LoRA Fine-Tuning Adoption
last edited
LoRA fine-tuning adoption shows three patterns: a clear winner in FaceID, a rough consensus on dataset size, and a split decision between DreamBooth and LoRA. Related histories appear in technique adoption timelines, when techniques became standard, and Civitai Training Post-Mortems.
Why FaceID won
IP-Adapter FaceID became a preferred character-consistency technique. The IP-Adapter-FaceID model card lists updates from December 27, 2023. The IP-Adapter repository records a January 4, 2024 release.
FaceID uses a face-recognition ID embedding rather than a CLIP image embedding. An auxiliary LoRA helps train face ID embeddings.
IPAdapter Plus, a ComfyUI extension, added FaceID support on December 22, 2023. Reference images can replace a separate LoRA for each character. IP-Adapter + ControlNet combined workflows documents pairing identity with pose control.
LoRA dataset size
Guidance for LoRA fine-tuning favors quality over quantity. The practical floor is 10–15 images. The preferred range is 20–30 images. Roughly 50 is the ceiling. LoRAs commonly learn one subject from 25–30 images.
One poor image among 20 can reduce LoRA output quality. Curated sets spanning poses, angles, expressions, and lighting can outperform 100 inconsistent images. Relevant software appears under training & fine-tuning tools.
DreamBooth versus LoRA
DreamBooth fine-tunes a full model around a subject. LoRA trains small adapter weights instead. DreamBooth remains stronger for realism, generalization, and faces.
LoRA fine-tuning offers operational advantages. Training takes minutes rather than tens of minutes. Files can approach 5 MB instead of several gigabytes.
A common division uses DreamBooth for high-fidelity likeness and LoRA for styles, rapid training, and compact sharing. Their spread appears in platform-specific technique spread and CivitAI.
Sources
- IP-Adapter-FaceID model card lists variants, helper adapters, and update dates.
- IP-Adapter repository records FaceID models and release history.
- Diffusers DreamBooth guide covers dataset preparation and training setup.
- Diffusers LoRA guide describes LoRA training workflows and checkpoints.