Best Realism SOTA Models for LORA Training | Users Weigh In

A lively discussion ignites among users regarding the best models for realism in LORA training after the recent release of Krea, Wan 2.2, and Qwen. With mixed opinions and varying experiences, users are eager to share best practices and recommendations.

Context and Significance

In the world of AI and image modeling, realism has become a hot topic following the introduction of new models. The excitement is palpable, yet confusion lingers over which model delivers the best results for trained character LORAs.

Key Insights from User Experiences

Here are the primary themes driving users' discussions about the models:

1. Model Performance Comparisons

• Krea: Users note it shows significant improvement but complains about color inaccuracies like yellow tints or washed-out images. “Can be fixed in post,” one user mentioned, indicating a workaround.

• Wan 2.2: Praised for its performance, yet users report needing to create multiple LORAs to achieve finer results.

• Qwen: Users are impressed with adherence above CFG 1, but it appears to compromise skin tones and aesthetics. “Great adherence, but at what cost?” questioned a participant.

1. Workflow Optimization

Users are exploring workflows between models, notably with Qwen leading into Wan 2.2's low noise option. The results have been decent, but questions remain about optimization.

1. Best Practices Shared

Several users have sought advice on best practices for employing these models for realism. They inquire about effective CFG settings and passing ratios between models. One commenter noted, "I wish to see a gallery of Qwen to Wan images for comparison.”

“This is a long-winded way of asking if anyone found a workflow that works well,” a user humorously stated, summarizing the community’s desire for clarity.

Key Takeaways

• 🔹 Krea shows potential but struggles with color accuracy.

• 🔹 Wan 2.2 may require multiple LORAs for optimal outcomes.

• 🔹 Users report Qwen’s adherence can affect realism in skin tones.

As discussions continue to stir excitement in this niche space, could user-shared workflows become the key to unlocking higher levels of realism in character creation?

Keep your eyes peeled for more user insights, as this story is still developing.

What Lies Ahead for LORA Models

There's a strong chance that the discussions surrounding Krea, Wan 2.2, and Qwen will lead to improved model updates in the near future. With user feedback driving innovation and developers eager to enhance realism, we might see models refine their color accuracy and tonal balance. Experts estimate around 70% likelihood that workflows combining these tools could become standardized practices among creators, paving the way for more effective character LORA generation. As the community continues to share tips and strategies, the collaboration could accelerate the pace of advancements in this niche field.

Historical Echoes in Innovation

Reflecting on the evolution of photography, the early days saw inventors wrestling with color reproduction much like today’s users confront realism in LORA models. Just as George Eastman’s Kodak revolutionized access to photography while prompting improvements in chemical processes, current models will likely evolve through user creativity and community collaboration. The pursuit of realism, once painstaking in traditional media, now finds resonance in digital tools, suggesting that progress often shares a familiar rhythm of trial, error, and eventual breakthrough.