Architecture-aware LoRA loading for Z-Image Turbo (Lumina2) in ComfyUI, with automatic per-layer strength calibration based on forensic weight analysis.

Background

LoRAs trained against Z-Image Turbo are commonly shipped with separate to_q, to_k, to_v projections — the standard diffusers export format. Z-Image Turbo's native architecture stores attention as a single fused QKV matrix. Loading these LoRAs without conversion means the attention weights never reach the model.

What the LoRA ships with	What Z-Image Turbo expects	What this pack does
Separate `to_q` / `to_k` / `to_v`	Fused `attention.qkv` `[11520, 3840]`	Block-diagonal fusion at load time
`to_out.0` naming	`attention.out` naming	Remaps automatically
Any prefix convention	Lumina2 exact key mapping	Uses `z_image_to_diffusers()`
Global strength only	Per-layer control	Interactive graph widget + auto-calibration

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/capitan01R/Comfyui-ZiT-Lora-loader.git

Nodes

Z-Image Turbo LoRA Loader

Input	Type	Description
`model`	MODEL	Z-Image Turbo / Lumina2 model
`lora_name`	dropdown	LoRA file from `models/loras`
`strength_model`	float	Global LoRA strength (-20.0 to 20.0)
`auto_convert_qkv`	boolean	Fuse separate Q/K/V to Z-Image's native fused QKV format
`lora_name_override`	string (link)	Optional — overrides the dropdown when connected
`layer_strengths`	string (link)	Optional — per-layer JSON from Auto Strength node

The graph widget shows 30 columns, one per transformer layer, each split between attention (purple, top) and feed-forward (teal, bottom). Drag to adjust. Shift-drag moves all active layers together. Click to toggle.

Z-Image Turbo LoRA Stack

Apply up to 10 LoRAs in sequence with independent strength, enable toggle, and QKV fusion per slot.

Z-Image LoRA Auto Strength

Reads the LoRA's weight tensors directly and computes per-layer strengths from the actual training signal in the file. One knob: global_strength.

Z-Image LoRA Auto Loader

Self-contained version of the above — analysis and application in one node. model in, patched model out.

How Auto Strength works

For every layer pair in the file:

ΔW = lora_B @ lora_A
scaled_norm = frobenius_norm(ΔW) * (alpha / rank)
strength = clamp(global * (mean_norm / layer_norm), floor=0.30, ceiling=1.50)

High-signal layers get pulled back, low-signal layers get nudged up, mean lands at global_strength. Layer discovery is fully dynamic.

Z-Image Turbo Architecture Reference

30 transformer layers
  attention
    qkv.weight      [11520, 3840]  (fused Q+K+V)
    out.weight       [3840, 3840]
    q_norm.weight    [128]
    k_norm.weight    [128]
  feed_forward
    w1.weight        [10240, 3840]  (SwiGLU)
    w2.weight        [3840, 10240]
    w3.weight        [10240, 3840]
  attention_norm.weight    [3840]
  ffn_norm.weight          [3840]
  modulation.linear.weight [15360, 3840]  (AdaLN)

dim=3840  n_heads=30  n_kv_heads=30  head_dim=128