiner's overall strength setting is doing. And as with every random-mode node in this pack: if a particular seed produces a merge you like, write it down. There's no way to reverse-engineer it from the output model afterward.

    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_OUTPUT_BLOCKSGM_OUTPUT_BLOCKS