Scholar Feed

Parameter-efficient fine-tuning (LoRA family)

ZipLoRA

ZipLoRA: Any Subject in Any Style by Effectively Merging LoRAs

Superseded baseline#24 of 1,113 most-superseded · first seen Nov 22, 2023

Superseded — cited as a baseline and beaten by newer methods

6 papers critique it · 5 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites ZipLoRA as a baseline.

Nevertheless, the merging process often leads to interference between the parameters of different adapters~Ortiz-Jimenez_Favero_Frossard_2023. This oversight in failing to optimally align the integrated parameters can result in a notable performance degradation of the merged model, leading to ineffective preservation of the distinct qualities of both content and style~yadav2023resolving.
— Break-for-Make: Modular Low-Rank Adaptations for Composable Content-Style Customization
“as the $$-prediction loss tends to capture broad concepts rather than the precise global structure.”
— ConsisLoRA: Enhancing Content and Style Consistency for LoRA-based Style Transfer
“However, such approaches, along with methods like B-LoRA, ComposLoRA, CMLoRA, remain heavily dependent on model-specific architectures rombach2022high,podell2023sdxl and domain constraints.”
— QR-LoRA: Efficient and Disentangled Fine-tuning via QR Decomposition for Customized Generation
“However, ZipLoRA requires an additional optimization stage for each new combination of content and style, thereby restricting the flexibility of reusing trained LoRA weights, which is LoRA's primary advantage.”
— Implicit Style-Content Separation using B-LoRA
“merging the attention layers of two LoRAs at the element level could lead to a smoothing of style details and textures, or even the loss of object characteristics.”
— K-LoRA: Unlocking Training-Free Fusion of Any Subject and Style LoRAs
“However, this approach requires several minutes per merge at test time, limiting its usability in real-time scenarios.”
— LoRA.rar: Learning to Merge LoRAs via Hypernetworks for Subject-Style Conditioned Image Generation

Beaten on benchmarks

Head-to-head results where a newer method reports beating ZipLoRA. Values are copied from the source paper's tables — verify against the cited paper.

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.