Update README with info about training process.

README.md

@@ -1,39 +1,46 @@
----
-tags:
-- text-to-image
-- flux
-- lora
-- diffusers
-- template:sd-lora
-- fluxgym
-widget:
--   output:
-        url: sample/cboyardee-dilbert-video-style_005100_00_20240917161546.png
-    text: cboyardee_dilbert a cartoon of a man drinking coffee in an office
--   output:
-        url: sample/cboyardee-dilbert-video-style_005100_01_20240917161556.png
-    text: cboyardee_dilbert a cartoon of a happy dog
--   output:
-        url: sample/cboyardee-dilbert-video-style_005100_02_20240917161606.png
-    text: cboyardee_dilbert a cartoon of a mcdonalds restaurant exterior
-base_model: black-forest-labs/FLUX.1-dev
-instance_prompt: cboyardee_dilbert
-license: other
-license_name: flux-1-dev-non-commercial-license
-license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md
----
-
-# cboyardee-dilbert-video-style
-
-A Flux LoRA trained on a local computer with [Fluxgym](https://github.com/cocktailpeanut/fluxgym)
-
-<Gallery />
-
-## Trigger words
-
-You should use `cboyardee_dilbert` to trigger the image generation.
-
-## Download model and use it with ComfyUI, AUTOMATIC1111, SD.Next, Invoke AI, Forge, etc.
-
-Weights for this model are available in Safetensors format.
-
+---
+tags:
+- text-to-image
+- flux
+- lora
+- diffusers
+- template:sd-lora
+- fluxgym
+widget:
+-   output:
+        url: sample/cboyardee-dilbert-video-style_005100_00_20240917161546.png
+    text: cboyardee_dilbert a cartoon of a man drinking coffee in an office
+-   output:
+        url: sample/cboyardee-dilbert-video-style_005100_01_20240917161556.png
+    text: cboyardee_dilbert a cartoon of a happy dog
+-   output:
+        url: sample/cboyardee-dilbert-video-style_005100_02_20240917161606.png
+    text: cboyardee_dilbert a cartoon of a mcdonalds restaurant exterior
+base_model: black-forest-labs/FLUX.1-dev
+instance_prompt: cboyardee_dilbert
+license: other
+license_name: flux-1-dev-non-commercial-license
+license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md
+---
+
+# cboyardee-dilbert-video-style
+
+A Flux LoRA trained on an NVIDIA 4070 Ti Super with [Fluxgym](https://github.com/cocktailpeanut/fluxgym). This is a **style transfer LoRA** designed to replicate the unique MSPaint style of cboyardee's artwork.
+
+Dataset is a collection of 32 frames extracted and cropped from [this video](https://www.youtube.com/watch?v=a6rGUlVvVJ4), tagged with [Florence-2](https://huggingface.co/microsoft/Florence-2-large) with some manual changes. 16 epochs were used with 10 repeat trains per image, resulting in a total step count of 5120.
+
+<Gallery />
+
+## Trigger words
+
+You should use `cboyardee_dilbert` to trigger the image generation.
+
+All tags began with "a cartoon of" but it seems like omitting that prefix still results in good outputs. In the future I want to retag with "MSPaint drawing" as it more accurately describes the style.
+
+**Note that "Dilbert" (or other characters) were not used in tagging, and so the LoRA will not be able to reproduce the character.**
+
+## Download model and use it with ComfyUI, AUTOMATIC1111, SD.Next, Invoke AI, Forge, etc.
+
+Weights for this model are available in Safetensors format.
+
+The [main safetensors file](https://huggingface.co/simpolism/cboyardee-dilbert-video-style/blob/main/cboyardee-dilbert-video-style.safetensors) achieves good outputs, but partial saves are also included with the number of epochs: 4, 8, 12. These files can be ignored unless you want to experiment with fewer-step models.