shaorax

AI Visual Designer | Turning imagination, symbolism, and emotion into visual worlds.

[shaorax @shaoraxRAC]

KREA2 being open source is almost unfair.

Straight 4K output, structured image generation, and the image quality is ridiculously good.

At this point, open source is no longer knocking on the door — it’s kicking it down.

[shaorax @shaoraxRAC]

@94vanAI 😅你怎么知道我也做了这样的事情 https://t.co/GDZMxk5ye3

[shaorax @shaoraxRAC]

@PhotogenicWeekE @seki_images lora强度我控制在0.75，提示词中加入Reversal film ,做为提示词，胶片质感���颗粒感很真实。 可能唯一不足是这个LORA 应该用了小分辨率512，768，���有训练1024 以上的分辨率。我发现尺寸超过一定像素范围以后生成效果会差很多，包括 奇怪的肢体，重复人像。 控制在一百万像素就没问题。

[shaorax @shaoraxRAC]

@PhotogenicWeekE @seki_images lora强度我控制在0.75，提示词中加入Reversal film ,做为提示词，胶片质感和颗粒感很真实。 可能唯一不足是这个LORA 应该用了小分辨率512，768，没有训练1024 以上的分辨率。我发现尺寸超过一定像素范围以后生成效果会差很多，包括 奇怪的肢体，重复人像。 控制在一百万像素就没问题。

[shaorax @shaoraxRAC]

@nanyuan0412 效果不错

[shaorax @shaoraxRAC]

@nanyuan0412 可怕，用krea2 我抽了10几张才能出类似的角度。还得一直强调提示词。

[shaorax @shaoraxRAC]

@BeautyVerse_Lab 优雅端庄，好，很好！

[shaorax @shaoraxRAC]

@jzaib4269 yes,it is

[shaorax @shaoraxRAC]

@underwoodxie96 自己的数字人

[shaorax @shaoraxRAC]

More references ≠ better results.

Less is more. The real trick is good images and proper tagging — not feeding your GPU a pile of chaos and hoping it becomes art.

[shaorax @shaoraxRAC]

Krea2 LoRA test: around 60 images, 2,000 steps, trained with AI-Toolkit.

For Natasha, I used image2 to generate multiple angles, then used a reference-image node to keep the character consistent with much less manual fixing.

Now just waiting for Krea2 + ControlNet. That combo could be scary good.

[shaorax @shaoraxRAC]

@icreatelife @X Hi, I’m working on creating a character I love and training a LoRA for her. I’m also really into art and design.

[shaorax @shaoraxRAC]

@PhotogenicWeekE 感谢！好东西呀，节点参数调大了会影响画质，这���可以绕过限制

[shaorax @shaoraxRAC]

@gosrum gpt-image-2，正常泳装三视图，已经描写泳装概念都难以正常出图呢。在我看来krea2的美学性能在无法出图的情况补足。当然是对于特殊内容了。审查才是开源模型和闭源模型的区别，又或者叫监管才是。

[shaorax @shaoraxRAC]

@underwoodxie96 好，krea2，复现成功

[shaorax @shaoraxRAC]

Today’s recommended node is an extended version of yesterday’s workflow.（Reference views: left side, right side, and a front low-angle shot.）

When you want to generate a similar image and adjust the angle, sure, you can try describing it with prompts. But with this node, you can adjust what you see in a much more visual and intuitive way.

Of course, since KREA2 doesn’t actually have true image-editing capability, this is still mainly useful for controlling angles during image generation — but for that, it’s honestly pretty handy.

Check the reference node graph, and if you have any questions, feel free to leave a comment.

[shaorax @shaoraxRAC]

@PhotogenicWeekE 非常不错的效果，请问使用了多少张进行训练呢？

[shaorax @shaoraxRAC]

@underwoodxie96 不得不说这个角度有点有趣

[shaorax @shaoraxRAC]

@AI_GIRL_DESIGN 美少女与黑猫？哦，是黑猫本人

[shaorax @shaoraxRAC]

Connect the image to the node, then feed it into the positive conditioning. No prompt needed, no negative prompt either — it can generate a similar image directly from the picture.

[shaorax @shaoraxRAC]

Want to get an equally gorgeous image with Krea2?
You just need to add one text box — with image-to-image, you can generate a nearly identical picture with ease.

Single Event Details