Driven by Araminta’s recent post on impressive LoRAs (Learn Once, Run Anywhere models), the author found motivation to resume training their own LoRA, called “Rouge Noir”, which had been left incomplete for some time. The style of this AI model is based on the dramatic contrast between red and black shades, producing silhouette images with reflections of the primary subjects.
The initial stages of training the model involved using a large dataset that consisted of over 50 images. However, the author’s RTX4080 computing system was unable to handle this extensive data load. In the search for a workaround, the author used Runpod to train the images. This didn’t involve the Rouge Noir model at this point. Instead, the author worked on a different model using Octane Render tokens that resulted in high-quality 3D rendered images.
The author’s passion for the Rouge Noir model was rekindled through observing Araminta’s work. Consequently, the author resumed and updated the Kohya_ss training, this time with a more selective dataset comprising 26 diversified images. Araminta’s tips were instrumental in refining the training settings, delivering a much-improved, cleaner training style for the Rouge Noir model.
That being said, the author is considering expanding the model’s capabilities by training it for the SD 1.5 model as well, contingent on the feedback and reception it receives. To give a glimpse of the model’s potential, the author shared a few sample images generated by this LoRA model. These images were created using Base SDXL, along with other refined SDXL models featured in the author’s workflow.
Evidently, the Rouge Noir model holds a special place in the author’s work, being one of their favourite LoRAs produced so far. The author encourages other creators to experiment and create fun images with this model and invites them to leave their queries or comments about the model.
