Krea-2-UAP-IR-LoRA

Krea 2 UAP IR LoRA

A style and concept LoRA for generating monochrome infrared, electro-optical, and military surveillance imagery with Krea 2 Turbo. It emphasizes degraded sensor texture, thermal contrast, targeting-camera graphics, distant aerial objects, and physically grounded impossible structures.

This is an independent community derivative. It is not an official Krea product and is not affiliated with or endorsed by Krea or the U.S. Department of Defense.

Release

• Version: 1.0
• Weight: Krea-2-UAP-IR-LoRA-v1.0.safetensors
• Released checkpoint: 1,250 training steps
• LoRA rank/alpha: 32/32
• Base model: krea/Krea-2-Turbo
• Trainer: ostris/ai-toolkit
• Training adapter: ostris/krea2_turbo_training_adapter

Trigger

No trigger word was used. The dataset was trained without captions. Describe the intended infrared or military sensor appearance directly in the prompt.

Usage

Load the LoRA on top of Krea 2 Turbo. A LoRA strength of 1.0 was used for the gallery. Lower values such as 0.7-0.9 can retain more of the base model's clean rendering while keeping the sensor-footage character.

Example prompt:

black and white infrared camera view of a giant mysterious fortress-like structure over the ocean, distant water horizon, grainy thermal surveillance footage

Another useful structure:

infrared black-and-white thermal reconnaissance footage still, grayscale only, no color, [subject and environment], distant observation framing, grainy thermal image texture, slight blur, no text, no overlays

Because there is no learned trigger token, ordinary subject and composition language remains important. Terms such as infrared, thermal reconnaissance, targeting camera, reticle, monochrome, and low-resolution sensor footage usually strengthen the learned appearance.

Gallery

All examples below were generated with the released final 1,250-step v1.0 weight at LoRA strength 1.0.

The original ComfyUI prompt and workflow metadata remains embedded in each PNG. examples_metadata.json contains extracted prompts, seeds, dimensions, sampler settings, LoRA strength, and checkpoint names for all eight examples.

Training Data

The dataset contains 127 manually selected frames from 27 public DVIDS videos. It is dominated by infrared footage, with additional full-motion video, electro-optical, and unspecified sensor imagery:

Candidate frames were extracted near the midpoint of each five-second interval. They were then reviewed in two human-curation passes: within each source video and across the complete candidate set. Training images are not redistributed in this model repository.

source_manifest.csv lists each selected frame's timestamp, source video, title, sensor label, and DVIDS source page. Local file paths have been removed.

DVIDS states that works prepared by U.S. Government employees as part of their official duties are generally not eligible for U.S. copyright protection, but some media on DVIDS may contain third-party rights. Check the rights information on each linked source page before redistributing source media. See the DVIDS copyright information.

Training Parameters

The LoRA was trained with AI Toolkit on an NVIDIA GeForce RTX 4070 Ti 12 GB using Krea 2 Turbo with the Krea 2 Turbo Training Adapter.

• Steps: 1,250
• Batch size: 1
• Optimizer: AdamW 8-bit
• Learning rate: 5e-4
• Precision: bf16
• Resolution targets: 512, 768, and 1024
• LoRA rank/alpha: 32/32
• Gradient checkpointing: enabled
• Text encoder training: disabled
• Captioning: none

The sanitized settings and release checksum are available in training_parameters.json.

Limitations and Responsible Use

• This LoRA learns the appearance of surveillance and sensor imagery; it does not detect, identify, or provide evidence of real aerial phenomena.
• It may overproduce reticles, overlays, noise, blur, compression artifacts, or small bright objects.
• Outputs can resemble authentic military footage. Generated images should be identified as synthetic and must not be presented as genuine government, military, intelligence, or UAP evidence.
• Follow the Krea 2 Acceptable Use Policy and apply appropriate content review or filtering when deploying the model.

License

This LoRA is a derivative of Krea 2 and is distributed under the Krea 2 Community License Agreement. The full agreement is included as LICENSE.pdf, and required attribution is provided in NOTICE.