CodeChris/EleetModel

Eleet Model: an anime style stable diffusion model

This model is also available on: Eleet Model - Civitai.

Eleet model is a block-weighted merged stable diffusion model aiming at generating good quality 2D anime style images that satisfy my personal taste and, hopefully, your taste.

ELEET is an Internet slang for the word 'elite', which dates back to 80s/90s. Its modified spelling in a digital form is 31337, which has been the most common Eta Noise Seed Delta (ENSD) value in stable diffusion probably since the NAI era. The model is named in honor of this tradition.

• Suggested settings
• Samples
• Merge ideas
• License

Suggested settings

For users who know little about stable diffusion settings, I recommend:

• Prompt: Start with masterpiece, best quality, aesthetic. Personally I also like to add photography phrases such as cinematic lighting, professional shadow, etc.

• Negative prompt:

  (worst quality, low quality:1.4), lowres, bad anatomy, (blurry), (text, logo, watermark, signature, username)

• Sampler: DPM++ 2M Karras

• CFG Scale: 6~9

• Steps: 16~30

• Highres Fix (optional): Latent sampler; 0.6~0.7 Denoising strength; 16 Highres steps.

• Clip skip: 2

• No external VAE needed

Samples

Here are 4 samples of the latest Eleet model version.

Sample 1 (Highres Fix from 512x800):

masterpiece, best quality, aesthetic, 1girl, solo, black eyes, green hair, low-tied long hair, school uniform, sitting, wariza, thighhighs, black thighhighs, from above, looking at viewer, (cityscape), cinematic lighting, professional shadow

Other common settings through all samples:

Negative prompt: (worst quality, low quality:1.4), lowres, bad anatomy, (child, loli), (blurry), (text, logo, watermark, signature, username)

Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 1.25, Hires steps: 16, Hires upscaler: Latent

Sample 2 (Highres Fix from 512x800):

masterpiece, best quality, aesthetic, 1girl, solo, red eyes, one eye closed, brown hair, long hair, grin, arms up, crop top, denim shorts, midriff, navel, athletic body, earings, (cowboy shot), looking at viewer, (waterfall), sunny, cinematic lighting, professional shadow

Sample 3 (Highres Fix from 832x512):

masterpiece, best quality, aesthetic, no humans, scenery, sky, cloud, outdoors, mountain, water, sunset, cloudy sky, lake, river, landscape, reflection, tree, nature, mountainous horizon, blue sky, evening, professional shadow, award winner photo

Sample 4 (Highres Fix from 800x512):

masterpiece, best quality, aesthetic, forest, grass, blue sky, cloud, outdoors, no humans, nature, scenery, railroad tracks, sunlight, sunbeam, lens flare, cinematic lighting, professional shadow

Additionally, here are samples of the previous model versions. Click to expand:

Eleet v1.0 samples

Sample 1: Scenery. txt2img+highres, (640x384) x1.5.

masterpiece, best quality, aesthetic, highres RAW photo, landscape photography, wide shot, from below, scenery, sunrise, blue sky, clouds, lake, reflection, sun, trees, floating leaves, ripples, foreground interest, depth of field, cinematic lighting, asymmetric composition, professional shadows, sharp focus, lens flare

Sample 2: Anime girl. txt2img, 576x832.

masterpiece, best quality, aesthetic, 1girl, :d, blue eyes, fox ears, gold hair, twin braids, crop top, off-shoulder jacket, blue pleated skirt, thighhighs, garter belt, skindentation, breasts, bare shoulders, midriff, (cowboy shot), looking at viewer, waterfall, mountains

Sample 3: Scenery. txt2img, 832x576.

masterpiece, best quality, aesthetic, highres RAW photo, cool color tone, wide shot, scenery, snow, blue sky, cityscape, skyline, buildings, rooftop, roads, cinematic lighting, professional shadows, sharp focus

Sample 4: Anime girl. txt2img+highres, (384x640) x1.5.

masterpiece, best quality, 1girl, solo, angry, brown eyes, blue hair, bangs, sidelocks, half updo, clenched hand, parted lips, [bodysuit|armored dress], long sleeves, gloves, armor, looking to the side, clenched teeth, (cowboy shot), night, full moon, forest, underexposure, professional lighting

Merge ideas

The weights for merging Eleet model were optimized through an automatic procedure with scoring, but I didn't necessarily pick the best-scored one as the final version. Instead, I will evaluate a number of high-scored candidates and score their outputs manually by myself.

I conducted the evaluation mostly on anime girls topics (no doubt) but also considered the model performance on scenery images. I will consider:

• Prompt response
• Image aesthetic quality over 3 scenarios:
  1. txt2img under --lowvram mode on a low-end GPU
  2. txt2img under normal or --medvram mode on a better GPU
  3. txt2img + highres fix on the previous GPU
• Image flaws (color shift, illogical drawing, etc.) over the above scenarios

License

CreativeML OpenRAIL-M.