Sign In

ThisIsReal

2.9k
29k
284
Verified:
SafeTensor
Type
Checkpoint Merge
Stats
2,338
Reviews
Published
Jan 19, 2024
Base Model
SD 1.5
Hash
AutoV2
B3DF0C8FF2

Pruned Version 6 is now available

Use this VAE : vae-ft-mse-840000-ema-pruned

Version 4

  • More details in the face

  • More realistic representation of people

  • Better incidence of light on the body

  • better representation of the hands

  • 2000 photos of people

  • 500 photos of faces

  • trained size 768x768, 1024x1024

Version 2/3

  • more face details

  • more body details

  • better incidence of light

  • better representation of multiple people

  • trained size 768x768, 1024x1024

When creating the model this time more attention was paid to the aesthetics as well as multiple people, and that it looks even more real

Version 1

  • First, I created an extensive base by reading in over 5000 photos of people of various nationalities, both clothed and nude.

  • Then the first model, AmiReal, was used to generate images.

  • After determining that the shadow cast was not correct when generating the people, I used over 500 images to refine the shadow cast on bodies and objects.

  • Then I used the next model, CyberRealistic, and combined the two models.

  • Unfortunately, this combination resulted in either strange faces or the same faces over and over again.

  • So I trained the model again and took over 3000 close-up shots of faces until the result was acceptable.

  • For my last combination, I decided to use the realisticVision and realhotMix models.

  • Then I tried again to refine with 500 images only hands and feet.