Version 4 is now live.
*Images are shown without any LORA's or EMBEDDINGS unless stated in their details.
This model is a custom blend of various models, (see below), giving many options for generating a wide variety of images, including NSFW. Based on Stable Diffusion 1.5.
Primarily focused on capturing skin texture and the feel of photography, and Polaroids in general, but experiment with other things. (This model is not perfect).
Though there are is no specific KEYWORD to 'turn it on', if you use words in the prompt such as 'polaroid' , 'analog' , 'film grain', 'perfect eyes' and so forth at the start of the prompt, you should get the look you're after.
The example pictures do not have any face/eye correction or upscaling. 768 x 1024px is quite stable at producing images without 'mutation' if the prompt and neg prompt are good.
The model needs a VAE. (vae-ft-mse-840000-ema-pruned is fine or you can use your own.)
Prompt example:
(selfie)+++, (bokeh)++, RAW photo, (intricate detailed eyes), alone, (shot from the side)+, close up portrait, cheeky smile, (Irish blonde woman), catholic tattoos, thicc, leaning against wall, (wearing undersized vest)++, (looking at the camera), intricate details, skin imperfections, halo of light, midday sunlight streaks spilling over face, face details, unbelievable intricate details, dark shadows, (graffiti covered ruins), real lighting, bloom, volumetric lighting, cinematic lights (ultra skin texture), (symmetrical eyes), light beams streaming through haze
Negative:
Child, Bad eyes, Overexposed, (semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, digital art, anime, painting, manga), (poorly drawn hands, poorly drawn face), (deformed iris), (deformed pupils), grit, man, distorted, male, skin rash, deformed body, mutated hands, poorly drawn eyes, head crop, smooth skin , deformed hands, deformed proportions , long face, long neck, looking away, puffy nipples, watermark, logo, text, multiple people, jewellery, Plastic skin, Overexposed, (semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, digital art, anime, painting, manga), (deformed iris), distorted body, (deformed pupils), grit, man, distorted, (extra arms), long body, male, skin rash, deformed body, mutated hands, poorly drawn eyes, head crop, smooth skin , deformed hands, (deformed proportions) , long face, long neck, looking away, puffy nipples, watermark, logo, text,
A non-exhaustive lost of models merged (because I didn't take notes): deliberate, clarity, dreamlike-photoreal, realistic vision, protogen (+ additional ones in version notes).
Made for fun, not for money or notoriety. Those who train the models from scratch are the ones deserving of applause.
Have fun! Can't wait to see what you come up with...
-
DEPRECIATED INFO:
(Version 3.1 is a hotfix, as the prune for v3 failed and was showing errors on peoples pictures.
(It is very dependant on negative prompts, so use the example images as a base.) The first image uses a light map as img2img so your results may very, but i thought it looked cool so i included it.
The model is better with a VAE. (vae-ft-mse-840000-ema-pruned is fine or you can use your own.)
* For some reason (for V2) the pruned version isn't showing as the first download and the hashes have got screwy. Please use the dropdown menu to choose the smaller version if desired *)
Description
From my testing and personal preference, this version has better eyes and more realistic skin. Less film grain and better lighting can be achieved without lora's. Its still not perfect, but i think its better than the previous version.
In addition to the base 1.5 model and those used in v1 and 2, the addition models merged include, endless, cyberrealistic and noise offsetlora (at a low level).
FAQ
Comments (6)
V3 tested and its a great realistic model. the easy to catch skin textures (my passion ..) is absolute on point.
My (starter) experience tipps: I copied the negative prompts from on of your reference postings (thats start with nose rings), that worked on my prompts better than the in your readme declared negative prompt. I get better results with 512 x 704 than 512 x 768 (which i normaly use, but it often was cropped ..) Great model. God Job. Thank you.
v3.1 is out now and will hopefully solve many issues people were having. apparently my testing was not robust enough and v3 suffered from a bad prune.
Been testing out V3 (as you know V2 is my go-to model) and it's great so far. The only odd thing I've noticed is that V3 tends to wants to show two people at 512x768 -- something I never recall seeing in V2 after generating thousands of images. Here's an example:
https://scottymac.s3.amazonaws.com/stable-diffusion/screenshots/avalon_comparison.png
hi, ive discovered that in my haste after the v2 hash to get v3 up that the prune was screwed up from the full version. the eyes arnt right either. and there may be other issues, like your example. basically, im not happy either and will be uploading v3.1 shortly
3.1 is up. hopefully that fixes everything. (i normally put 'alone' in the prompt, as it helps with multiple people, especially in landscape wide shots.)
The contrast on the faces looks really nice in V3 and 3.1. I kinda love that the last photo goes 1 → 2 → 3 people: https://scottymac.s3.amazonaws.com/stable-diffusion/screenshots/avalon_comparison_v1_2_3.png
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.