To Train in CIVITAI

1. Click CREATE->Train a Lora-> Writ the name of your new LoRa and click the type, e g VIDEO/ACTION if you want to train for minimax H3.

2. Click some boxes.

3. Upload your images/Dataset.

4. Copy in your caption on each clip. Auto-Caption did not work for me.

5. Click Next.

6. You get to choose for what base to train, e g, Minimax H3.

7. Pay the buzz. Currently, the cost is 2750 buzz. It may go up later.

8. Next it will train, and the next day there will be a selection LoRa to choose from.

Example DataSet provided:

I made a LoRa here H3_BJ_Gesture_V1 - V1 | MiniMax H3 LoRA | Civitai

but it lacks a tongue-in-cheek bulge from the other side of the hand, so I tried to make datasets to train those. I failed. inside the LoRa post, I have now included the two FAILED datasets in the link above. I cannot see what is wrong with them. Perhaps too many, or wrong clips somehow? Perhaps trhe wrong Caption?

For Caption, I used the same caption for every clip. I only want to train one movement-combo, and this has worked for me in other LoRas.

If you want to make your dataset

1. The best dataset may be one that is AI-generated.

2. If you want to pick from the internet, that's probably fine. For clips, 4 seconds seems to work well. However, you may not be able to share the dataset if you suspect it might have copyright issues, and the LoRa must avoid recognizable real characters, dead or alive.

3. You should write in the description something about where they came from.

You should write some suggested prompt(s) and strength in the description. On my first LoRa, I did not, and CIVITAI blocked the LoRa.

CAPTION OR TAG?

Choice of Caption and Tag depends on the purpose and model of the LoRa. The training in CIVITAI will suggest if you get it wrong. Auto-tag or manual? Also depends. Auto-tag worked on my images for style and character LoRA.

I am still a noob when it comes to how to write Captions or tags, so I am going to have to ask you to Google it, or possibly ask an LLM.