HomoDiffusion 2 Anime

Simple merge of MatureMaleMixV1.4 and ToonYouV5 Unstable with HomoDiffusion 2.0

Mix is capable of producing same quality images as HomoDiffusion 2.0 the only difference is due to merging via SmoothAdd MT it inherited style of both aforementioned models.

Both style of prompting will work i.e anime tags and normal, natural language.

Recommended VAE:

vae-ft-mse-840000-ema-pruned

General Tips

Sentence or prompt structure matters in both SD Next and InvokeAI 3.0. They both handle them differently due to different pipelines, parser and underlining optimisations. Additionally depending on WebUI and it's version some of prompts will be restricted to 75 tokens, meaning anything above that count will be either completely ignored or just split up. Latest version of InvokeAI & SD.Next support up to 225 tokens.

For example:

If you structure your prompt the same way like in first example i.e description of quality, lighting condition, blurriness (Gaussian, Bokeh, depth of field, aperture, focus, ISO, DLSR or mirror-less make & model and so forth), resolution properties, different types of compression artefacts and by-products (i.e JPEG artefacts) and then the subject i.e muscular/skinny/lean male/man/men, facial& body features, clothing preferences, details regarding background you will end up with gens like the one below.

Example no.1:

Prompt:
RAW photo, best quality, masterpiece, dramatic, intricate, high res, photorealistic, cinematic lighting, HDR, a 25-year-old man baking a cake in a cosy kitchen, detailed face, wrinkled face, short scruffy brown hair, hard ???????, coral eyes, beard, hairy body, ????????? pose, ???????? facial expression, thanksgiving atmosphere, shirtless, pyjama bottoms, (rear shot)+, (seen from the back)+, back view, looking at viewer, sunset, close-up,

But if you restructure the whole sentence and first describe person/objects and background, then add all the other descriptors, there's high chance you'll end up with something complete different.

Example no.1:

Prompt:
a 25-year-old man baking a cake in a cosy kitchen, detailed face, wrinkled face, short scruffy brown hair, hard ???????, coral eyes, beard, hairy body, ????????? pose, ???????? facial expression, thanksgiving atmosphere, shirtless, pyjama bottoms, (rear shot)+, (seen from the back)+, back view, looking at viewer, sunset, close-up, RAW photo, best quality, masterpiece, dramatic, intricate, high res, photorealistic, cinematic lighting, HDR, 

Example no.2:

Prompt:
RAW photo, best quality, masterpiece, dramatic, intricate, high res, photorealistic, cinematic lighting, HDR, (rear shot)+, (seen from the back)+, back view, looking at viewer, close-up, a 25-year-old man baking a cake in a cosy kitchen, detailed face, wrinkled face, short scruffy brown hair, hard ???????, coral eyes, beard, hairy body, ????????? pose, ???????? facial expression, thanksgiving atmosphere, shirtless, pyjama bottoms, sunset,

Both examples were made in Invoke AI 3..0.0 using txt2img.

Recommended settings:

SD Next:

Step Count:

From 20 to 70

Samplers:

Euler A, DPM++ a Karras, DPM++ 2M/2S a Karras

Second Pass & Upscaler:

4x-UltraSharp with denoising strength between 0.1 to 0.3, 5 to 50 steps. For second pass sampler you can use combination of sampler mentioned above. Some of them will handle better adding details, textures, some might cause extra issues like double glands, elongated limbs, missing limbs or blurriness.

Refiner:

Since SD Next lets you mix&match not only samplers but refiner for SD 0.9 XL with SD 1.5 base models I do recommend to at least play with new refiners to increase fidelity of generations, get crispy details or to fix some issues.

aDetailer:

Face either mediapipe mesh/full or yolov8s,

Detection threshold varies from 0.3 to 0.45,

Steps count for aDetailer model 12-30,

Resolution for in-painting from 512x512 to 768x768 - it will slowdown generation significantly.

Resolution:

from 512px to 1024px with aspect ratio from 4:3, 3:4 or 16:9, 9:16.

InvokeAI 3.0

Step Count:

15-75

Samplers:

DPM++2M SDE Karras, DPM++2M Karras, DPM++2S, Euler Ancestral, Euler Karras, Huen, Heun Karras,

Upscaler:

RealESRGAN_x4plus

Resolution:

768px up to 1280px.

Portrait, Free or Wide.

Disclaimer:

Some of prompts are not mine. They're either found here, on CivitAI or taken from user of Unstable Discord #men-only channel.

Model, recommended resources, hint and description of each of my model may change over time with new LoRA/TI tested or/and with new versions of HD2.