Buckets:
| # Changelog | |
| ## Invisible Man v0.1 (2026-07-05) | |
| First specialist LoRA — a narrow edit adapter trained on top of the v0.3 generalist. | |
| - Rank-8 in-context edit LoRA; responds to the trigger instruction `make person invisible`. | |
| - Makes the person in the source image vanish while keeping clothing, props and scene. | |
| - comfy/loras/NK2E-InvisibleMan-v0.1.safetensors (ComfyUI). Start around 1.0 strength. | |
| - See comfy/loras/README.md for triggers and usage. | |
| ## v0.3 (2026-07-04) | |
| Recommended release. | |
| - Rank-64 in-context edit LoRA, continued to 24k steps. | |
| - Training data: the v0.1 mix plus UltraEdit and HQ-Edit (~72k instruction/edit pairs). | |
| - comfy/v0.3/NK2E-v0.3.safetensors: recommended (ComfyUI). | |
| - train/v0.3/NK2E-v0.3-train.safetensors: native fp16, for training specialists on top. | |
| ## canny v0.1 (2026-07-04) | |
| First control LoRA — a separate, ControlNet-style in-context adapter. | |
| - Rank-32, 12k steps, trained on ~8k COCO canny-edge/image pairs. | |
| - 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 the edit LoRAs. | |
| - comfy/canny_v0.1/NK2E-canny-v0.1.safetensors (ComfyUI). | |
| - Softer than a pixel-locked ControlNet. Depth and pose are planned, same recipe. | |
| ## v0.2 (2026-07-02) | |
| - Rank-64 in-context edit LoRA, continued from v0.1 to 12k steps. | |
| - Training data: the same ~72k mix (MagicBrush, OmniEdit, InstructPix2Pix, UltraEdit, HQ-Edit). | |
| - comfy/v0.2/NK2E-v0.2.safetensors (ComfyUI). | |
| - train/v0.2/NK2E-v0.2-train.safetensors: native fp16, for training specialists on top. | |
| ## v0.1 (2026-06-30) | |
| First Hugging Face release. | |
| - Rank-64 in-context edit LoRA for Krea 2 Raw. | |
| - Trained on about 37k instruction/edit pairs (MagicBrush, OmniEdit, InstructPix2Pix). | |
| - comfy/v0.1/NK2E-v0.1.safetensors: recommended (ComfyUI). | |
| - comfy/v0.1/NK2E-v0.1-step8000.safetensors: earlier checkpoint, less baked. | |
| - train/v0.1/NK2E-v0.1-train.safetensors: native fp16, for training specialists on top. | |
Xet Storage Details
- Size:
- 1.99 kB
- Xet hash:
- 65bb2a2ad5988fb409e17dcff33236107f532e18a454840930a6cf527f547b45
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.