𝗔𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗻𝗴 𝗟𝗼𝗥𝗔 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗛𝗲𝗿𝗺𝗲𝘀 𝗔𝗴𝗲𝗻𝘁

𝗔𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗻𝗴 𝗟𝗼𝗥𝗔 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗛𝗲𝗿𝗺𝗲𝘀 𝗔𝗴𝗲𝗻𝘁

Creating datasets for LoRA training is boring. You find images. You check licenses. You write captions. I used Hermes Agent to automate this for FLUX.2 [klein].

I needed a style where the base model is weak. I chose medieval marginalia. These are weird creatures from old manuscripts.

Hermes did the heavy lifting:

• Found 199 images on Wikimedia Commons.
• Filtered for licenses and resolution.
• Removed duplicates and noise.
• Picked the best 30 images for the set.

Good captions describe the subject. They do not describe the style. The model learns style from pixels. Hermes built a pipeline to write these.

I used ai-toolkit on a RunPod RTX 4090. I trained on the base model. The best results appeared at step 1000. The loss curve kept dropping, but the images got muddy after step 1000.

Tips for your next LoRA:

• Use 20 to 40 images for style.
• Describe objects in captions, not the look.
• Use a trigger word not found in English.
• Trust your eyes, not the loss curve.

If an agent handles sourcing and captioning, let it. It saves time.

Source: https://dev.to/stephen_btl/training-a-lora-on-flux2-klein-with-hermes-agent-2k05

Optional learning community: https://t.me/GyaanSetuAi