ComfyUI FLUX.2 Klein LoRA Loader

Architecture-aware LoRA loading for FLUX.2 Klein (9B) in ComfyUI, with automatic per-layer strength calibration based on forensic weight analysis.

Background

LoRAs trained against FLUX models are commonly shipped in diffusers format — separate to_q, to_k, to_v projections per attention layer. FLUX's native architecture stores these as a single fused QKV matrix, and single blocks fuse attention and MLP gate into a single linear1 projection. Loading these LoRAs without conversion means most attention weights never reach the model.

| What the LoRA ships with | What FLUX expects | What this pack does | |---|---|---| | Separate to_q / to_k / to_v | Fused img_attn.qkv / txt_attn.qkv | Block-diagonal fusion at load time | | Separate single block components | Fused linear1 [36864, 4096] | Fuses [q, k, v, proj_mlp] correctly | | Global strength only | Independent img/txt + per-single-block | Interactive graph widget + auto-calibration |

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/capitan01R/Comfyui-flux2klein-Lora-loader.git

Nodes

FLUX LoRA Loader

| Input | Type | Description | |---|---|---| | model | MODEL | FLUX.2 Klein / FLUX.1 model | | lora_name | dropdown | LoRA file from models/loras | | strength_model | float | Global LoRA strength (-20.0 to 20.0) | | auto_convert | boolean | Convert diffusers-format LoRAs to native FLUX 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 double blocks (8 columns, img purple / txt teal, split top/bottom) and single blocks (24 columns, green). Drag to adjust. Shift-drag moves all bars in a section. Global strength shown as a reference line.

FLUX LoRA Stack

Apply up to 10 LoRAs in sequence with independent strength, enable toggle, and auto-convert per slot.

FLUX LoRA Auto Strength

Reads the LoRA's weight tensors directly and computes per-layer strengths from the actual training signal in the file. Double blocks are analyzed with img and txt streams independently. One knob: global_strength.

FLUX 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)

Double blocks are processed with img and txt streams independently. Mean layer lands at global_strength.

Diffusers format fusion math

A_fused = cat([A_q, A_k, A_v], dim=0)          [3r × in]
B_fused = block_diag(B_q, B_k, B_v)            [3·out × 3r]

Alpha/rank scaling is pre-baked into B_fused before patching.

FLUX.2 Klein Architecture Reference

Double blocks (8 layers)
  img stream:
    img_attn.qkv    [12288, 4096]  (fused Q+K+V)
    img_attn.proj   [4096, 4096]
    img_mlp.0       [24576, 4096]
    img_mlp.2       [4096, 12288]
  txt stream:
    txt_attn.qkv    [12288, 4096]
    txt_attn.proj   [4096, 4096]
    txt_mlp.0       [24576, 4096]
    txt_mlp.2       [4096, 12288]

Single blocks (24 layers)
  linear1    [36864, 4096]  (fused Q+K+V+proj_mlp)
  linear2    [4096, 16384]

dim=4096  double_blocks=8  single_blocks=24