Flux Lora Training Using Diffusion-Pipe - Help Wanted
The user is seeking help with training a Flux Lora model using the Diffusion-Pipe tool. They have been using the tool for a while but are facing challenges with the training process, particularly with a Flux Lora for an AI girl character.
Why it matters
Improving the training process for Flux Lora models can benefit the Stable Diffusion community, as Loras are an important component in customizing and enhancing the capabilities of the Stable Diffusion model.
Key Points
- 1The user has been using Diffusion-Pipe to train Loras for various projects
- 2They are currently training a Flux Lora for an AI girl character, but the training is taking a long time
- 3The train/epoch loss is still above 0.22, although it is decreasing
- 4The user is looking for help with the Flux.toml configuration to improve the training process
Details
The user has been using the Diffusion-Pipe tool to train Loras for various projects, including Hunyuan, Wan, Z-Image, SDXL, and Flux. They have created a lot of Loras using this tool. However, when they went back to retrain a Flux Lora for one of their AI girl characters, the training process is taking a very long time, up to 30 hours, and the train/epoch loss is still above 0.22, although it is decreasing. The user is running the training on an RTX4090 GPU, using only 15GB of VRAM and no spillover to DRAM. They have provided their training settings, which include 1200 epochs, a micro-batch size of 4, and various other configurations. The user is seeking help from the community to share their Flux.toml content and provide any feedback on their current settings, as they believe there might be some inefficiencies in their setup.
No comments yet
Be the first to comment