Model Merge ZIT Blocks (DEPRECATED)

ModelMergeZITBlocks is a deprecated node, and the deprecation story here is a good one: it was the per-layer Z-Image Turbo merge node, and instead of just killing it, the pack folded it into ModelMergeZIT. That node has a granularity selector - set it to blocks and you get every slider this node had, byte-for-byte behavior, under the same hood. This page exists for people who meet the old name in a downloaded workflow.

What it did

Same fundamentals as ModelMergeZIT: merge two Z-Image Turbo / Lumina2 models by giving every part of the 30-layer transformer its own ratio from model1 to model2 (0.0 = model1, 1.0 = model2). The difference was granularity - this node exposed all 30 layer_0…layer_29 sliders individually, plus the embedders (x_embedder, t_embedder, cap_embedder), the refiners (context_refiner, noise_refiner), final_layer, norm_final, and other. No grouping, no training wheels. That's a lot of sliders, and it's exactly why the successor added the grouped mode - four group sliders for the 95% case, and the per-layer view for the rest.

If you still have this node in a saved workflow: it loads, it runs, the ratios mean the same thing. But for anything new, reach for ModelMergeZIT with granularity set to blocks - same sliders, same semantics, one less node to keep around, and you get the grouped mode as a bonus when you want it.

The practical guidance carries over unchanged: your base usually goes in model1, your style model in model2, and the layer ranges map to real jobs - layers 0–5 encode the prompt, 6–14 build composition, 15–23 carry detail, 24–29 refine aesthetics. Pull composition layers toward 0 to keep model1's structure, leave the late layers at 1 for model2's finish. And save your good results with the pack's DonutModelSave, because a hand-tuned per-layer merge is exactly the thing you don't want to lose to a reload.

Install

Part of ComfyUI-DonutNodes: ComfyUI Manager → search "DonutNodes" → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/DonutsDelivery/ComfyUI-DonutNodes.git donutnodes
cd donutnodes
python -m pip install -r requirements.txt

Dependencies are the pack's usual opencv-python-headless, scipy, matplotlib, psutil, tqdm, requests - no downloads beyond that, since the merge works on checkpoints you already have.