Hi all. welcome back to the post about SPLINTERLANDS WEEKLY ART CHALLENGE. This time I will try to make a character LORA. And the character that I will make her LORA this time is LIRA THE DARK.
To make a proper LORA, you need at least 20 reference images (be it characters, costumes, artsytle). But I was desperate and tried to make LORA with only 1 image above. The steps to make LORA came from here, but because I used only 1 image, there are some steps that were changed.
Making Datasets
The first step is to create datasets. To create it we need help from this Google colab.
Type the project name (liradark) and select folder structure >> organize by project. After that press the play button.
Put the reference image into the liradark folder (project name)
We skip the second and third steps (because we select and sort the reference images ourselves). The fourth step we will create tags (descriptions) for each reference image (here only 1)
set tag_threshold to 0.5 because we will be creating character LoRAs (0.35 for artstyles LoRAs, 0.5 for everything else). Delete everything in blacklist_tags. Press the play button.
In the fifth step, we will create a new description that will replace the general description in the fourth step. The point is that the program can better understand the character when we type the description later in the LoRA testing process.
If we skip this fifth step, when typing in the description (1girl, long hair, purple hair) there is a possibility that the result that will appear is not Lira the Dark.
Before the sixth step, we scroll down to the EXTRA section to check if the new description (created in the fifth step) has replaced the old one.
Click the play button on the sixth step
The process of creating datasets is complete. Don't forget to disconnect Google colab before proceeding to the next step.
Creating LORA
Use this Google colab in the process of creating LoRA.
The project name and folder structure are the same as the previous step.
Use the settings as below when we only have 1 reference image. (This is just my reference)
Click the play button.
LoRA will be saved in "lora_training/output/liradark". Don't forget to disconnect after finishing the process.
Testing LORA
Use this Google colab for this process.
Change BRANCH to Master.
Put the links from the liradark folder into custom_urls. You can also insert models other than those provided by this google colab into custom_urls (go to CIVITAI)
click the play button.
Select the link in the public url
Results
To know how the program looks and how it works, you can check it out HERE. In that post, I made some backgrounds and did some upscaling..
Here I will just share the output and what prompt I used.
| 1 |  | prompt: liradark, lira-dress, lira-accessories, lira-hood, full body, looking at viewer, <lora:liradark-02:1> Negative prompt: (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2) Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 1077252573, Size: 512x768 Model: AnyLine-fp16, Denoising strength: 0.5, Clip skip: 2, CFG Rescale phi: 0 Hires upscale: 1.5, Hires upscaler: 4x-UltraSharp, Lora hashes: "liradark-02: 2be19d987c83", Version: v1.4.0 |
| 2 |  | prompt: liradark, lira-dress, lira-accessories, lira-hood, full body, dark background, <lora:liradark-01:1> Negative prompt: (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2) Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 4108523014, Size: 512x768, Model: AnyLine-fp16, Denoising strength: 0.5, Clip skip: 2, CFG Rescale phi: 0, Hires upscale: 1.5, Hires upscaler: 4x-UltraSharp, Lora hashes: "liradark-01: b1bcd4aa82fd", Version: v1.4.0 |
| 3 |  | prompt: liradark, lira-dress, lira-accessories, lira-hood, glowing eyes, (evil smile:0.4), upper body, black background, dungeon, <lora:liradark-01:1> Negative prompt: (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2) Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 12832323, Size: 512x768, Model: AnyLine-fp16, Denoising strength: 0.5, Clip skip: 2, CFG Rescale phi: 0, Hires upscale: 1.5, Hires upscaler: 4x-UltraSharp, Lora hashes: "liradark-01: b1bcd4aa82fd", Version: v1.4.0 |
| 4 |  | prompt: liradark, lira-dress, lira-accessories, lira-hood, hood up, shaded face, glowing eyes, (evil smile:0.4), upper body, black background, dungeon, <lora:liradark-02:1> Negative prompt: (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2) Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 2206889952, Size: 512x768, Model: AnyLine-fp16, Denoising strength: 0.5, Clip skip: 2, CFG Rescale phi: 0, Hires upscale: 1.5, Hires upscaler: 4x-UltraSharp, Lora hashes: "liradark-02: 2be19d987c83", Version: v1.4.0 |
| 5 |  | prompt: (liradark:1.3), solo, lira-dress, lira-accessories, lira-hood, glowing eyes, (evil smile:0.4), cowboy shot, (black background:1.3), <lora:liradark-03:.8:1,0,0,0,1,1,1,1,1,1,1,0,0,0,0,0,0> Negative prompt: (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2) Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 4109916847, Size: 512x768, Model: AnyLine-fp16, Denoising strength: 0.5, Clip skip: 2, CFG Rescale phi: 0, Hires upscale: 1.5, Hires upscaler: 4x-UltraSharp, Lora hashes: "liradark-03: adc3ce8d2f7b", Version: v1.4.0 |
| 6 |  | prompt: (liradark:1.3), solo, lira-dress, lira-accessories, lira-hood, glowing eyes, (evil smile:0.4), cowboy shot, (black foreground:1.3), <lora:liradark-03:.8:1,1,1,1,0,0,0,0,1,1,1,0,0,0,0,0,0> Negative prompt: (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2) Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 1089849151, Size: 512x768, Model: AnyLine-fp16, Denoising strength: 0.5, Clip skip: 2, CFG Rescale phi: 0, Hires upscale: 1.5, Hires upscaler: 4x-UltraSharp, Lora hashes: "liradark-03: adc3ce8d2f7b", Version: v1.4.0 |
| 7 |  | prompt: (liradark:1.3), solo, lira-dress, lira-accessories, lira-hood, glowing eyes, (evil smile:0.4), cowboy shot, (black foreground:1.3), <lora:liradark-03:1:1,1,1,1,0,0,0,0,1,1,1,0,0,0,0,0,0> Negative prompt: (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2) Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 3542798437, Size: 512x768, Model: AnyLine-fp16, Denoising strength: 0.5, Clip skip: 2, CFG Rescale phi: 0, Hires upscale: 1.5, Hires upscaler: 4x-UltraSharp, Lora hashes: "liradark-03: adc3ce8d2f7b", Version: v1.4.0 |
| 8 |  | prompt: liradark, lira-dress, lira-accessories, lira-hood, (purple hair:1.2), solo, glowing eyes, (evil smile:0.4), fighting stance, hand up, (black foreground:1.3), <lora:liradark-01:1:1,1,1,1,0,0,0,0,1,1,1,0,0,0,0,0,0> Negative prompt: (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), bad-hands-3, hair between eyes Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 231649986, Size: 512x768, Model: AnylineUPfp16 x 1.1 + El Zipang_pruned_fp16 x -0.1, Denoising strength: 0.5, Clip skip: 2, Hires upscale: 1.5, Hires upscaler: 4x-UltraSharp, Lora hashes: "liradark-01: b1bcd4aa82fd", Version: v1.4.0 |
| 9 |  | prompt: liradark, lira-dress, lira-accessories, lira-hood, (purple hair:1.3), solo, shaded face, glowing eyes, (evil smile:0.4), fighting stance, hand up, (black foreground:1.3), dungeon, <lora:liradark-01:1:1,1,1,1,0,0,0,0,1,1,1,0,0,0,0,0,0> Negative prompt: (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), bad-hands-3, ((hair between eyes, magic)) Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 1170271957, Size: 512x768, Model: AnylineUPfp16 x 1.1 + El Zipang_pruned_fp16 x -0.1, Denoising strength: 0.5, Clip skip: 2, Hires upscale: 1.5, Hires upscaler: 4x-UltraSharp, Lora hashes: "liradark-01: b1bcd4aa82fd", Version: v1.4.0 |
In numbers 1 and 2, the output is still the same as the reference image. But unfortunately, there is a misunderstanding, where the hood changes to the hair color and the "flower" on the dress disappears..
So it can be concluded that the result of LoRA fails to match the reference image and instead creates a new character (or a glimpse of the one in the reference image).
And here are some results using other models that have not been upscaled yet.
 |  |
|---|---|
| prompt: masterpiece, best quality, BREAK, liradark, lira-dress, lira-accessories, lira-hood, (purple hair:1.3), solo, shaded face, glowing eyes, (evil smile:0.4), fighting stance, hand up, anime gacha SSR Character, <lora:liradark-01:1:0,0,0,0,0,0,0,0,1,1,1,0,0,0,0,0,0>, (dark theme, horror (theme)) Negative prompt: (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), bad-hands-3, (hair between eyes:1.4, magic:1.5) Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 3056152762, Size: 768x512, Model: AnylineUPfp16 x 1.1 + El Zipang_pruned_fp16 x -0.1, Clip skip: 2, Lora hashes: "liradark-01: b1bcd4aa82fd", Version: v1.4.0 | prompt: masterpiece, best quality, BREAK, liradark, lira-dress, lira-accessories, lira-hood, (purple hair:1.3), solo, shaded face, glowing eyes, (evil smile:0.4), fighting stance, hand up, anime gacha SSR Character, <lora:liradark-01:1:0,0,0,0,0,0,0,0,1,1,1,0,0,0,0,0,0>, (space, galaxy, dark theme, horror (theme)) Negative prompt: (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2), bad-hands-3, (hair between eyes:1.4, magic:1.5) Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 3145971320, Size: 768x512, Model: AnylineUPfp16 x 1.1 + El Zipang_pruned_fp16 x -0.1, Clip skip: 2, Lora hashes: "liradark-01: b1bcd4aa82fd", Version: v1.4.0 |
From the 8 images, I selected 2 images that I enlarged using the img2img tab (an example of the process is in the previous post).
That was the process of making LoRA from Lira the Dark. Sorry if what I explained earlier there are some that you can't understand (because I'm not good at explaining things). Thank you for visiting and see you in posts about LoRA from other characters...