I love the way it matched the color scheme on this. Sometimes when I use NovelAI, I feel like I'm 'playing dress up' with the character but it's just so much fun.
Firstpass 0x0 just means it's automatic and that's what I'm using too. I have one weird idea left. When mixing, set model A as trinart2_step115000, B as nai-animefull-final-pruned and multiplier to 0.8. You can also check "save as float16" even though I honestly doubt it changes much.
It's "better anatomy" but yes that's the one. I mixed it 50% with r34_e4 and then 0.65 of the result with trinart. Finally I did nai * 0.7 + the_previous_mix * 0.3. Not sure if it changes much though, I was just messing around. Results are very similar to a nai * 0.8 + trinart * 0.2.
Thank you very much. I hope the drawing of ai will be more abundant
Is the f111 model used in this picture the same as 'Zeifer F111 Female Nude (better animation)? Thank you.
It's "better anatomy" but yes that's the one. I mixed it 50% with r34_e4 and then 0.65 of the result with trinart. Finally I did nai * 0.7 + the_previous_mix * 0.3. Not sure if it changes much though, I was just messing around. Results are very similar to a nai * 0.8 + trinart * 0.2.
realistic, masterpiece, high quality, highres, cute afraid loli goblin goat slave with wavy hair, cleavage cutout, small breasts, detailed face and eyes, (frills:1.2) (solo focus:1.2) pov, aroused, embarrassed, antlers, by Jeremy Lipking, by Antonio J. Manzanedo, by (Alphonse Mucha) digital painting Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, out of focus, censorship, ugly, old, deformed, amateur drawing, odd, fat, adult
Steps: 20, Sampler: DDIM, CFG scale: 14, Seed: 1404394528, Size: 512x960, Model hash: 45a4bfe3, Denoising strength: 0.6, First pass size: 0x0 highres fix Upscale latent space image when doing hires. fix Stop At last layers of CLIP model:1
Still got a different image - -! Maybe "First pass size: 0x0" cause?
Firstpass 0x0 just means it's automatic and that's what I'm using too. I have one weird idea left. When mixing, set model A as trinart2_step115000, B as nai-animefull-final-pruned and multiplier to 0.8. You can also check "save as float16" even though I honestly doubt it changes much.
Must be a change in how exactly the models are merged but it shouldn't matter anyway. I did a merge on the current version of the ui and got the same hash as yours [45a4bfe3]. It still generates the same image. I used highres fix with latent space upscale and CLIP:1 so maybe it's one of these settings?
My parameters:
realistic, masterpiece, high quality, highres, cute afraid loli goblin goat slave with wavy hair, cleavage cutout, small breasts, detailed face and eyes, (frills:1.2) (solo focus:1.2) pov, aroused, embarrassed, antlers, by Jeremy Lipking, by Antonio J. Manzanedo, by (Alphonse Mucha) digital painting Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, out of focus, censorship, ugly, old, deformed, amateur drawing, odd, fat, adult
Steps: 20, Sampler: DDIM, CFG scale: 14, Seed: 1404394528, Size: 512x960, Model hash: 45a4bfe3, Denoising strength: 0.6, First pass size: 0x0 highres fix Upscale latent space image when doing hires. fix Stop At last layers of CLIP model:1
Still got a different image - -! Maybe "First pass size: 0x0" cause?
This picture is so awesome, I want to copy it but the result is different Where you download model and What's your novelai and trinart model hash? I download animefull-final-pruned.ckpt[925997e9] and trinart2_step1150000.ckpt[f1c7e952],then merge them in stable-diffustion-webui,result animefull-final-pruned_0.8-trinart2_step115000_0.2-Weighted_sum-merged.ckpt[45a4bfe3] which not same with your model hash d0d8a399
Must be a change in how exactly the models are merged but it shouldn't matter anyway. I did a merge on the current version of the ui and got the same hash as yours [45a4bfe3]. It still generates the same image. I used highres fix with latent space upscale and CLIP:1 so maybe it's one of these settings?
This picture is so awesome, I want to copy it but the result is different Where you download model and What's your novelai and trinart model hash? I download animefull-final-pruned.ckpt[925997e9] and trinart2_step1150000.ckpt[f1c7e952],then merge them in stable-diffustion-webui,result animefull-final-pruned_0.8-trinart2_step115000_0.2-Weighted_sum-merged.ckpt[45a4bfe3] which not same with your model hash d0d8a399
Either my shit is bootleg or I'm doing something wrong. I can't seem to replicate 95% of this work. Or should I even be able to? What about the rest of you?
Where can I download this model? > nai-aminefull-final-pruned_0.8-trinart2_step115000_0.2-Weighted_Sum-merged Your art is amazing!
- get trinart and novelai from the links in the OP of any of the diffusion threads on /g/ or other boards - merge them (automatic1111's webui has a nice tool for this)
Forgot to add to the description, but found it interesting: I got the 'spoken hearts' by combining 'dirty talk' and 'heart' and it made the little hearts appear in speech bubbles. I added dirty talk to see if it would change her expression (which it also did)
So to better understand what am I doing in the program, I, out of curiousty, decided to use previous prompt but utilize custom resolution that's 1.5 larger than standard. And because of NN black box result differs much more then if I take a random tag and add {{}} (do note that if I would add :**number** to exact same tag results would also shift randomly)
1girl, (((cum))), brown hair, black eyes, nipples, small breasts, upper body, solo focus, bangs, bukakke, facial, orgy, (((cum on body))), open mouth, multiple boys, multiple penises, happy, smile, cum drip, cumpool
I cant get the same result, what am i missing? I'm already merging full pruned model 0.2 with the same trinart model 0.8 but still cant get the result, and the arstyle. Closer artstyle is when merging model around 0.3 instead of 0.2.
After a lot of tweak finally i get the the closer result with model NAI 0.75 + Trainart2 0.25 hash e7c39565 Can show me how you upscale this picture? Im new with upscaler so i can't get better result
For upscaling I used the same exact model, prompt, negative prompt and CFG, but with DMP2 a Karras at 60 steps, denoising: 0.2, upscaler: ESGRAN_4x. Not sure if this is optimal, but it gives me reasonable results.
