Inference_Core_ModelMergeSDXLDetailedTransformers

This is the deepest level of checkpoint merging control this pack offers, and it's worth saying plainly upfront: most people don't need it. If Inference_Core_ModelMergeSDXL's per-block ratios already do what you're trying to do, stop there. This node exists for the narrower case where block-level isn't precise enough and you need to reach inside a specific transformer block and adjust a specific attention sub-layer on its own.

What "detailed" actually means here

Inference_Core_ModelMergeSDXL gives you one ratio per named UNet block. This node explodes several of SDXL's input blocks (the ones that actually contain transformer layers - SDXL doesn't put attention in every block) down into their individual components: for every transformer_blocks.N inside a block, you get separate sliders for norm1, the self-attention projections (attn1.to_q, attn1.to_k, attn1.to_v, attn1.to_out), the feed-forward layer (ff.net), norm2, the cross-attention projections (attn2.to_q, attn2.to_k, attn2.to_v, attn2.to_out), and norm3. Blocks with more transformer sub-layers (SDXL's deeper blocks stack several transformer_blocks entries) get correspondingly more sliders. Every one of them is the same FLOAT control as elsewhere in this pack's merge nodes: 0 to 1, step 0.01, default 1, meaning "how much of model2 to mix in at that exact sub-layer."

Why anyone bothers with this level of control: attn1 (self-attention) governs how the image attends to itself - internal coherence, texture, composition. attn2 (cross-attention) is where the model actually reads your text prompt and lets it steer generation. Those are genuinely different jobs, and a specific, documented reason to touch one without the other - say, wanting model2's prompt-following behavior via its cross-attention weights while keeping model1's internal image quality via its self-attention weights untouched - is exactly the kind of thing block-level merging can't express and this node can.

model1 and model2 are your two SDXL checkpoints. The output is a single merged MODEL.

The honest advice

Don't start here. Start with Inference_Core_ModelMergeSDXL's block-level control, and only drop down to individual attention layers once you've confirmed block-level granularity genuinely can't do what you're after - with this many independent sliders, tuning by trial and error is impractical, and you'll get much further with a specific hypothesis about which layer type you want to isolate than by nudging things at random.

Installing it

Ships with the pack. ComfyUI Manager: search "ComfyUI-Inference-Core-Nodes". Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/LykosAI/ComfyUI-Inference-Core-Nodes

then install.py, or pip install -e . and restart - pure tensor arithmetic on loaded checkpoints, no acceleration extras required.

Troubleshooting

Same architecture-mismatch rule as the other merge nodes here: this is SDXL-only, and an SD1.5 checkpoint will fail on the tensor shapes immediately rather than merge incorrectly. Given the sheer number of parameters, the more realistic failure mode is losing track of what you've actually changed - if a merge stops making sense, it's usually faster to reset every slider back to its default and reintroduce changes one attention type at a time than to debug a graph with dozens of non-default values already applied.