The next iteration in the LimitlessVision evolution. Excels in photorealism, analog, and cinematic. Will give accurate representations of both male and female anatomy and those between.
Description
FAQ
Comments (27)
Hey! Any chance you could upload an FP16 version of this model? Can't run BF16 models on mac, unfortunately.
I can try, I apologize it is not compatible. You could try it anyway. I labeled it bf16 all my dataset training was done bf16. Not sure that makes the model completely bf16? Please let me know and I will look into converting if possible.
Excellent, was looking forward to this! Does this require a VAE or is there one already baked in?
Vae is not baked in, I wanted to leave that option open to users. I use the basic SDXL VAE
can't get a decent penis
Penises can be tricky, but the more simple you prompt the better usually. Also negative prompt can mess it up as well.
Remove hand related prompts from the negatives
Try the new version
HHhmmm... could be my new favorite checkpoint!
One question: I am getting often vaginas when I render men. How can I avoid that?
Hi Silas, thank you so much for sharing the SDXL new model. I’ve given it a try, but unfortunately, I haven’t been as successful as I hoped. The images I’m generating are not as good as the ones I got from your SD1.5 model, which produced excellent results. The images I’m getting now are somewhat blurry and not as visually appealing. You can see an example in the link below. With prompt 'Vietnamese, cute face, gay, lying on leaves autumn theme, (penis), photography ultra realistic, masterpiece, best quality, highres, cinematic film grain, iso'
https://i.postimg.cc/h4r89hyK/2023-11-23-17-48-17-50.png
With your SD1.5 model using similar setting and same prompt I get something like this
https://i.postimg.cc/hthNkkz7/2023-11-15-21-39-28-40.png
I suspect there might be an issue with my Python code. Below is a snippet of it. I’ve tried using 50 iterations and even 100, but I haven’t been able to produce a single satisfactory image compared to your SD1.5 model. I also attempted to use the base XL VAE, but it didn’t make a difference. I’ve experimented with different image sizes as well, but to no avail. It’s quite surprising because the images you’ve shared are wonderful. So, I’m inclined to think that I might be doing something wrong…
from diffusers import StableDiffusionXLPipeline,StableDiffusionPipeline, DPMSolverMultistepScheduler
import helpers
MODEL_PATH = 'limitlessvisionxl_v1.safetensors'
PROMPT_PATH = '/home/me/Pictures/cnGuy/prompt.txt'
NEG_PATH = '/home/me/Pictures/cnGuy/neg.txt'
PATH = '/home/me/Pictures/cnGuy/' #path where image will be saved
SIZE1= 768 #script will randomly use as height or with
SIZE2 = 768#1024 script will randomly use as height or with
STEPS = 50 #number of step (20 min recommended with new DPMSolver algo)
pipeline = StableDiffusionXLPipeline.from_single_file(MODEL_PATH, torch_dtype=torch.float16)
pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
pipeline.safety_checker = None #to aovid black sensored image if NSFW
pipeline.requires_safety_checker = False #to aovid black sensored image if NSFW
pipeline = pipeline.to("cuda") #Nvida gpu mandatory with cuda installed
for prompt in PROMPT_LIST:
output = pipeline(prompt,
height=SIZE1,
width=SIZE2,
negative_prompt=NEG,
num_inference_steps=STEPS)
image = output.images[0]
Firstly, LVXL is not nearly as developed as it's 1.5 counterpart. The reasoning comes down to time and effort. I can train 10 datasets for a 1.5 model in the time it takes to train 1 on SDXL. Second, prompting is not the same for SDXL as it is for 1.5. Identical prompts and settings will always yield different results (not always worse), but different. LVXL is a baby in development as of now. I am continuing to work on it daily to be as flexible and accurate as it's the 1.5 model. Long story short, it is not a fair comparison yet. I appreciate your thorough input. I will look at the info you supplied and see if there are hotfixes to implement.
Dear Silas,
Thank you for your response. I understand that the maturity level of the models is not the same yet. I appreciate your hard work and will definitely try your model again in the future. If you need any assistance, please let me know. Although I am not an expert, I have my own Nvidia1080ti (11gb) so I can try to help if needed.
Best regards.
what image sizes did you use?
@rime11 not sure if the question is directed to me, but in training, I go by a loose rule of MINIMUM 1024x1024. If the native image resolution is less than that on either width or height, I do not use it ( yes, I do not upscale or modify training images). At least 50% of the images are 4k format with cinematic aspect ratio. I do have a good chunk of portrait and square, but SDXL already does pretty well at those aspects. I intentionally do not stick to a specific size, yes that means my buckets are huge in training, but I feel that it helps give more stability in allowing users to select their own preferred sizes and aspects when using.
Can we get even fatter hirsute supersized men in next version please, this model is nearly perfecto for my needs.
lol, I will make sure to get some more hefties in the next!
please keep working on this, right now it appears to be the best SDXL checkpoint at 3rd-person male genitalia, but even then, it's not super great at it. But still, that's a niche only this checkpoint occupies right now, I would run with that
Thank you for your input! Do not worry, this project will see more evolutions. I agree at the not super great aspect. I am working on it being as flexible as it's 1.5 predecessor. Sadly, developing SDXL is more resource consuming than 1.5. Hopefully the model gains a bit more popularity in order to help justify the commitment.
@SilasAI6609 the situation actually changed within the past 2 days since i made that comment, and now there's a homosexuality-focused checkpoint called Himeros which does it better. But still, that's only 2. just so you're aware.
@shapeshifter83 Do you think Himeros is better overall? It does penises a lot better, but otherwise it has problems with faces and similar, or at least I'm doing something wrong. This model has been much more consistent when it comes to bodies/faces/etc, it's actually still my preferred model for naked males. Hopefully it gets a worthy successor, this one is an amazing job.
@jtabox idk, there's also been some LoRAs that have come out for SDXL in the past month that have allowed me to use whatever checkpoint i want, so i haven't used much of either after my last comment
@SilasAI6609 I love this checkpoint because you don't have to use a lora to create a penis. Please keep working on it and improving it. I hope you keep making newer versions.
Hint: penis shows up if negative prompts are removed & you specifically prompt for it (this was buried in comments). Great work! THANK YOU for this model!
Thank you for you review and input. I am also working diligently on the next model versions. I have a large number of projects going on at the moment, but LimitlessVision and LVXL are high on priority!
this is giving me a lot of mutations, extra body parts, piles of body parts, phantom limbs, and genitalia switcharoo.
Thank you for your feedback. Have you tried using the supplied prompts in the examples?