This LoRA was trained with AI Toolkit.
To use this AI Toolkit Krea 2 Edit LoRA in ComfyUI, install Ostris's custom node pack:
ComfyUI-Krea2-Ostris-Edit
https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit
This LoRA was trained with KV Cache enabled in AI Toolkit. When using it in ComfyUI, set kv_cache to ON in the Krea 2 Ostris Edit Model Patch node.
Using the example workflow with KV cache enabling
Prompt 1 – Natural photorealistic conversion
Convert the anime image into a natural photorealistic live-action photograph while preserving the exact subject count, age category, identity-defining features, pose, composition, clothing, objects, background, lighting, and colors.
```
Prompt 2 – High-end Cosplayer conversion
Convert the anime image into a premium high-end cosplay photograph of clearly adult cosplayers with polished beauty retouching, refined social-media filtering, cinematic color grading, realistic skin and fabric, while preserving the exact subject count, pose, composition, character features, costume design, objects, background, lighting, and colors.
H3 Anime-to-Live-Action LoRA (Experimental)
An experimental style-transfer LoRA built on H3, designed to transform anime-style video sequences into realistic live-action footage.
Important Usage Notes
Pair with Video Captioning (Recommended):
For best results, pair this LoRA with a video-captioning model (VLM) or provide a thorough description of every camera cut and shot transition in your prompt.
Multi-Shot Limitation:
If the prompt only covers a single frame or misses subsequent cuts, the LoRA may only apply the live-action conversion to the first shot (or the first two shots), failing to stylize later scenes.
You may need to use an audio refiner node to fix the audio
Tested Environments
Native H3 Base Model: Fully tested and verified.
H3 + Accelerated / Lightning LoRAs: Tested and verified; maintains style consistency even under fast, low-step setups.
Recommended Workflow
Captioning: Run your source anime clip through a video VLM to extract a full description covering scene transitions and camera dynamics.
Prompt Setup: Structure your prompt to account for every cut (e.g.,
Shot 1: ... -> Cut to Shot 2: ...)
For Qwen Image, you may use 3 three different prompt:
Natural photorealistic live-action style：
LumiReal, natural photorealistic live-action style.
Recreate the anime reference in a natural photorealistic live-action style, preserving the subjects, distinctive character features, pose, composition, clothing, objects, setting, lighting, and color palette.
Cosplay photography style：
LumiReal, polished high-end cosplay photography style.
Recreate the anime reference in a polished high-end cosplay photography style, with refined beauty retouching, cinematic color grading, realistic skin and fabric, while preserving the subjects, distinctive character features, pose, composition, clothing, objects, setting, and color palette.
Cinematic live-action style：
LumiReal, premium cinematic live-action style.
Recreate the anime reference in a premium cinematic live-action style, with attractive character casting, refined yet natural facial features, realistic skin texture, professional film lighting, believable environments, and scene-appropriate cinematic effects, while preserving the subjects, distinctive character features, pose, composition, clothing, objects, setting, and color palette.
Description
FAQ
Comments (14)
Interesting, gonna try this with Ostris' wf and see what happens.
I just tested it, and the results are good; adding SeedVR upscaling might yield even better results.
@MengY_Madrigal yeah, already tried it, it's quite good, thanks!
Hi, sorry, but once again, there’s no point in sharing your LORA if an image doesn’t contain any data for this model. Yes, there is a workflow, but it’s in JSON format; however, we want a workflow based on the image—it’s simpler that way :)
Add one image with workflow to the showcase. I couldn't do that earlier because the graphics card in my computer had broken, so I had to wait for the cloud platform to update its node plugins before I could run the tests.
go to - .\ComfyUI\custom_nodes\comfyui-krea2-ostris-edit\workflow\Krea2_Ostris_Edit.json
VERY GOOD! this is the best one. for anime to real.
Can I use it with draw things?
感谢大佬，很好用。一个小缺点就是眼睛好像都是looking at viewer的状态，闭眼或者背面或者看向旁边的原图也会看向镜头。
啊，这个确实没注意到，提示词加个眼睛状态看能不能修下
Looks pretty good at first glance, I'll test it.
Whenever I use this LORA I get really strange blotches/artifacts on my image. Check out the one I posted to your gallery. I'm using your exact workflow, so I'm not sure what's going on or how to fix it. Any thoughts?
did you update the node to the latest version? It looks like the KV Cache node's error, it should be opened.
@MengY_Madrigal Yo, that was it! Thanks for your help.
ComfyUI manager was showing that I was updated to the latest version of the Ostris node, but I wasn't. It all seems to be working nicely now. Great LORA btw! Thanks for your hard work on it. I'll try to post some images later to your gallery.