Your LoRA is 17 sliders, not a light switch

What it does Standard LoRAs apply a single strength value across all neural layers. This extension for AUTOMATIC1111’s WebUI exposes the internal blocks—17 for standard LoRA, up to 61 for Flux—so you can set weights per layer via prompt syntax. You can also schedule when a LoRA starts and stops during inference, or let random values shake specific blocks.

The interesting bit The real payoff is surgical control: a character or composition LoRA often does its job in the early sampling steps and then starts meddling with style. By cutting it off at step 10—or zeroing out the output blocks—you keep the pose and lose the unwanted paint-job. The built-in XYZ plotter and Effective Block Analyzer turn guesswork into a grid search.

Key highlights

• Supports block-level weights for LoRA, LyCORIS, SDXL, and Flux formats (12 to 61 blocks depending on the model).
• Prompt syntax like lbw=IN02 or comma-separated masks lets you target specific layers without touching others.
• Start, stop, and step-range scheduling lets you limit LoRA influence to specific diffusion timesteps.
• Random (R, U) and dynamic (X) placeholders for weights enable preset randomization or runtime overrides.
• Includes an XYZ plotting function and an Effective Block Analyzer to visualize which layers actually matter for a given LoRA.

Caveats

• The README warns that StabilityMatrix users may need to disable API mode to avoid unexpected behavior.
• XYZ plotting disables Hires. fix and forces batch size/count to 1.
• LyCORIS syntax is still shifting because upstream is under active development.

Verdict Worth installing if you’re tired of LoRAs leaking style into backgrounds or oversaturating outputs. If you only ever use single-strength LoRAs and don’t care which layer does what, you can skip it.