Many thanks for such a great Widowmaker and Mercy embeddings goes to Master_Corneo
Also i really appreciate luisa_pinguin for great guide about hypernetworks training that helped me beat my fear of training tab.
So, about hypernetwork. It was trained on some Midjourney generations so it makes anime generations slightly more semi-realistic and adds some bloom and blurred background with some soft light.
Trained on Anything.v3 with clip_skip 2 so it should work with almost any other anime-based model.
Showcase images was generated with AbyssOrangeMix2_hard and clip_skip 2.
In general this hypernetwork is a child of luck and many tries, so if you know exactly how to train it properly and what all the parameters means, please share your knowledge.
Description
FAQ
Comments (7)
This looks awesome! The things people are accomplishing with hyper networks is blowing me away!
I case you don't already have it, I'd totally recommend the kl-f8-anime2 VAE recommended and available on the Kenshi model.
Hey, quick question about your widowmaker images: are the provided prompts the originals from txt2img, or are they the final from the inpainting? The results are awesome. What changes did you make to achieve them? Do you recall, roughly?
Hey. Yeah, txt2img, parameters was like:
dark night scene, best quality, 1girl, (upper body:1.2), digital illustration, beautiful shapely (corneo_widowmaker_v2) aims her rifle at the viewer, standing in the dark, (beautiful face), (athletic:1.3), (wide hips:1.1), (thick thighs:1.1), (night Paris with Eiffel Tower in detailed background:1.3), dark aesthetic, (smirk:1.2), film grain, lens distortion, (chromatic aberration), erotic, (muscular:0.7), torn tights, (elbow protection pad), (lips), intricate, unreal engine 5, volumetric lighting, realistic, realistic lighting, cinematic, 4k, cinematic lighting, depth of field, masterpiece, perfect, award-winning, hyper-detailed, photorealistic, ultra realistic, realistic light, hard lighting, intricate details, stop motion, hyperfocus, tonemapping, sharp focus, hyper detailed, (fog:0.7), a character portrait by senior character artist, trending on Artstation,<hypernet:sxz-bloom:0.8>
Negative prompt: (bad_prompt_version2:0.7), (loli), african, cartoon, low quality, close up, old, lowres, text, error, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, black and white, monochrome, censored, (out of frame:1.3), (short skirt), walking, gold, weapon, knife, sword, bride, oversaturated, (fat), ((overexposed)), bad eyes, twisted hands, white skin, pale
Steps: 35, Sampler: DPM++ 2M Karras, CFG scale: 12, Seed: 2624500747, Size: 704x512, Model hash: 0fc198c490, Model: AbyssOrangeMix2_hard, Hypernet: sxz-3bloom, Hypernet hash: 727e2da6, Hypernet strength: 0.8, Batch size: 3, Batch pos: 0, Clip skip: 2, ENSD: 31337
You may just copy all this stuff and paste it in prompt box, and then just click arrow button in automatic1111, and all parameters will be transfered to suitable boxes.
@sadxzero cool, thanks!
@sadxzero What hardware are you running? I have a good card, but I can't even begin to get it to upscale that high. I feel like I'm missing something.
lmao, nevermind, I'm just an idiot. I left hi-res fix on and thought you were doing the same to generate at like 2850x3680 or something. Then I looked at your image again and saw it was way smaller than that. My bad!