HOME | DD

Description
I have had armour on my bucket list for a long time, and finally got around to it. I tried to re-train the embedding and hypernetwork for this one, but it seems like the likeness got slightly worse. On one hand the consistency was really good, but on the other hand it was just consistently not quite the face I was aiming for. Perhaps it was just unlucky RNG this time. I also noticed that the embedding really lost flexibility with more training, so I will try to re-train it again for the next image, this time using alpha channel filter to see it it helps flexibility by removing the distractions. All that said, I think the output is pretty good here, hope you enjoy it.

Prompt: a photo of GirlNextDoor2E-500 wearing full body plate armour, in a medieval courtyard, (masterpiece), (extremely intricate:1.2),(realistic), professional photograph of a stunning woman detailed, sharp focus, dramatic, award winning, cinematic lighting, volumetrics dtx, (film grain, bokeh, blurry foreground, blurry background)

Negative prompt: BadDream, (UnrealisticDream:1.2), nude, nsfw

Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 7, Size: 1920x1080, Model: absolutereality_v181, Clip skip: 2

Workflow: Train embedding and hypernetwork on the same 10 images. Generate samples in 960x544 to get the composition right, using early checkpoints of embedding to not hinder flexibility. Move to img2img and upscale to 1920x1080. Run more img2img to sharpen detail. Mask the face and inpaint with later checkpoint of embedding until likeness is all right. Done.

Description
I have had armour on my bucket list for a long time, and finally got around to it. I tried to re-train the embedding and hypernetwork for this one, but it seems like the likeness got slightly worse. On one hand the consistency was really good, but on the other hand it was just consistently not quite the face I was aiming for. Perhaps it was just unlucky RNG this time. I also noticed that the embedding really lost flexibility with more training, so I will try to re-train it again for the next image, this time using alpha channel filter to see it it helps flexibility by removing the distractions. All that said, I think the output is pretty good here, hope you enjoy it.

Prompt: a photo of GirlNextDoor2E-500 wearing full body plate armour, in a medieval courtyard, (masterpiece), (extremely intricate:1.2),(realistic), professional photograph of a stunning woman detailed, sharp focus, dramatic, award winning, cinematic lighting, volumetrics dtx, (film grain, bokeh, blurry foreground, blurry background)

Negative prompt: BadDream, (UnrealisticDream:1.2), nude, nsfw

Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 7, Size: 1920x1080, Model: absolutereality_v181, Clip skip: 2

Workflow: Train embedding and hypernetwork on the same 10 images. Generate samples in 960x544 to get the composition right, using early checkpoints of embedding to not hinder flexibility. Move to img2img and upscale to 1920x1080. Run more img2img to sharpen detail. Mask the face and inpaint with later checkpoint of embedding until likeness is all right. Done.