It is one of my first models that I trained thanks to this guide:
[Guide] Make your own Loras, easy and free | Stable Diffusion Other | Civitai
There are LoRas and Checkpoints that reference the VHS style, and after this inspiration, I came up with the idea of being able to create a LoCon of a VHS stylized taking references to interviews, television programs, movies and commercials from the late 80s and mid 90s to create an effect not so destroyed or glitched by the quality of the VHS of that time. It was intended for portraits more than for other purposes, for landscapes and others, I am not 100% sure that it works well.
It was trained by approximately 140 portrait-type photos extracted from video fragments and other image sources.
In the training I used this checkpoint with SD 1.5 as a base, so I also recommend that you use this checkpoint in conjunction with the LoCon to obtain the most faithful result to the one I had.
(Surely there will be better combinations, better options, but that I have not tested or documented.)
How to install LoCon:
How to use it:
Prompts: 90VHS, vhs still from tv, + whatever-you-want
I also recommend using epiNoiseoffset_v1.
Negative Prompts: rz-neg-foranalogportrait, cartoon, 3d, ((disfigured)), ((bad art)), ((deformed)), ((poorly drawn)), ((extra limbs)), ((close up)), ((b&w)), weird colors, blurry
You can experiment with negative prompts, but for me ‘rz-neg-foranalogportrait’ give very good results.
I recommend these parameters:
LoCon Strength: 1.0
Size: 512 x 512 - Truly I have not tried with other resolutions, I think the most stable is this measure.
Sampling Steps: between 20 and 40 is the most stable.
Sampling Method: DPM++ SDE Karras works very well for portraits.
Using this LoCon in img2img interesting results are achieved maintaining 512x512, and with these settings:
Same settings as txt2img but with these differences:
CFG Scale: 26 ~
Denoising strength: 0,5 - sometimes 0,8
I invite you to experiment with checkpoints, mix them with other Loras and share them!
Don’t forget to put the prompts you used, thank you!
Description
How to use it:
I recommend this Providence - 2.1 - 1.2 - Anchor | Stable Diffusion Checkpoint | Civitai
because it was trained with it on 768x768
Works better with complex prompts.
Prompts: 90VHS, vhs still from tv, vhs from documentary, vhs commercial, ((yourprompt on VHS style image like 80s)), ((high quality:1.2, masterpiece:1.2)),high details, detailed and intricate, intricate details, high intricate details, absurd amount of details,((upper body:1.2)), portrait, Realism, Ambient Occlusion, Natural Lighting,Volumetric Lighting, Volumetric Light, Volumetric
Negative Prompts: rz-neg-foranalogportrait, Neg_Facelift768, cartoon, 3d, asian, ((disfigured)), ((bad art)), ((deformed)), ((poorly drawn)), ((extra limbs)), ((close up)), ((b&w)), weird colors, blurry
I recommend these parameters:
LoCon Strength: 1.0
Size: 768 x 768 - Truly I have not tried with other resolutions, I think the most stable is this measure.
Sampling Steps: between 20 and 40 is the most stable.
Sampling Method: DPM++ SDE Karras works very well for portraits.
Using this LoCon in img2img interesting results are achieved maintaining 512x512, and with these settings:
Same settings as txt2img but with these differences:
CFG Scale: 26 ~
Denoising strength: 0,5 - sometimes 0,8
FAQ
Comments (9)
RuntimeError: The size of tensor a (768) must match the size of tensor b (1024) at non-singleton dimension 1
nvm im rarted
used the 2.1 version instead of 1.5 by mistake
@user8923747 no drama!
I'm having the same issue with the new version? 🤔
@squidboimusic maybe you are using the new version (SD 2.1) with a SD 1.5 checkpoint, try it with another one!
@Rolfoso I'll give that a try, and let you know how it goes! 👍
@Rolfoso Actually I think all my models are 1.5 versions. I dunno why, but I'm still getting an odd error. Maybe I will reinstall the LoCon extension? (This is the first time I have used a Lycois model.)
It would have been the World Trade Towers in the 90's though. :o
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.