huehuehue9001 Posted April 15 Posted April 15 (edited) I am braindead. Can't think anything creative. Have a slop. Youtube recommended me a gothgirl. Now it's full of them. Spoiler Edited April 22 by huehuehue9001 9
huehuehue9001 Posted April 22 Posted April 22 (edited) Spoiler Some anon requested for non twin braids. The one above is too muscular. My mistake was I prompted the muscle in the base image. It's better to prompt muscle in the upscale not in the base. You get the svelte, slender outlines and then add in the muscle tone. I'm not sure what car it is, but I think it's a Mercedes. Default tanlines are bigger. To get them small you have to inpaint and it needed a lot of help. It's better to start with no tanlines. I did some Z image test. Model is Moody Wild Mix on civitai.red. Faces inpainted with illustrious. No DOAXVV lora yet. Spoiler Qwen 2512 upscale. z image + ultimate sd upscale. Started with z-image from the beginning. Only face is swapped as the last step. I am using moody wild mix and the provided workflow, ZIB + ZIT. The ultimate sd upscale part introduced a lot of errors, so I gave up on that. The hands and feet will still get errors. It took 12 gens to get this image. I didn't want to inpaint and showcase what z image can do. The nipples and genitals are also a hit and miss. You will need to prompt the details you want. If lazy, use gemini or qwen LLM. It's progress and looks better than base model. Moody mix doesn't do anime or CGI. I tried to prompt but it ignored. In my opinion, if you want the cgi look, qwen is better and sharper. If you want to create real life looking from scratch, z image is really good. The NSFW is more advanced too. You need to use ultimate upscale sd for z image. if you just upscale 4k image with z image, it will be blurry. qwen just works. Spoiler I don't know why it gave a lace swimsuit, but lets just roll with it. Maybe because of the lora with orchid swimsuit. Mai confirmed in DOA6LR. If the character is sideways, the AI really struggles with faces. So I had to go in the gimp rotate 90 degrees. anime >make it as 3dpd as possible. the ocean and balcony looks fake though. i had trouble with this one. things didn't look good. Spoiler No good idea. Have a slop. Ayane's VA died. https://news.livedoor.com/article/detail/31276342/ Spoiler Spoiler oops Spoiler forgot to prompt the eyes. but is her eyes blue or grey? it's not this blue in xvv. //edit, when you tilt the camera and look from below you can see her eyes tint blue. the prompt is (blue grey eyes) for illustrious. model will vary. qwen can't do underwater. it's too clean. i have water caustics in the base image, but it got deleted. I can't get AI to do a hanging on cliff pose. I got one, but ended up giving up, the pose still looked weird. I tried some stick figures, but I can't even draw sticks. It came out incoherent. I gave up and did something simple. It wasn't you owner-kun, it was me Lei Fang. ok. I figured out how to do the weird angle faces. Just give up on it looking the same as the game. drop lora for face swap all the way to 0.5 and let the model handle it. low denoise at 0.35. the base image looking good helps. Edited June 1 by huehuehue9001 14
huehuehue9001 Posted June 1 Posted June 1 (edited) Previous post is full. Anyway, I heard Nvidia PID can upscale to 4096x4096. I tried it. It's not the magic bullet like all the demo showed. Spoiler Ideogram V4 is good at region prompting. Spoiler Spoiler The idea was >Waitress >Hungry >Exhausted >Sleep >Yuuki Rito Fully clothed is easy. Can you guess where this is going? Spoiler Halt! Criminal scum! The face is a bit different. This one was a lot of redo. Nudes with poses are hard. You can pass off clothes as random weird fashion but you can recognize body deformations easily. I wanted to make crotch visible, shined by the water reflection, but nothing good came up and I'm giving up. The likeness is very close. So I went with it. This is the qwen upscale. Spoiler Surprise! No good idea. Have a slop. I learned a new word, "mechabare". It didn't do anything much though. The half mecha face really didn't want to come up. I didn't download a lora but just inpaint half the face and use high denoise = 0.69. It has to go inpainting twice, first ask for "metal skin" to remove the human skin, then the second round you'll get better mecha face. Too high denoise the first time you'll just get disproportionate face. The base image is flux2. Illustrious can't do diving poses by prompt. Flux2 is still whacky but with the distilled version I can gen at 9seconds per picture. So I can brute force it and get a nice one. It has variety to the image even with the distilled version and you can prompt it to give 3d cgi style art. better than z image. So it's, flux2 t2i > illustrious i2i > illustrious upscale > qwen > illustrious inpainting. They say the devil is in the details. There are weird parts but I didn't want to spend too much time on it. It's shark week. Spoiler The gold pieces are very random. This is with Dea Marina lora. Keeping with the farming theme... Anyway, it's flux2 base image for the pose > inpainting > qwen i2i because inpainting is still broken > inpainting again > upscale > qwen > inpainting. If you say, just draw it from scratch. I would say I can't. For flux2 base. Only 2 out of 64 gen is useful. I think I'll try anima next time. moo Spoiler Massive dump. It's just slops. Edited June 22 by huehuehue9001 12
huehuehue9001 Posted June 24 Posted June 24 (edited) Spoiler When it's a ballroom dress, it has to be Urushihara Satoshi fashion. The face looks derpy here. Idea was looking in the mirror, while dressing up, but it looks too clear. Not sure who to be maid, settled with Misaki. She's a jobber character. Wanted to make her look modest, but AI kept giving her sexy maid outfit. Now she looks better than the boss. Never outshine the boss. She's fired. the writer's block is real. i think i need a break. update. well if i can't do it. why not just let AI write the prompt. I used gemini before but it's censored. gemma4 have uncensored version. it's open source. gemma4 e4b is ok. but it can be dumb at times and make things up. gemma4 26b a4b is noticeably better. but it's big. i'm using a 4bit quantized version to fit in 16gb of vram. even highly compressed it's still better. i haven't tried qwen. the new 3.6 is all very big. Edited yesterday at 10:44 AM by huehuehue9001 9
5250 Posted yesterday at 04:47 AM Posted yesterday at 04:47 AM 15 hours ago, 割れMOD最強 said: This image is really good. Would you tell me which checkpoint are you using?
Recommended Posts
Create an account or sign in to comment
You need to be a member in order to leave a comment
Create an account
Sign up for a new account in our community. It's easy!
Register a new accountSign in
Already have an account? Sign in here.
Sign In Now