Not really—there’s no absolute “best.” It mostly comes down to personal preference. The model names just reflect that they’re all built on the same base model. I’m simply experimenting with merges to create a style that I personally enjoy.

@asdasdsad Thanks, I’m glad you like Sayaka. As I mentioned in my other reply, full-body shots are naturally harder to keep as detailed as close-ups, especially for skin texture. For now, I’d suggest avoiding distant full-body standing poses and using compositions like crouching, sitting, or curled-up poses instead. That can help work around the model’s current limitations as much as possible.

The samplers mentioned in the model description are just recommendations. Those options are available in Stable Diffusion WebUI Forge, and they’re simply the ones I personally prefer. That said, you’re totally free to try other samplers too—there’s always a chance they might work even better for you.

I’ve noticed that too when generating full-body images. Part of the reason is that diffusion models usually struggle with very small faces in wide or full-body compositions, so facial details become less stable. Also, the Illustrious base is more focused on illustration and character knowledge than specifically fixing long-distance face quality. Since this model series is merged from multiple checkpoints and LoRAs, biases from the source models—especially toward close-up shots—can become even more noticeable. So if the original models were not trained or tuned for strong full-body detail, it is difficult to fully solve for now. I’ll keep experimenting and see if I can find a better way to improve it.

I’ve also tried using img2img inpainting to refine those details, and it works quite well. It does take a bit more time, but the results are fairly good, so that might be worth trying as well.

Thank you, I’m glad you like it. The real credit should go to the creators behind these excellent base models. I’m mostly just experimenting with different merges and style combinations. I’ll continue testing different directions and see what works well for a semi-real anime style.

I haven’t personally used WeiLin, so I’m not fully sure about its exact node behavior.

But if the WeiLin prompt node accepts a normal text input, I would suggest placing Impact Pack’s ImpactWildcardProcessor before it:

ImpactWildcardProcessor → WeiLin Prompt Editor → positive prompt / conditioning

Write your wildcard syntax, such as __your_wildcard_name__, in ImpactWildcardProcessor, then pass the processed text output to the next text / positive prompt input.

The idea is to expand the wildcard first, then send the plain text downstream. Also make sure your wildcard files are in Impact Pack’s wildcard folder, for example:

ComfyUI/custom_nodes/ComfyUI-Impact-Pack/custom_wildcards

If the wildcard file was just added or edited, try restarting or refreshing ComfyUI.