Crafting Character Lora | New Debate on Quality Emerges
A growing trend of people creating their own character Lora faces using AI tools has sparked fresh debates over image quality and consistency. Recent comments on forums reveal varying opinions about the reliability of datasets that many users are currently employing.
The Mechanics of Creating Lora Characters
People are sharing methods to generate face images using tools like Z-image combined with BFS (Best Face Swap) Lora and the Flux2Klein model. A common guideline suggests about 20 images for training a dataset. However, many argue this is far from sufficient.
One forum member stated, "You can make a Lora with 10 maybe less if the dataset quality is decent," while another disagreed, asserting, "20 might not be enough to get consistency." Interestingly, users are also offering insights into how variation can occur among generated images. A comment mentioned, "By the standards of THIS group, your images rate about 2/10 due to easily visible artifacts."
Quality Issues Resurface
Concerns about the quality of generated faces persist. Many users have pointed out that changes in facial structure can lead to drastic discrepancies, with one likening the output to posting "home snapshots to a top tier photography group." Users highlight how inconsistent results can undermine the overall project.
"You have at least a 15% variation. So you're gonna get ROASTED on your images," warned one commenter.
Moreover, users debate the blend of real and synthetic images, with one suggesting, "What would be great would be a training set preparation workflow that could take X number of genuine images and then generate a batch of synthetics." This reflects the community's frustration about how to effectively train consistent models.
Mixed Reactions from the Community
Responses reveal a mix of encouragement and skepticism. Some laud the progress, with one noting, "This method does work. I’ve used it a couple of times when the AI had an episode and spit out a distinct face." Others maintain a critical stance, suggesting dissatisfaction remains high.
Key perspectives include:
"I diversify my datasets with various hairstyles, clothing, lighting, backgrounds."
"Flak aside, this method does work."
Potential Directions Forward
As the community embraces AI-generated art, discussions about quality control are intensifying. Some members foresee growing community demand for standardized guidelines concerning dataset sizes, with initiatives likely pushing towards the production of more refined outputs.
Key Points to Consider:
🔼 Many argue 20 images are not sufficient for reliable results.
⬇️ Critics emphasize significant quality concerns in the generative process.
🤔 "The mix of real and synthetic in your training set is key."
As more people join the trend, the quest for improved quality in AI-generated character creation will likely intensify, driving innovation and setting new benchmarks for users in this rapidly evolving digital art landscape.