Seeking Clean AIO Models Without Embedded Lightning LoRA And Other Mix version
#33
by
ioritree
- opened
Many WAN models now come with Lightning LoRA already merged into the base model. While this improves inference speed and simplifies deployment, it introduces serious drawbacks:
- LoRA contamination is too severe: The merged LoRA permanently alters the base model’s behavior, often overriding its original style and flexibility.
- No control over LoRA influence: Since the LoRA is baked into the model, users can’t adjust its strength or disable it. This removes the ability to fine-tune outputs based on prompt or context.
- Inconsistent results: The merged LoRA can cause unpredictable outputs, especially when prompts unintentionally trigger specific styles or features.
- Loss of modularity: Users lose the freedom to experiment with different LoRA combinations or swap them dynamically.
I’m looking for a clean version of WAN models—no embedded Lightning LoRA, no baked-in accelerators—just the original base model with full flexibility to load LoRA manually when needed.
My models are designed for speed and simplicity, which is why they include everything people need to get started quickly. Everyone is free to pick and use whatever models they want individually elsewhere. "Clean AIO" is a bit of an oxymoron in this case, imho.
Phr00t
changed discussion status to
closed