我想知道，有没有你推荐的触发词，保证得出给更好的写实风格，我常常用你的图片的参数进行设置，尤其是负面tag和正面tag里的修饰词进行设置，但总是会出现比较偏向于2d,3d,blender,之类的画风，不够真实。所以很需要一个准确万用的描述词，方便我准确出现真实的写实风格

Hi, thanks as always for your efforts and all the great model releases.

A few questions regarding these models, if you have a moment:

1. is it trained for v_pred (velocity) or eps (epsilon) prediction type?

2. is Noob v_pred version or eps version used as base / for merge?

3. is it trained with zero terminal signal-to-noise ratio (SNR)?

4. when you release LoRA for this base model, are any trained for v_pred and/or zero-SNR?

is the noob and IL a different version or is it kinda upgrade version?

Finally ! A model that can do both realistic and cartoonish NSFW. Can be improved on both ends (hard to have a very photographic or cartoonish result sometimes) but a good start for sure, I'm gonna try some Loras for that.

*IMPORTANT* From my testing (On ComfyUI) this model does not like high emphasis "()" at least in the negative prompts section, for example" (anime,anime screencap:1.4)" caused all my images to turn single color blue or green. Might just be a skill issue but wanted to point this out

Can this model use the controlnet of noob_eps?

This model is extremely perfect，have you considered the vpred version?

The noob-V1.5-VAE version is easily the best semi-realistic checkpoint I've used, by far, for anime/game/cartoon/etc characters. It's excellent at translating the characters built into Illustrious/Noob into a realistic style without eliminating their unique traits.

Just make sure to use a good amount of realism tags in the positive prompt and non-realism tags in the negative prompt (see below for the ones I tend to use), and you might be amazed at what this checkpoint can pull off, without the need for any additional LoRAs.

Grade: A+ and a gold star ⭐

Tags I use:

In the positive prompt:

masterpiece, best quality, newest, highres, absurdres, realistic, photorealistic, photo \(medium\), dslr, 4k, real life, photo background, depth of field, blurry background, bokeh,

In the negative prompt:

lowres, worst aesthetic, bad quality, worst quality, bad anatomy, jpeg artifacts, scan artifacts, lossy-lossless, unfinished, ugly, poorly drawn, greyscale, (illustration, 2d, 3d, painting \(medium\), toon \(style\), sketch, anime,:1.1) watermark, text, extra digits,

Typically, I've had success using Euler A, Normal scheduler, at CFG 4.00 and 30 steps in ComfyUI.

Also, assuming you're using this locally, feel free to try some of the common larger aspect ratios. This checkpoint can actually produce some amazing results even at those higher resolutions (including 1024x1280 / 1280x1024, 1152x1408 / 1408x1152, 1024x1360 / 1360x1024).