adding special feature for micro training

MICROMODELS.MD

@@ -0,0 +1,37 @@
+# Micro models
+
+Micro models allow training quickly on a small data set in minimal time.  The technique is simply training on the new data with no preservation.  If you add too much data or train for too long you will damage the model notably, but it is an easy and quick way to get one character into a model.
+
+To train a micro model, only place one subfolder of 20-50 images into a training root folder
+
+Example:
+
+    /training_samples/ted_bennett/ted_bennett
+    /training_samples/ted_bennett/ted_bennett/a close up of ted bennett in a suit and tie.jpg
+    etc
+
+Training root in this example will be "/training_samples/ted_bennett" and the extra subfolder is required but can be named anything.
+
+## Example
+
+You can recreate the above model by downloading the dataset or try your own.
+
+Dataset:
+
+    https://huggingface.co/panopstor/ff7r-stable-diffusion/blob/main/ted_bennett.zip
+
+Model:
+
+    https://huggingface.co/panopstor/ff7r-stable-diffusion/blob/main/ted_bennett_259step_13minutes.ckpt
+
+Unzip the file into your /training_samples folder.  It will create a subfolder on its own to work with the trainer.
+
+On the command line, use v1-finetune_micro.yaml which is setup for this case.  You may wish to slightly change the REPEATS for your data set, with 31 images and repeasts 50 the output is fairly good and completes in just 13 minutes on an RTX 3090:
+
+    python main.py --base configs/stable-diffusion/v1-finetune_everydream.yaml -t  --actual_resume v1-5-micro.ckpt -n ted --gpus 0, --data_root training_samples\ted_bennett
+
+Prompt is simply "ted bennett"
+
+![ted bennett as ironman](./demo/ted_bennett_ballet.png)
+![ted bennett as ironman](./demo/ted_bennett_ironman.png)
+![ted bennett as ironman](./demo/ted_bannett_superman.png)

README-FF7R.MD

@@ -0,0 +1,46 @@
+This is a finetuning of the RunwayML 1.5 ckpt. 
+
+Info: https://huggingface.co/panopstor/ff7r-stable-diffusion
+
+Model: https://huggingface.co/panopstor/ff7r-stable-diffusion/blob/main/ff7r-v5-1.ckpt
+
+
+This fine tuning includes over a dozen concepts trained in 1636 images with individual captions on each image, plus 1636 images scraped from the Laion dataset for model preservation. 
+
+See the Hugging Face link for more info and a downloadable model you can try.
+
+Everything should be fairly natural to prompt with the exception of Biggs and Wedge which were trained with "ff7r" as it if were their surname.
+
+Prompt guide:
+
+    cloud strife
+    tifa lockhart
+    barret wallace
+    aerith gainsborough
+    biggs ff7r
+    wedge ff7r
+    in midgar city
+    slums district of midgar city
+    business district of midgar city
+    train station in midgar city (optionally in a district name from above)
+    streets of midgar city (optional in a district name from above)
+    shinra security officer holding an assault rifle
+    rooftops of the midgar slums district
+    food truck in the slums district of midgar
+
+You can mix and match all the above!  Try it out! 
+
+Preservation tests.
+
+1.4 is far left, 1.5 far right.
+
+FF7R 4.1 was trained on Kane Wallmann's repo with 1400 images training and 1400 images of regulation generated out of SD itself, similar but not quite equal to dreambooth technique.
+
+FF7R 5.1 was trained on EveryDream with 1636 images training and 1636 images of Laion ground truth without DreamBooth techniques.
+
+(warning huge file!)
+
+![a](./demo/mega_test01.webp)
+
+New characters tests:
+![a](./demo/mega_test_characters01.webp)