Author: Hongke Because most LoRAs on Liblib are trained on FLUX FP8, and most Checkpoints on the market are basically mergers of the official model + LoRA.

Suppose the official model is A, the merged LoRA is x, and the LoRA you loaded is B, then the output result equals A+x+B. This leads to the situation shown in Example 3 of the image: because the training set of the "influencer" base model consists of a bunch of attractive women, it inevitably biases the model weights toward that typical influencer face. Consequently, even if the prompt specifies a little girl, the output will still feature an influencer face. Unless you train on top of the influencer model, which I won't do.

Xingmang, while basically leaving the face shape and composition unchanged, slightly enhances the skin texture and details of the generated images. That is pretty much the purpose of this model.

Prompt Suggestions:

Tips: There is no fixed template. Just write the scene you want in natural language. It is recommended to send a reference image to ChatGPT to get a descriptive prompt and create based on that—it works much better than writing it from scratch.

Recommended Online WebUI Parameters: Simply keep the defaults for the most part.

Sampling Method: Euler

CFG: 3.5

Steps: 30

Resolution: Width 1024 x Height 1024, ideally kept as multiples of 64, such as: 1024x1536, 1152x768, 1344x896.

Keep the resolution around 1024. When the character occupies a small proportion of the frame, structural errors may occur. You can enable Hires. fix with a denoising strength of 0.3–0.6 and a 2x upscale. The higher the denoising strength, the more noticeable the changes in the image will be.

Recommended ComfyUI Workflow:

https://www.liblib.art/modelinfo/5171f6fa89304df5b45e87e3118ec3b3