EimisAnimeDiffusion_2.0v

Diffusion model

This model is trained using base model as the previous version with way bigger dataset.<br> There are two versions of it:<br> EimisAnimeDiffusion_2-0 (original)<br> EimisAnimeDiffusion_2-0_alternative (original + orangemix:0.2 + even bigger dataset).<br> Read the end to choose the one you want the most.<br> At the beginning all the examples will be using "EimisAnimeDiffusion_2-0". <br>

Sample generations

Of course this model works well with anime style, magic, and a bunch of different effects. A couple of examples:<br>

Postitive:(1girl), sky, cloud, battle, armor, cape, boots, duel, scenery, outdoors, gloves, sunset, long hair, mountains, ice mountain
Negative:lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, blurry, flat, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 8, Seed: 4027860244, Size: 1024x768

<img src=https://i.imgur.com/Pvykviv.png width=75% height=75%>

Positive:1girl, solo, water, blue hair, red eye, winter, village, magician, magic circle, medium breasts, snowing
Negative:lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, blurry, flat, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 4016818418

<img src=https://i.imgur.com/BLXctxZ.jpg width=75% height=75%>

Positive:1girl, solo, ahoge, bangs, blush, bridal gauntlets, capelet, closed mouth, crossed bangs, white long dress, final fantasy, winged capelet, yellow hair, hair band, hair between eyes, hair ornament, highres, jewelry, looking at viewer, extra short hair, beautiful detailed background, solo, upper body, shoulder wing, white gold theme, indoor, royal palace, glowing light, wind, flowers
Negative:lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, blurry, flat, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 2762179779

<img src=https://i.imgur.com/C3SDGCd.jpg width=75% height=75%>

Positive: 1girl, wavy hair, medium hair, magician, blue eyes, black hair, :d, (magic circle:1.2), (black coat), full body, (ancient ruins), (scenery), sky, outdoors, landscape, stars,
Negative: lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, blurry, flat, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 477438759

<img src=https://i.imgur.com/sumnvfW.jpg width=75% height=75%>

Scenery

Positive: moon, night, tree, scenery, sky, fantasy, cloud, moonlight, outdoors, castle, mountain, tower, forest, nature, house, bridge, building, gate, bush, grass, pagoda, water, field, cliff, full moon, night sky, star (sky), starry sky, bare tree, cloudy sky, (no humans), mountainous horizon, city
Negative: lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, blurry, flat, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 561959925

<img src=https://i.imgur.com/gskGUSv.jpg width=75% height=75%>

Positive:cloud, scenery, sky, day, outdoors, grass, fantasy, landscape, mountain, (floating island:1.5), blue sky, cloudy sky, river, flowers
Negative: lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, blurry, flat, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 222763192

<img src=https://i.imgur.com/fDbLPCB.jpg width=75% height=75%>

Small comparison with v1

Right V2, Left V1.

Positive:bubble, rating:safe, underwater, jellyfish, 1girl, jacket, solo, bangs, boots, water, submerged, thighs, gloves, air bubble, bubble blowing, silver hair, very long hair, black footwear, thigh cutout, red eyes, long sleeves, black jacket, thigh strap, looking at viewer,  bare shoulders, black gloves, hair between eyes, magic
Negative:lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, blurry, flat, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 4205949473

<img src=https://i.imgur.com/ie07l2V.png width=75% height=75%>

Positive:1girl, cloud, sky, solo, magic, clock, sunset, moon, outdoors, dress, tower, sun, frills, electricity, lips, blonde hair, cloudy sky, long hair, hair ornament, wavy hair, purple eyes, looking at viewer, fire, fire magic, fire effect, electricity
Negative:lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, blurry, flat, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 1143293364

<img src=https://i.imgur.com/FhSmXqs.png width=75% height=75%>

For more in depth testing between these two:<br> Better face structures (eyes fixed)<br> Higher resolution (new data was trained on 768x768 instead of 512x512)<br> Better looking characters, animations, enviroment, effects and way more <br>

Which model to choose

EimisAnimeDiffusion_2-0 is trained on smaller dataset, however it keeps the style better.<br> It might be worse on some aspects like hardly getting specific prompts or some other small issues, however<br> it has way better quality, effects and keeps the style I wanted way better.<br> EimisAnimeDiffusion_2-0_alternative on the other hand understands better way more prompts (especially in comparison with some NSFW prompts).<br> However, way worse with style, effects, details.<br> Also sometimes not as smooth, some stuff be random, btu still really great alternative model. <br> Example:<br> Left normal, right alternative:<br>

Positive:1girl, solo, gloves, smile, tree, outdoors, :d, signature, sleeveless, skirt, breasts, hakama, bangs, fang, flower, petals, shirt, blush, standing, day, animal ears, long hair, open mouth, fox ears, cherry blossoms, japanese clothes, looking at viewer, black gloves, arm up, very long hair, bare shoulders, animal ear fluff, hakama skirt, medium breasts, cowboy shot, sleeveless shirt, grey hair, thick eyebrows, red eyes, half gloves
Negative:lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, blurry, flat, lowres, text, error, cropped, worst quality, low quality
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 518161897

<img src=https://i.imgur.com/UqDXt6X.png width=75% height=75%>

Might not be the best example, but original does have a bit more detail and more flying leaves.<br> It is way mroe noticable with magic or element effects. Also with architecture and background in general.<br> But it does understand better some characters and specific prompts. <br> For example, Hatsune Miku:

<img src=https://i.imgur.com/ivXHVbR.png width=75% height=75%>

As you can see, the alternative way better on some prompts.

Some more info

New datasets trained on clip skip 1, but clip skip 2 also works decently (not as crispy though).<br> Orangemix model link that was used in the alternative:<br> https://huggingface.co/WarriorMama777/OrangeMixs