GodnessMerger_OutputBlocksExperimental

This is GodnessMerger_OutputBlocks with one addition: a model input the plain version doesn't have. Everything else on the surface - seed, min, max, type, custom_value, and the GM_OUTPUT_BLOCKS output type - is identical between the two. So what does having the actual model buy the "Experimental" version? Honestly, the README (which is screenshots, not prose) doesn't say, and I'd rather tell you that plainly than invent an answer. My best-guess read: output_blocks in a real U-Net isn't one tensor, it's a stack of several sub-blocks (each with its own resnet and, in the higher-resolution parts, attention layers), and a node that can see the actual model could in principle assign a ratio per sub-block instead of one flat number for the whole decoder. That's inference from the shape of the problem, not something documented anywhere - flagging it as a guess rather than passing it off as fact.

What's solid ground

output_blocks is the U-Net's decoder - the upsampling path that reconstructs spatial detail after the bottleneck. It's where a lot of fine texture and rendering character lives, as distinct from middle_block's outsized pull on overall composition. That part comes from how the classic Stable Diffusion U-Net is actually structured, and it's true regardless of which of these two nodes you use.

The inputs

• model - the thing this variant needs and its sibling doesn't. Wire in one of the checkpoints headed into your merge.
• type - custom for an exact ratio, random for a reproducible one.
• min / max - bounds for the random draw.
• custom_value - read only in custom mode.
• seed - keeps random mode reproducible.

Output: GM_OUTPUT_BLOCKS, headed for a merge combiner elsewhere in the pack that isn't part of this batch of node briefs.

Installing it

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

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

Restart ComfyUI once it's cloned. No model download for this node itself - it takes a model you've already loaded elsewhere in the graph.

Where this bites

Forgetting to wire the model input is the one mistake this node makes easy to avoid noticing: unlike the plain OutputBlocks node, which will happily sit there producing a ratio nobody uses, this one needs a live model connected and will error immediately if you leave that slot empty - which, honestly, is more helpful than the alternative.

Beyond that: treat "Experimental" as the author's own warning label, not marketing. If you're not chasing a specific improvement over the plain OutputBlocks node, there's no reason to reach for this one first - start simple, and only swap in the model-aware version if you have a concrete reason to think the flat single-ratio approach isn't giving you what you want.