On-Site Lora Training Questioned | Users Concerned About Outdated Models
Edited By
Carlos Mendez
A growing group of people are raising concerns that on-site Lora training is using outdated base models. The issue has caused confusion among users, particularly regarding character likeness transfer in fine-tuning processes.
Multiple individuals have noted that when training with the pony base or SDXL base, character likeness does not transfer effectively. One user remarked, "If I train directly on a modern fine-tune, it carries over to all other finetunes perfectly." This implies that those using older models may experience inconsistencies in character representation.
Models Under Fire
Pony Model Allegations
Many believe the pony model is a key concern, stating that it fails to carry over likeness correctly, raising questions about its effectiveness in modern settings.
SDXL Model Issues
Users mention that SDXL appears to be simply the original version and lacks any recent updates. "You’re getting exactly what it says on the tin," one commenter noted, pointing to the stagnation in SDXL's development.
Training Method Options
While people can choose specific fine-tunes for training at a higher cost, many feel it shouldn't be the only way to access updated models.
"I've done fewer tests on SDXL, but my main suspicion is on the pony model," stated a user, indicating shared frustration.
Mixed Sentiment
Overall, feedback reflects a blend of skepticism and determination. While some users believe the current system falls short, others seek solutions through targeted training choices. Many are pushing back against perceived limitations in model updates.
Key Insights
△ Users suspect older models are hindering performance.
▽ Training on modern fine-tunes shows better character likeness.
※ "If I train directly on a modern fine-tune, it carries over to all other finetunes." - User comment
As controversy intensifies, attention remains on whether platforms will address these concerns or continue with outdated offerings. Will users deal with ineffective models indefinitely, or is change on the horizon? Only time will tell.
Shifting Models Ahead
With the growing discontent among people regarding outdated models in on-site Lora training, there's a strong chance we will see platforms responding in the coming months. Many users are vocal about their dissatisfaction, which could lead to developers prioritizing updates to address these concerns. Experts estimate around a 60% probability that we will see newer models introduced or significant updates to existing ones, targeting more effective character likeness transfer by mid-2026. This shift may also encourage more competitive pricing strategies for fine-tunes, allowing broader access for all people.
Reflecting on Historical Change
A unique parallel can be drawn to the evolution of photography in the early 20th century when photographers transitioned from glass plates to film. Initially, many clung to outdated techniques, believing glass yielded better results. However, as discontent grew among artists and studios, the film industry surged forward, leading to groundbreaking advances in capturing images. Just like those early photographers, today’s AI developers might embrace innovation when faced with mounting demand for improved performance, pushing them to abandon outdated practices for something more effective.