Ernie test

Since I can't do a full-fledged train yet, I created two Asian female LoRAs and an anime LoRA and merged them into the base model. It has been adjusted with the assumption that you will be using Distill LoRA, so be sure to use it. Since the CFG scale is assumed to be around 1.2 to 2.0, negative prompts can also be used. Testing is mainly done using CFG scale 1.5. Those who want faster speed should set it to 1.0 (negative prompts will no longer be available).

There are parts of the workflow that are a little strange, so I'll explain them. ernie-image-promptenhancer.safetensors used in the Prompt Enhancement part seems to be a fine-tuned version of mistral-3-3b.safetensors. In fact, they can be used interchangeably. Although it doesn't make much sense, I tried to make it possible to mix the two text encoders in any ratio. With a merger node, you can only specify a range from 0 to 1.0, but since I have a float primitive, you can specify a wider range by connecting the value part to it. (see also PromptEnhancer Selector )

If you want to turn off Pronmpt Enhancer, simply bypass the subgraph entirely.

There is a part inside Pronmpt Enhancer that specifies the number of vertical and horizontal pixels, but it is unclear what it does. It can be set independently of the actual image aspect ratio, and changing it will greatly change the composition and pattern. I can't understand what that means. I don't feel like I have any control over it.

*These output examples have the same prompts, seed and other parameters, but only the vertical and horizontal sizes of each internal section have been changed.