Hypernetworks Or Embeddings?

[NaughtDZ]

When training in the style of a certain painter, I found that the same basic parameters (prompt, number of sample pictures), only the training speed of Embedding is 0.0005, and the Hypernetworks is 0.0000005. With the same 25000 steps, the effect of embedding is much better than Hypernetworks. It seems that Hypernetworks cannot grasp the details of image style. Do i need to further optimize the prompt, or expand or reduce the sample image?

[aleksusklim]

How many images?

What vector/token size for embedding?

[NaughtDZ]

400 pictures，8 vectors per token for embedding.

[aleksusklim]

I'll try with 8 then. My first attempt was 200 images (converted to ~800 after preprocessing) and 1 vector. And looks like I need less images, best-picked manually (instead of dumping everything I got).

[OrenjiVR]

Am I missing something?
One learning rate is 1000 times higher than the other. Doesn't logic dictate that of course the higher learning rate learns faster (in exchange for it being more prone to errors)?

[NaughtDZ]

I think I know why I said this phenomenon. I looked at the text message on the training page and found that during training,It will use the prompt in your previous txt2img to generate the number of steps for each set log image. If the prompt of your txt2img is empty or has nothing relation with your training, the training log image will also generate something irrelevant.

[NaughtDZ]

Interestingly, there was a mutation in the training logs of 15000 and 15500 when I retraining it from 0 step. The two training sessions were held on October 16 and October 17, respectively. I did not change any parameters and data related to the training. Maybe the sdUI is updated or I accidentally adjusted something in the settings?The three pictures are 15000 logs and 15500 logs respectively, and the artist draws the original picture by hand

[Heathen]

Nice artist, reminds me a bit of Kentaro Miura.

[NaughtDZ]

I think I know why I said this phenomenon. I looked at the text message on the training page and found that during training,It will use the prompt in your previous txt2img to generate the number of steps for each set log image. If the prompt of your txt2img is empty or has nothing relation with your training, the training log image will also generate something irrelevant.

[1960sy]

Is there any regulation about the ratio between pictures number and the number of vector per token?