Practical AI Guides

English

简体中文

English

简体中文

Appearance

LoRA Usage: Style Stacking and 4-Step Acceleration ​

ComfyUI LoRA lets you switch styles without changing the base model, and can also compress sampling from twenty-plus steps down to 4 steps. This page walks you through three techniques in 10 minutes.

Goal ​

In one sentence: chain LoRAs into your existing text-to-image workflow — first load a single LoRA to switch styles, then stack multiple LoRAs, and finally use a Lightning-type acceleration LoRA to compress sampling steps down to 4 steps.

Preparation ​

  • Have the default workflow from your first generation working.
  • Download 1–2 LoRA(s) in .safetensors format from Civitai or Hugging Face, and note which base model they are made for.
  • If you want to practice acceleration: also prepare a Lightning-type acceleration LoRA matching your base model; for video (e.g. Wan), look for lightx2v acceleration LoRAs.

Follow Along ​

Step 1: Place the File and Hook Up the Load LoRA Node ​

Put the LoRA file into the ComfyUI/models/loras/ folder, then return to the interface and refresh (or restart ComfyUI). Double-click the canvas, search for and add the Load LoRA node: connect Load Checkpoint's MODEL and CLIP outputs to its model and clip inputs respectively, then connect its MODEL and CLIP outputs to where KSampler and CLIP Text Encode originally were. Keep strength_model and strength_clip at the default 1.0 for now.

Expected result: the queue runs normally and the output carries the LoRA's style; lower the strength to 0.5 and run again — the style visibly fades, showing the LoRA is taking effect.

Step 2: Chain Multiple LoRAs to Stack Styles ​

Drag out another Load LoRA node, connect the first node's MODEL and CLIP outputs to the second node's inputs to form a chain. Select different LoRAs for the two nodes, setting each strength to around 0.7.

Expected result: with the same set of prompts, the output carries the traits of both LoRAs at once; if one style is too strong, lower that one's strength alone.

Step 3: Use a Lightning-Type LoRA for 4-Step Acceleration ​

Swap in a Lightning acceleration LoRA, lower the KSampler's steps to 4, set cfg to the low 1–2 range@@I18N_KEEP_000031@@, and choose the sampler and scheduler according to that LoRA's instructions (Lightning commonly uses euler + sgm_uniform@@I18N_KEEP_000032@@). The step count must match the LoRA version: for the 4-step version, use steps=4. The same applies to video models — the Wan series can use lightx2v acceleration LoRAs, with CFG set to 1@@I18N_KEEP_000033@@.

Expected result: with the same prompts, generation time is greatly shortened, and image quality is close to the original twenty-plus-step sampling.

Verify ​

  • Single LoRA takes effect: generate one image each at strength 0 / 0.8 / 1.2; the style should get progressively stronger.
  • Stacking is controllable: set one LoRA's strength in the chain to 0, and the corresponding style should disappear from the image.
  • Acceleration takes effect: record the time and step count before and after acceleration with the same prompts, and confirm that 4-step output is usable and not overexposed.

Common Pitfalls ​

  • LoRA not found in the dropdown: the file must be placed in models/loras; after placing it, refresh or restart ComfyUI.
  • Base model mismatch: a LoRA is trained on a specific base model; pairing it with the wrong base model gives poor results or even weird images, so check the compatibility notes when downloading.
  • Acceleration LoRA without lowering CFG: keeping your usual CFG of around 7 will cause overexposure and breakdown; the step count must also match the 4-step/8-step versions.
  • Multiple LoRAs all maxed out: with both strengths at 1.0, they easily clash — start each at 0.7 and fine-tune from there.

Next Lesson ​

Once you can freely stack styles, the next step is ControlNet, which uses line art, depth, and pose to directly control composition: ControlNet Basics: Line Art, Depth, and Pose for Composition Control.

Recommended for practice: Qwen-Image Text-to-Image, Qwen-Image-Edit Instruction-Based Editing, and Wan2.2 14B Text-to-Video (for video acceleration, pair with lightx2v).