In the image and video ecosystem, LoRA became the primary distribution unit for consistent characters, brand art direction, and camera motion styles on Civitai, Serverless GPU & Inference Clouds, and ComfyUI—supporting hot-swappable multi-LoRA serving over shared FLUX or video base Checkpoints.
LoRA & QLoRA
Also known as: LoRA · QLoRA · Low-Rank Adaptation · PEFT
Plain English
A lightweight 'patch' file (often tens of megabytes instead of gigabytes) that teaches an existing model a new style, character, domain, or skill without retraining or duplicating the entire base model.
Technical Overview
Low-Rank Adaptation freezes base model weights W and injects trainable low-rank decomposition matrices ΔW = A·B (rank r ≪ d) into attention and MLP projections. QLoRA pairs LoRA with 4-bit Quantization of the frozen base model, enabling 70B LLM fine-tuning on a single GPU; in image/video tools like ComfyUI and Civitai, multiple LoRAs can be stacked dynamically onto one checkpoint.
Chronological Entries (3)
Dettmers et al. introduced QLoRA (NF4 4-bit base Quantization + double quantization + paged optimizers), democratizing fine-tuning of 65B+ open LLMs and establishing Hugging Face PEFT as a standard training workflow.
Hu et al. at Microsoft introduced LoRA, demonstrating that weight updates during adaptation have low intrinsic rank. Training only tiny adapter matrices cuts trainable parameters by 1,000–10,000× while merging cleanly back into base weights with zero inference latency overhead.
Direct Relations (4)
- Alternative toCheckpoints
- UsesQuantization
- UsesSafetensors
- Part ofCivitai