Phr00t

/

Qwen-Image-Edit-Rapid-AIO

like 1.92k

I can't really help without the workflow you used.

v21 does include more realism LORAs to improve skin detail, but I was still able to get plenty of anime and cartoon results when prompting for those styles.

Could you post a screenshot of the workflow？

Of course, here is the workflow I used. Version 21 completely ruined the anime style.

How’s i5 look without the lighting 8steps Lora? I think it’s already part of the merge, no?

i believe its because you have terms "anime" & "digital art" in the negative prompt. check those over, remove and try again

I'm check your workflow, it's "Anime" and "digital art"in your negative prompt. try remove it.

After my test, there is really no way to do anime and comics. Let's say use a real photo and let him turn it in. Ghibli style or comic style, he will fail to transform. Or partial comics, others are still mainly realistic

Of course, here is the workflow I used. Version 21 completely ruined the anime style.

There is so much going wrong with this workflow.

First, you are adding another (incorrect) accelerator LORA. Do not use any additional accelerators, they are already baked into the "all in one".

I have no idea why you are using the "next scene" LORA when you are just trying to remove some clothes?

The "ModelSampling" and "CFGNorm" stuff only reduced quality in my experience. I recommend removing those.

You literally put "anime" in your negative prompt. However, this should theoretically be ignored because you have a CFG of 1.

I haven't tested using the "Clownshark sampler" and would recommend simply using the basic KSampler with euler_ancestral/beta.

Finally, another big problem: you are not even prompting for an anime style.

I'm surprised you got even anything close to working with this setup before.

Of course, here is the workflow I used. Version 21 completely ruined the anime style.

There is so much going wrong with this workflow.

First, you are adding another (incorrect) accelerator LORA. Do not use any additional accelerators, they are already baked into the "all in one".

I have no idea why you are using the "next scene" LORA when you are just trying to remove some clothes?

The "ModelSampling" and "CFGNorm" stuff only reduced quality in my experience. I recommend removing those.

You literally put "anime" in your negative prompt. However, this should theoretically be ignored because you have a CFG of 1.

I haven't tested using the "Clownshark sampler" and would recommend simply using the basic KSampler with euler_ancestral/beta.

Finally, another big problem: you are not even prompting for an anime style.

I'm surprised you got even anything close to working with this setup before.

我在任何工作流或者是官方工作流都尝试将真实照片转换成动漫或者其他卡通类的风格。它的失败率占据在80%以上，有可能人物转换成功，可背景还是写实。或者服装还是原来的，脸变成卡通。

I'm check your workflow, it's "Anime" and "digital art"in your negative prompt. try remove it.

When cfg = 1, the negative prompt is ignored.

Of course, here is the workflow I used. Version 21 completely ruined the anime style.

There is so much going wrong with this workflow.

First, you are adding another (incorrect) accelerator LORA. Do not use any additional accelerators, they are already baked into the "all in one".

I have no idea why you are using the "next scene" LORA when you are just trying to remove some clothes?

The "ModelSampling" and "CFGNorm" stuff only reduced quality in my experience. I recommend removing those.

You literally put "anime" in your negative prompt. However, this should theoretically be ignored because you have a CFG of 1.

I haven't tested using the "Clownshark sampler" and would recommend simply using the basic KSampler with euler_ancestral/beta.

Finally, another big problem: you are not even prompting for an anime style.

I'm surprised you got even anything close to working with this setup before.

I implemented all the recommendations in my workflow, but V 21 still tries to incorporate a bit of realism.

v19

v21

Here are some more examples of how the model performs terribly in anime style. These were created using the workfow I attached above.

v19

v21

Of course, here is the workflow I used. Version 21 completely ruined the anime style.

There is so much going wrong with this workflow.

First, you are adding another (incorrect) accelerator LORA. Do not use any additional accelerators, they are already baked into the "all in one".

I have no idea why you are using the "next scene" LORA when you are just trying to remove some clothes?

The "ModelSampling" and "CFGNorm" stuff only reduced quality in my experience. I recommend removing those.

You literally put "anime" in your negative prompt. However, this should theoretically be ignored because you have a CFG of 1.

I haven't tested using the "Clownshark sampler" and would recommend simply using the basic KSampler with euler_ancestral/beta.

Finally, another big problem: you are not even prompting for an anime style.

I'm surprised you got even anything close to working with this setup before.

I implemented all the recommendations in my workflow, but V 21 still tries to incorporate a bit of realism.

v19

v21

While you did find some improvements, you still are not prompting for an anime style.

In version 19, without specifying an anime style in the prompt, the generation was adjusted to the style of the drawing.

In version 19, without specifying an anime style in the prompt, the generation was adjusted to the style of the drawing.

Then use v19 or specify what style you want. I will never guarantee that the exact prompts will always work better with different merges.

Even if i prompt version 21 to have an anime style, it still makes the drawing look realistic.

v20 also doesn't have this problem. The anime style is very precise and high-quality.

Try v22.

Try v22.
Thanks, I ran tests today with version 22. It generates anime perfectly, thanks.