CyberKrea Enhancer Advanced

CyberKrea Enhancer Advanced

A standalone ComfyUI module containing one node: CyberKrea Enhancer Advanced. Based on the Advanced node from the supplied ComfyUI-Krea2T-Enhancer archive by capitan01R. Other nodes are not included or imported.

The node enhances Krea2's text-fusion path and provides an independent control for text tokens after txtmlp. The original enhancement profile and defaults are preserved. This version improves error handling and cleanup of temporary patches; improved image quality has not been demonstrated.

Installation

Comfy Registry

Find CyberKrea Enhancer Advanced by cyberdelia in ComfyUI-Manager, or install with Comfy CLI:

comfy node install comfyui-cyberkrea-enhancer

See the Comfy Registry page. New releases may take time to become installable while Registry checks are pending. Restart ComfyUI after installation.

Git or ZIP

Clone the repository into your ComfyUI custom nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/cyberdeliaAI/ComfyUI-CyberKrea-Enhancer.git

Then fully restart ComfyUI and search for CyberKrea Enhancer Advanced.

Alternatively, install the ZIP release:

1. Download ComfyUI-CyberKrea-Enhancer-1.0.0.zip from the GitHub release and extract it into ComfyUI/custom_nodes/.

2. Check the folder structure:

   ComfyUI/
     custom_nodes/
       ComfyUI-CyberKrea-Enhancer/
         __init__.py
         nodes.py
         fusion.py
         ...
   

3. Fully restart ComfyUI.

4. Search for CyberKrea Enhancer Advanced when adding a node. Its category is CyberKrea/conditioning.

No additional Python packages are required in a working ComfyUI installation with native Krea2 support. The module uses ComfyUI's existing PyTorch installation. It does not download models or make network requests.

You can remove the original module by moving its folder out of custom_nodes. First check whether other workflows still need it. This module uses its own node ID and can be loaded alongside the original. Avoid using both enhancers on the same model path, as their effects will be combined.

Connections

Model loader → optional LoRA loader(s) → CyberKrea Enhancer Advanced → sampler (MODEL)

The node has one MODEL input and one MODEL output. Your workflow's existing CLIP, conditioning, latent and VAE connections follow their usual paths. No additional text encoder or special prompt format is required.

Settings

| Setting | Default | Description | | --- | --- | --- | | enabled | true | Enables or disables this enhancer. | | strength | 1.0 | Text-fusion enhancement, range 0.0–2.0. 0 disables this component; 1 uses the original Advanced profile. | | text_scale | 1.0 | Multiplies text tokens after txtmlp, range 0.25–4.0. 1 is neutral for this component. This is not a CFG setting. | | debug | false | Writes diagnostics to the ComfyUI console. Adds processing overhead. |

Fully neutral: enabled=false, or strength=0 together with text_scale=1. With strength=0 and a different text_scale, text scaling remains active.

Start with the defaults to match the original Advanced node's behavior. Compare against enabled=false using the same seed, prompt, resolution, sampler and LoRA settings. Change one setting at a time. Higher values are not automatically better: composition, color and details may change as well.

Improvements

• A dedicated node ID: CyberKreaEnhancerAdvanced. Workflows using the old node ID are not converted automatically.
• The fallback for an unexpected tensor shape calls the actual saved original forward function. In the archive, the helper referenced the attribute name used by the basic enhancer instead of the Advanced version.
• An unexpected tensor rank is also passed to the original forward function. This avoids a failure inside the enhancer itself; the original model may still correctly reject invalid input.
• Temporary forward patches are restored exactly, including after exceptions. An inherited method remains inherited; no extra forward instance attribute is left behind.
• The model configuration no longer receives a temporary _active flag. A separate runtime context prevents repeated application during nested calls.
• Reapplying the node replaces its own wrapper and preserves other wrappers. No wrapper is registered when disabled or fully neutral.
• Step/sigma diagnostics are only computed when debug is enabled. This avoids the associated GPU-to-CPU synchronization during normal use.
• An incompatible model produces a clear error when enhancement is active, rather than silently continuing without enhancement.

Preserved behavior

The original 24 chunk weights, global multiplier 1 + strength × 14, two-pass text-fusion calculation and per-token relative delta cap of 0.75 are preserved. The cap limits the difference from the original text-fusion output before txtmlp and text_scale; it does not limit the complete model output or the visible effect.

With nonzero strength, the text-fusion path runs twice: once for a reference and once for an enhanced candidate. The complete image generator does not run twice. When only text_scale is active, the regular text-fusion path is used.

Compatibility and test status

Requires ComfyUI with native Krea2 support, the 12 × 2560 text layout and the keyed model-wrapper API. Use a Krea2 text encoder in the rest of your workflow. This enhancer targets Krea2, rather than Flux, SDXL or other architectures. No arbitrary minimum version number is specified: the required interfaces and model structure are checked directly.

CPU regression tests cover the enhancement profile, bypass, text scaling, fallback, wrapper registration and recovery after errors. Numerical comparison against a preserved excerpt from the archive uses synthetic text-fusion modules in float32, float16 and bfloat16. See VALIDATION.md for the results and their precise limitations.

No complete Krea2 image generation has been performed here. Visual quality, CUDA/MPS inference, GGUF or other custom loaders, torch.compile and combinations with other nodes that replace the same forward methods have not been validated. The node is intended for inference, not training.

The lock protects concurrent patched calls made through this module. It does not make arbitrary concurrent use of the same underlying model by other modules safe. Use normal sequential ComfyUI execution.

Running the tests

From a source checkout of this repository, using ComfyUI's Python environment:

python -m unittest discover -s tests -v

By default, the tests use a small test double for ComfyUI's wrapper API. To test with the actual patcher_extension.py:

CYBERKREA_PATCHER_EXTENSION=/path/to/ComfyUI/comfy/patcher_extension.py python -m unittest discover -s tests -v

The second option is also a component test, not a complete ComfyUI session. Development tests and GitHub workflow files are excluded from Registry packages.

Attribution and license

Original code: capitan01R, ComfyUI-Krea2T-Enhancer, MIT. The original LICENSE is preserved; see also NOTICE. This fork is not an official release by the original author.