if you like the models, please consider supporting me on, i will continue to upload more cool stuffs in the future
http://patreon.com/PromptSharingSamaritan
https://ko-fi.com/promptsharingsamaritan
Description
This one is way better than SDXL model of previous version. The model was trained using 768x768 images so the minimum resolution should be 768x768
Recommended prompt:
Close up photo of <....>
Negative prompt: cartoon, painting, illustration, (worst quality, low quality, normal quality:2)
I use cfg scale 3 vae-ft-mse-840000-ema-pruned VAE. Resolution is 768 x 1152
to avoid getting the same overtrained face, try to use random name in the prompt and remove the word 'woman" in the prompt.Sometimes it works
if you like the models,please consider supporting me on, i will continue to upload more cool stuffs in the future
http://patreon.com/PromptSharingSamaritan
https://ko-fi.com/promptsharingsamaritan
FAQ
Comments (41)
Looks great in terms of quality, however, judging by the example images it seems to be plagued with the same face issue as the other model. Any workaround?
try using random woman's name in the prompt and remove the word 'woman" in the prompt
Dynamic prompts and CADS extensions can help with missing variability
does it do nsfw? nipples, genitals, sex poses?
This model was a nice surprise. Absolutely fantastic for embeddings and loras. Great work!
Very good results. Thanks for your Model!
This is my actual main go-to. Love the results!
@PromptSharingSamaritan the detail and realism in the new 1.5 model examples images is incredible! did you make an improvement or is 1.5 just better at this? interested in hearing more about the details of the model - dataset, steps, parameters, etc, if you would be willing to share. thanks for the model!
yes but it has the same face problem, i use 61 steps using one trainer
I keep getting Getty Images watermarks on my results even after I put logo, watermark, text on the negative prompt. Any solution for it?
What prompt are you using? I haven't seen something like that yet.
This model, among other photorealistic models, looks completely unusual and original. It probably trained for the most part in close-up photographs and tends to do exactly that in the generations. I really liked the fresh, unusual from other models faces , the excellent expressiveness of the portraits, the cool natural skin texture, the studio composition and lighting.
Yes, this model is very peculiar and not very responsive, inclined more towards a female face, but it is definitely worth paying attention to. Very original cool work!
is there a way to run this model on Google Colab? Best regards
i don't know how to google colab, maybe someone else can help
https://github.com/arenasys/qDiffusion has a backend that can run on Collab
you can use fooocus there you import this model and will run perfectly, everything will be present in the GitHub repo.
super cool
pls post it without XL
I've been testing the model for a few days now, and it's really good if you need photorealism, just natural. The skin, lighting, etc. looks really good! This is exactly the model I was looking for!
Google says that SDXL is better than SD1.5, so why the change ? It means that Fooocus users only get to use it as a Refiner.
1. Don't trust Google, make your own decisions.
2. There are XL versions to download, just click the version at the top.
@Madsen1981 I only see v3.0 (SD1.5), v2.0 (the version that comes with Fooocus) and v1.0, listed at the top
@JohnnyWu22 If you click on the version you can see in the right column what their Base Model is. V.2 and V.1 is built on SDXL.
@Madsen1981 I already have v2, I was hoping to get an SDXL version of v3.0. Is there one ?
@JohnnyWu22 I would think it's coming soon maybe. XL takes a bit longer to train.
Having trouble adding this as a Lora to my replicate model. Keep getting "invalid lora checkpoint". Link is download link + API key. Help! (I'm not a programmer, just a rookie)
I may have found a way to remove sameface from this checkpoint without losing too much else: refiner. I used Realistic Vision as the only other 1.5 checkpoint I know that claims to handle 768x1152, switching at 0.2 - 0.3. The result retains most of the composition and details from the base image. Slightly softer or blurrier though. Or less oversharp, if you prefer. See post.
Face ADetailer brings back the overtrained face. I guess it runs on the base checkpoint, not sure if it can be switched to the refiner.
I believe using a more extensive neg prompt could help with the refiner part. More testing is needed in any case.
can always inpaint or use a lora of a character or person
Wouldnt IPAdapter / face ID be able to fix the face problems?
@dadadies yes, then you would need to use the face ip adapter to replace the face.
please continues your work
i'm waiting for the next big model
illustriousXL version pls
Why is your 3.0 version 1.5?
This model is pretty useless by itself - no NSFW, no styles. But combined with other Clip it surprisingly capable to do pretty good looking art images.
it's pretty interesting for non nsfw photos
for some reason it always give small dicks to girls tho
What specific prompt serves to generate realism in this model?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.