Depth ControlNet LoRA for Krea 2 — better structure, layout, and depth-guided generation without needing a full ControlNet.
Krea 2 Depth ControlNet LoRA adds depth-based structural control to Krea 2, helping guide composition, object placement, pose, and scene layout while keeping the model’s native image quality and style flexibility.
It is useful for workflows where you want stronger control over generation using depth maps, especially in ComfyUI or custom inference pipelines.
Trained for Krea 2 workflows and designed to be lightweight compared to a full ControlNet.
Check it out here :
Huggingface : https://huggingface.co/Patil/Krea-2-depth-controlnet
Github : https://github.com/Tanmaypatil123/Krea-2-controlnet
HF space : https://huggingface.co/spaces/hugging-apps/krea-2-depth-controlnet
Comfy ui details : https://github.com/facok/comfyui-krea2-controlnet
For more details on comfy ui on lower vram gpus check this : https://www.reddit.com/r/StableDiffusion/comments/1ummd2h/krea_2_depth_controlnet_lora/
For feedback, collaborations, or custom model work, reach out to me on X/Twitter: https://x.com/TanmayPatil79
Description
First release
FAQ
Comments (17)
So, we need 48GB VRAM to run this?
Naah not need .. You can use it on consumer gpus .
You will find details over here https://www.reddit.com/r/StableDiffusion/comments/1ummd2h/krea_2_depth_controlnet_lora/
48 gb requirement is for full precision only .
@Tanny007 Oh got it, thanks.... last noob question: does this go in Models/loras, or models/controlnet ?
@ArtiOtaku damn I am confused now .. This is not controlnet architecture basically like you see in sd15 models .. it's more like lora like controlnet you can consider .
@ArtiOtaku put it in your lora folder. I get ~80s generations on an 8gb 3070. if you download the 3 panel picture on the github page, it has the comfyui workflow
https://github.com/facok/comfyui-krea2-controlnet
@wondier got it, thanks for your assistance.
I tried using this with the Turbo LoRA, but the results are overcooked... Does this only work with the RAW model?
No this does work with turbo also .
I just tried it with RAW/Base (BF16, 52 steps, CFG 3.5) - worked like a charm. Same workflow with the Turbo model (BF16, 8 steps, CFG 1.0) > overcooked and not adhering to the depth mask. I thought the culprit was the Turbo LoRA, but it was not. Also deactivated all other LoRAs, but the result wasn't changed.
I will try it again with a more simpler workflow.
@redhome Yeah i know there might be difference as the checkpoint was trained with raw model not with turbo but the lora does work with turbo but not like how it works with raw .
@Tanny007 I tried a few things - also using the provided workflow, but it seems the Turbo models doesn't behave the way RAW does.
@redhome Yeah I know .. will try to do something about this in next training run :) but for now thanks for feedback and trying the checkpoint.
Hey bro, is there any workflow to use it ? and the depth map made by this LORA ?
https://www.reddit.com/r/StableDiffusion/comments/1ummd2h/krea_2_depth_controlnet_lora/
you find details in this subreddit or you can also check this :
Amazing one! It works well. Thx for your hard work！
Thank you
Thanks for the LoRA, it works better in RAW, even with RAW + Distill LoRA, but it impacts the quality of characters' LoRAs resemblance. Decreasing the LoRA strength doesn't help.