tungsten-logo
FULLY_REAL_XL

FULLY_REAL_XL

14 stars
1 running now
README

FULLY_REAL_XL

Creator: pAInCREAT0R

It does a great job with photos and portraits. Like most SDXL models, it retains a lot of flexibility for other styles as well. I tend to use more steps than needed, HiresFix, and occasionally Adetailer.

Version Notes: (Newest to Oldest)

  • V10.0 Directly trained the checkpoint for the first time in a few iterations. Training sets of high definition images and people. Trying to obtain the photo realistic results and more natural poses. Trained it 700 epoch at almost 1 mil steps. Then remixed the model with V9 to smooth results.

  • V9.0 More LoRA training. Less tightness of focus. Still struggles with facial expression variety. Probably the last version for awhile unless something comes to surface too amazing to resist. I plan to focus on other areas/models for some time.

  • V8.0 I do not normally push new versions this quickly, but the people had spoken. There was concern over the focus in V7 being too soft. I took the hint (eventually lol) and addressed it. It was time to refresh the bloodline a bit in the model. I think this version improves the focus without giving up the training improvements made along the way. Thank you for those who took time encourage.

  • V7.0 More of a fine tune than a full release. I used some methods I learned creating the Dark Arts Images checkpoint to adjust the contrast and brightness. To me this achieves a more photographic and less generated look in general. Other slight improvements.

  • V6.0 Returning the FULLY_REAL_XL model to published status by uploading my current iteration. It is slightly improved over V5.1. This was a result of merging another of my models I use for LoRA testing at a .25 weight with version 5.1.

  • V5.1 More human like representation (less artificial) and some improved lighting. Also, slightly less push to portrait mode by default. Result of a more detailed and better captioned training set.

  • V4.0 Added an additional trained LoRA of my own, a small amount of about a dozen other LoRA's, and some of a DPO training LoRa. Then I remixed multiple versions of the resulting checkpoints until I hit a version that gave the improved results I was looking for.

  • V3.0 Trained an additional LoRA for a more photogenic look. Merged the LoRA and then remerged the base to the desired effect. Goal was more of a photogenic look and less of a 'rendered' look. Until I get a better dataset and/or get better at checkpoint training, this is probably the last version for awhile.

  • V2.0 I moved from merging to training with this version. I am experimenting with a technique of training LoRA's for an effect and then merging them into the base model. (Instead of directly training the model.) For version 2.0 I created a LoRA based on 282 extremely high resolution photos of various subjects. Then I merged it at a lower weight with the base model to enhance the outputs. I am still learning, but I am happy with the results so far.

  • V1.0 The original version was a massive group of merges of various realism focused models. When I landed on a combination one day that achieved superior results (to my eye anyway), (pAIn)FULLY_REAL_XL was born.

If you try it and like it, please leave some pictures in the gallery.

Bild1.webpBild2.webpBild3.webpBild4.webpBild5.webpBild6.webpBild7.webpBild8.webpBild9.webpBild10.webpBild11.webpBild13.webpBild12.webpBild20.webpBild14.webpBild15.webpBild16.webpBild17.webpBild18.webpBild19.webp

FULLY_REAL_XL

Creator: pAInCREAT0R

It does a great job with photos and portraits. Like most SDXL models, it retains a lot of flexibility for other styles as well. I tend to use more steps than needed, HiresFix, and occasionally Adetailer.

Version Notes: (Newest to Oldest)

  • V10.0 Directly trained the checkpoint for the first time in a few iterations. Training sets of high definition images and people. Trying to obtain the photo realistic results and more natural poses. Trained it 700 epoch at almost 1 mil steps. Then remixed the model with V9 to smooth results.

  • V9.0 More LoRA training. Less tightness of focus. Still struggles with facial expression variety. Probably the last version for awhile unless something comes to surface too amazing to resist. I plan to focus on other areas/models for some time.

  • V8.0 I do not normally push new versions this quickly, but the people had spoken. There was concern over the focus in V7 being too soft. I took the hint (eventually lol) and addressed it. It was time to refresh the bloodline a bit in the model. I think this version improves the focus without giving up the training improvements made along the way. Thank you for those who took time encourage.

  • V7.0 More of a fine tune than a full release. I used some methods I learned creating the Dark Arts Images checkpoint to adjust the contrast and brightness. To me this achieves a more photographic and less generated look in general. Other slight improvements.

  • V6.0 Returning the FULLY_REAL_XL model to published status by uploading my current iteration. It is slightly improved over V5.1. This was a result of merging another of my models I use for LoRA testing at a .25 weight with version 5.1.

  • V5.1 More human like representation (less artificial) and some improved lighting. Also, slightly less push to portrait mode by default. Result of a more detailed and better captioned training set.

  • V4.0 Added an additional trained LoRA of my own, a small amount of about a dozen other LoRA's, and some of a DPO training LoRa. Then I remixed multiple versions of the resulting checkpoints until I hit a version that gave the improved results I was looking for.

  • V3.0 Trained an additional LoRA for a more photogenic look. Merged the LoRA and then remerged the base to the desired effect. Goal was more of a photogenic look and less of a 'rendered' look. Until I get a better dataset and/or get better at checkpoint training, this is probably the last version for awhile.

  • V2.0 I moved from merging to training with this version. I am experimenting with a technique of training LoRA's for an effect and then merging them into the base model. (Instead of directly training the model.) For version 2.0 I created a LoRA based on 282 extremely high resolution photos of various subjects. Then I merged it at a lower weight with the base model to enhance the outputs. I am still learning, but I am happy with the results so far.

  • V1.0 The original version was a massive group of merges of various realism focused models. When I landed on a combination one day that achieved superior results (to my eye anyway), (pAIn)FULLY_REAL_XL was born.

If you try it and like it, please leave some pictures in the gallery.

Bild1.webpBild2.webpBild3.webpBild4.webpBild5.webpBild6.webpBild7.webpBild8.webpBild9.webpBild10.webpBild11.webpBild13.webpBild12.webpBild20.webpBild14.webpBild15.webpBild16.webpBild17.webpBild18.webpBild19.webp
Actions
RUN
About
FULLY_REAL_XL
Created February 2024
sparks/fully-real-xl
Popularity
14 stars
1.9K runs
0 pulls
Maintainers
sparks