tori29umai/krea2-controlnet

Krea2 ControlNet Collection

Language: English | 日本語

This repository is a collection of Control LoRA models for Krea-2. Each model has its own control behavior, training data, and training configuration. More models, including lineart controls, may be added over time.

Models

Model	Control behavior	Status
krea2-anythng	Loosely follows the overall silhouette, pose, and broad composition	Available
krea2-lineart	Lineart-guided generation	Planned

krea2-anythng

krea2-anythng is the model that loosely follows the input image's overall silhouette and broad composition. It leaves room for the prompt and Krea-2 to determine appearance and detail. It is not an exact edge-tracing or pixel-perfect structure model.

File

• krea2-anythng_step_007000.safetensors
• Base: krea/Krea-2-Raw
• Rank: 64
• SHA-256: 264011d01fa4821959547689a44f2b261a5beaf384d177c60c605caf3371b4b6

The file contains the Control LoRA weights and expanded input projection only. A compatible Krea-2 base model is required.

Inference with ComfyUI

Use comfyui-krea2-controlnetPlus.

Input and output example

Input control image	Output image

Control range setting example

These example images are provided by the comfyui-krea2-controlnetPlus repository.

1. Install the custom node according to its repository.
2. Put the safetensors file in ComfyUI's models/loras directory.
3. Prepare an image whose broad silhouette/composition should guide generation.
4. Encode it with Krea2 Control Plus Image Encode. For krea2-anythng, start with channel_mode=rgb, normalize=none, invert=false, and match_latent_size.
5. Load the model with Krea2 Control Plus LoRA Loader. Start with start_percent=0.0 and end_percent=1.0.
6. Attach the control latent with the required Krea2 Control Plus Apply node.
7. Send the resulting model to the normal Krea-2 sampler and use your text prompt.

The custom node repository provides example_workflows/Krea2Controlnet.json.

krea2-anythng training reference

This configuration applies specifically to krea2-anythng; future models in this repository may use different datasets and settings.

The model was first trained for 6,000 steps at a learning rate of 1e-4. Training was then resumed from the step-6,000 weights with a lower learning rate of 3e-5 and continued to step 7,000. The published file is the final step-7,000 checkpoint.

• Dataset: 106,786 aligned silhouette-control / RGB-target pairs
• Captions: per-image natural-language English captions (*_nl.txt)
• Preprocessing: matched approximately 1-megapixel aspect buckets, center crop, Qwen-Image VAE encoding

base_model: krea/Krea-2-Raw
control_type: silhouette
rank: 64
batch_size: 1
gradient_accumulation: 8
caption_dropout: 0.1

initial_training:
  steps: 6000
  learning_rate: 0.0001
  warmup_steps: 200

extended_training:
  resume_from_step: 6000
  final_step: 7000
  learning_rate: 0.00003
  warmup_steps: 50

precision: bfloat16
gradient_checkpointing: true
optimizer: AdamW
adam_betas: [0.9, 0.99]
weight_decay: 0.0

The 13B base was frozen. Training updated rank-64 LoRA weights across the 28 DiT blocks and the expanded input projection for the VAE-encoded control latent.

krea2-anythng limitations

• Intended for rough whole-image silhouette, pose, and layout guidance.
• Fine contours, faces, small objects, and exact spatial relationships may differ.
• Prompt, seed, sampler, and control start/end percentages affect adherence.
• Outputs inherit Krea-2's behavior and limitations.
• Experimental; evaluate outputs before production use.

License and attribution

Use each checkpoint in accordance with the Krea-2 base model and inference software licenses and terms. Thanks to the Krea-2 ControlNet training implementation and ComfyUI Krea2 Control Plus integration.