ModelMergeSDXLDetailedTransformers

Open this node and you'll get roughly 700 input fields. No, that's not an exaggeration, and no, you don't need all of them. ModelMergeSDXLDetailedTransformers takes the per-transformer-block control of its sibling and drills one level deeper: inside every transformer block, you get separate weights for the query, key, value, output projections, the feed-forward network, and the normalization layers. It is the most granular SDXL checkpoint merge ComfyUI has ever had, and it exists because sometimes a merge recipe is "model2's keys and values, but model1's queries."

How it works

Same engine as the rest of the family - it subclasses ComfyUI's ModelMergeBlocks and blends model1 and model2 per key, matched by longest prefix. The difference is how deep the argument tree goes. Each transformer block in SDXL is expanded into 12 sub-weights:

norm1, attn1.to_q, attn1.to_k, attn1.to_v, attn1.to_out, ff.net, norm2, attn2.to_q, attn2.to_k, attn2.to_v, attn2.to_out, norm3

That's self-attention (attn1) and cross-attention (attn2), each split into the q/k/v projections and the output projection, plus the feed-forward and the norm layers. Apply that to the 10-transformer stacks at blocks 7/8 and the middle block, the 2-block stacks at 4/5, the mirrored output side, and the totals get silly fast.

The good news: because weights inherit from their enclosing prefix, you don't have to touch 700 things. Set input_blocks.8.1.transformer_blocks.3. and every sub-weight inside it follows. You only go down to ...attn2.to_k when you specifically want model2's keys there.

Inputs and output

• model1, model2 - the two checkpoints.
• The same coarse structure as ModelMergeSDXLTransformers - time/label embeddings, input/middle/output blocks, out. - with each transformer block expanded into the 12 sub-weights above.
• All FLOAT, 0–1, default 1.0 (which, as always in ComfyUI merging, means "take from model2").

Output is a single MODEL, ready for a sampler or SaveCheckpoint.

Who this is actually for

People doing checkpoint chemistry as a hobby, researchers, and anyone who's found that attention-level blends beat whole-block blends for a specific result - for example, mixing the key/value projections of a style model into a base while keeping its own query projections, which changes how the model attends without replacing its semantics wholesale. There's real technique here.

It's also a trap for beginners, and I'll say it plainly: you do not need this node to make a good merge. The regular ModelMergeSDXLTransformers gets you 95% of the benefit with maybe 30 knobs. If you open this, get overwhelmed, and never touch it again - that's the intended audience being filtered. Start coarse, and only drill in when you can name exactly what the extra control is buying you.

Installing it

ComfyUI Manager → search ComfyUI_experiments, or:

cd ComfyUI/custom_nodes
git clone https://github.com/comfyanonymous/ComfyUI_experiments

Restart ComfyUI. No dependencies, no model downloads; the node is under advanced/model_merging.

The honest take

It's the power tool in a pack of experiments, and it still feels like an experiment - a great one, but undeniably a kitchen sink. Fine for tinkering, essential for nobody. When you do reach for it, work top-down from the coarse weights, save a preset when you find something good, and remember that 700 sliders all at 1.0 still just means "model2."