Krea2_Anything2RealCharacters (V3 Release)
Krea2_Anything2RealCharacters-V3 is a high-resolution Image-to-Image / stylization LoRA model fine-tuned on top of Krea2_Turbo. It is specifically optimized to translate any 2D, anime, or artistic input into a high-fidelity, photorealistic style with fine skin micro-textures and precise facial details.
💡 What's New in V3
Complete Realism Conversion: Fixed the incomplete realism translation in V2.5, delivering significantly more authentic and natural photorealistic results.
Enhanced Character Consistency: Achieves vastly superior character feature retention and identity consistency compared to the original V1 release.
Released as LoRA for Maximum Flexibility: Unlike V2 and V2.5, V3 is released as a LoRA. This provides much stronger compatibility when stacking external Krea2 style or character LoRAs, opening up endless possibilities for custom aesthetics.
Special thanks to our community group members for their valuable test feedback! Unless critical issues arise, V3 will serve as the final release for the foreseeable future. For commercial inquiries, feel free to reach out. 🙏
🚀 Workflows & Online Demo
Try Online Demo: https://www.runninghub.ai/post/2091224831916716033/?inviteCode=rh-v1559
Download ComfyUI Workflows: Krea2_Anything2RealCharacters-V3-Workflows - v1.0 | Krea 2 Workflows | Civitai
✨ Key Features
Native 2048px Resolution Training: Trained on a full dataset at 2048×2048 resolution, allowing high-resolution direct outputs without additional upscalers.
Micro-Texture Detail: High-precision rendering of skin pores, hair strands, and realistic lighting.
LoRA Stacking & Compatibility: Fully compatible with external Krea2 Text-to-Image (T2I) LoRAs and face/character LoRAs for adjusting photography styles, lighting, and facial genetics.
Embedded Style & Feature Triggers: Fine-tuned support for custom film filters, lighting setups, and specific feature keywords.
Dual-Image Identity Transfer (Experimental): Supports transferring real-world identities from a reference image onto anime/stylized poses.
⚙️ Recommended Inference Settings
ParameterRecommended ValueDescriptionBase ModelKrea2 TurboCore architectureLoRA Weight0.8 – 1.0Adjust based on style intensitySampler Steps8 – 10Matches the Krea2 Turbo architectureCFG Scale1.0Prevents over-saturation and artifactsTarget Resolution2048 × 2048Native 2K outputCLIP / VAEStandard ComfyUI ModelsDownload official weights from ComfyUI Repository
🔑 Prompting & Trigger Words
Primary Trigger
Include the main trigger phrase at the start of your positive prompt:
transform the image to realistic photograph
Auxiliary Triggers & Filters
Mix and match these trained keywords to modify style, lighting, or body features:
Demographics:
caucasian,woman,man,muscularGaze & Angles:
looking at camera,look down,unfocused faceLighting & Aesthetics:
backlight,Film filter,hazy,photorealistic,highly detailedAttire & Features:
black stockings,white stockingsAnatomical Details:
pubic hair,labia
🧪 Dual-Image Persona Conditioning (Experimental)
For transferring a real-world face onto a 2D/anime pose:
Figure 1: Base anime image (determines pose and structure).
Figure 2: Reference person image (determines real-world identity).
Dual-Image Trigger Phrase:
with the character looking exactly like the one in Figure 2
⚠️ Note: Dual-image feature transfer is experimental and results may vary depending on the seed and input alignment.
Description
FAQ
Comments (23)
Lora works great! hope for a V2 that doesn't just generate asian
Does this LoRA have two functions? One is anime-to-real-life conversion, and the other is clothing removal?
That's right, but removing the effect requires a glitch, which I will solve in the V2 version.
Excellent lora, thank you!
If anyone is getting glitchy output, make sure your Ostris Edit nodes are up to date:
ComfyUI-Krea2-Ostris-Edit
And switch the "Krea 2 Ostris Edit Model Patch" node to True
Thank you, we are currently completing more concepts and training at 2048 resolution. With the support of Krea2, it can be said to be a god level Lora.
i'm guessing we have to use this with the edit lora right?
No need to edit Lora. I have provided a workflow.
The online link has workflow downloads, and the configuration file also has workflows
Great but too much asian vibe and it's make some mistakes in outfit and pose. klein9b make more identical image.
V2 will solve the problems of Asian faces and incorrect clothing. The main reason for the incorrect clothing is that the resolution setting is too high. The prior version is only trained at a resolution of 1280, and consistency issues may occur when the resolution is higher than 1280.
@AIGC_Singularity Ok. Wish you luck with that! Thanks for sharing!
@AIGC_Singularity needs more Asian variety, especially Korean cosplayers/ cosers. Some popular cosplayers found on twitter come to mind.
@kitekholin V2 will add facial reference function, but this is not strong consistency and can only guide facial features appropriately.
It works pretty well as far as realism goes but it makes colors very dark and muted. A person with vibrant auburn hair looks like they have dark brown hair afterwards.
You can add a color adjustment node when outputting the original image to reduce saturation to 20 or more.
这个lora好强。用了你之前的qwen edit anime to real的lora，效果一般。但是这个krea2的lora，速度比Qwen edit快，质量远高于qwen，效果非常好，不用二次重绘。
This is great!
Is it possible to make a similar lora for Illustrious or Pony?
渲染效果偏向细腻唯美，不过有个很大的问题是对比原图来说，会明显改变服饰设计特征（对比其它A2Rlora来说）
I'm confused by remove clothes, the two keywords triggers are a bit different from each other, they should be used in stack (both)? Or one or the other. Also remove clothes is also overlapping in meaning with undressing, what happens when used? Or it's supposed to remove clothes with this lora. Please explain.
這個LoRA很實用。但能不能和其他LoRA混合使用?另外，可不可以順便修圖? 比如改變瞳孔顏色?
Prepare V3 version, which should be able to load Lora for use at this time
nope. this Lora does not work, nor does the workflow you provided. tried several different images and different parameters, never got the expected results. It also makes everyone asian.
The workflow link should have been added above. What faces are needed to load the corresponding style of text and image Lora to solve this problem? Currently, the V3 version is being prepared, and the effect should be better.