Token and Class words are what guide the AI to produce images similar to the trained style/object/character.
Include any mix of these words in the prompt to produce verying results, or exclude them to have a less pronounced effect.
There is usually at least a slight stylistic effect even without the words, but it is recommended to include at least one.
Adding token word/phrase class word/phrase at the start of the prompt in that order produces results most similar to the trained concept, but they can be included elsewhere as well. Some models produce better results when not including all token/class words.
3k models are are more flexible, while 5k models produce images closer to the trained concept.
I recommend 2k/3k models for normal use, and 5k/6k models for model merging and use without token/class words.
However it can be also very prompt specific. I highly recommend self-experimentation.
These models are subject to the same legal concerns as their base models.
Comparison
Epoch 5 version was earlier in the waifu diffusion 1.3 training process, so it is easier to produce more varied, non anime, results.
