Nynxz's Krea 2 Edit (NK2E)
A set of in-context LoRAs for Krea 2. The edit LoRAs take a source image and a text instruction (like "give her pink hair" or "add sunglasses") and edit the image. A separate canny control LoRA takes an edge map and generates an image that follows it.
Experimental and unofficial. Results are rough and inconsistent, and vary a lot by edit type. Not affiliated with or endorsed by Krea.
Demo
Try it in your browser — no install required — on the Hugging Face Space:
→ https://huggingface.co/spaces/nynxz/nk2e-krea2-editor
Links
- Live demo (Space): https://huggingface.co/spaces/nynxz/nk2e-krea2-editor
- ComfyUI Node Pack: https://github.com/Nynxz/ComfyUI-NK2E
- Training Toolkit: https://github.com/Nynxz/NK2E
Use (ComfyUI)
- Download a LoRA from
comfy/v0.3/(recommended) into your ComfyUImodels/loras/. - Install the node pack (link above) and load its example workflow.
Strength around 0.7 is a good start. The reference is the source image (or, for the canny LoRA, an edge map).
Versions
| Version | File | Rank | Steps | Training data | Notes |
|---|---|---|---|---|---|
| v0.3 | comfy/v0.3/NK2E-v0.3.safetensors | 64 | 24k | v0.1 mix + UltraEdit + HQ-Edit (~72k pairs) | recommended (newest) |
| v0.2 | comfy/v0.2/NK2E-v0.2.safetensors | 64 | 12k | same ~72k mix | continued from v0.1 |
| v0.1 | comfy/v0.1/NK2E-v0.1.safetensors | 64 | 8k | MagicBrush + OmniEdit + InstructPix2Pix (~37k) | first release |
| canny v0.1 | comfy/canny_v0.1/NK2E-canny-v0.1.safetensors | 32 | 12k | 8k COCO canny pairs | control LoRA (see below) |
See CHANGELOG.md for history.
What it does well (and not)
- Good at: keeping identity and composition, localized edits (hair, accessories, objects), and global style or lighting changes.
- Weaker at: recolouring a whole large region, and scene or structure changes (swapping a background, adding a person, re-posing).
It is not a mask or inpaint model; identity is preserved without a mask.
Canny control (experimental)
comfy/canny_v0.1/ is a separate control LoRA (ControlNet-style, in-context). Feed a
canny edge map as the reference instead of a source image: structure comes from the
edges, content from the text prompt. Same node setup as editing. It follows the outline
but is softer than a pixel-locked ControlNet. Depth and pose are planned, same recipe.
Specialist LoRAs
comfy/loras/ holds narrow, single-purpose edit LoRAs trained on top of the generalist.
Each responds to one trigger instruction rather than general edits, and loads with the
same node setup. Because they are low-rank and concept-specific, they usually want a higher
strength (start near 1.0).
| Specialist | File | Base | Trigger |
|---|---|---|---|
| Invisible Man | comfy/loras/NK2E-InvisibleMan-v0.1.safetensors | v0.3 | make person invisible |
See comfy/loras/README.md for details and triggers.
Training a specialist on top
comfy/ holds ComfyUI-format weights for inference; train/ holds native-format weights
for training. Point the toolkit's base_lora at train/v0.3/NK2E-v0.3-train.safetensors
(the latest generalist native) to train a narrow specialist on top — this is how the
specialists in comfy/loras/ were made. See the training toolkit (link above).
License
Krea 2 Community License (see LICENSE and NOTICE); commercial use only under $1M/yr revenue. Not an official Krea project.