Chinese documentation: README_zh.md

Native-style ComfyUI nodes for Krea2 Control LoRA inference. The plugin keeps ComfyUI's built-in Krea2 inference path intact, uses the expanded input projection from the LoRA checkpoint during sampling, and injects a VAE-encoded control latent.

• Krea2 Control LoRA Loader: loads a Krea2 Control LoRA from models/loras, applies compatible block LoRA weights to the Krea2 model, prepares the expanded input projection, and registers the sampling wrapper.
• Krea2 Control Image Encode: encodes any control IMAGE with the supplied Krea2/Qwen VAE. It can consume outputs from comfyui_controlnet_aux preprocessors such as Depth Anything, Canny, OpenPose, lineart, and normal maps, but it does not import or call comfyui_controlnet_aux.
• Krea2 Control Apply: converts the encoded control latent into the Krea2 latent space and attaches it to the model after the Control LoRA has been loaded.

1. Prepare a control image with normal ComfyUI nodes, or connect a comfyui_controlnet_aux preprocessor output.
2. Encode that control image with Krea2 Control Image Encode using the Krea2/Qwen image VAE. Keep the default match_latent_size and connect the sampler latent to this node's latent input.
3. Load the Krea2 Control LoRA with Krea2 Control LoRA Loader.
4. Attach the encoded latent with Krea2 Control Apply.
5. Send the resulting model to your sampler.

Krea2 Control Apply is required after the loader. If the Control LoRA is loaded without an attached control latent, sampling fails instead of silently running a partially patched model.

Block LoRA weights are applied through ComfyUI's ModelPatcher so normal model loading, offload, and low-VRAM behavior still apply. During the Krea2 diffusion forward call, image tokens still pass through the native first projection so regular LoRA patches on the base model remain active; the Control LoRA contributes only the control-token half of the expanded projection. The temporary projection state is restored immediately afterwards, so removing the node does not leave the base Krea2 path patched.

LoRA block matching reads the live module weight shapes instead of relying only on state_dict() shapes, which improves compatibility with quantized/GGUF UNET loaders whose state dict may expose storage-shaped tensors.

match_latent_size is the default because the reference flow resizes the control image to the final generation size before VAE encoding. Switch to keep_control_image_size only if you already resized and cropped the control image elsewhere.

Krea2 Control Image Encode is generic and does not run depth, canny, pose, or other preprocessors. Its image options are lightweight tensor operations:

• channel_mode: keep RGB controls as-is, or convert to grayscale and repeat back to RGB before VAE encoding.
• normalize: per_image_minmax normalizes each image independently and is useful for matching the reference depth flow.
• invert: flips [0,1] control values when the preprocessor convention is reversed from the LoRA's training convention.
• batch_mode: independent_images keeps image batches as separate samples when using 3D Krea2/Qwen VAEs; video_frames preserves ComfyUI's default video-style VAE behavior.

For the public depth LoRA, a good starting point with a Depth Anything output is channel_mode=grayscale, normalize=per_image_minmax, and invert=false. Turn invert on only if the depth preview shows near objects as dark instead of white.

Public depth LoRA weights: Patil/Krea-2-depth-controlnet

Other LoRA types such as canny, pose, lineart, or normal should usually use channel_mode=rgb, normalize=none, and invert=false.

The control type is determined by the LoRA checkpoint. A depth LoRA needs a depth-like control image; pose, canny, and normal LoRAs need the matching preprocessor image.

The encode node returns normal VAE-space latents. The apply node performs the same Krea2 latent-format normalization that ComfyUI applies to the sampler's main latent before the DiT sees it.

Thanks to Krea-2-controlnet for documenting the reference Krea2 control-LoRA inference pipeline. Thanks to Patil/Krea-2-depth-controlnet for providing the public depth Control LoRA weights.