Update README.md

README.md

@@ -2,22 +2,22 @@
 license: openrail++
 ---
 
-# Terminus XL Gamma (v2 preview)
+# Terminus XL Otaku (v1 preview)
 
 ## Model Details
 
 ### Model Description
 
-Terminus XL Gamma is a new state-of-the-art latent diffusion model that uses zero-terminal SNR noise schedule and velocity prediction objective at training and inference time.
+Terminus XL Otaku is a latent diffusion model that uses zero-terminal SNR noise schedule and velocity prediction objective at training and inference time.
 
-Terminus is based on a similar architecture to SDXL, and has the same layout. It has been trained on fewer steps with very high quality data captions via COCO and Midjourney.
+Terminus is a new state-of-the-art model family based on SDXL's architecture, and is compatible with (most) SDXL pipelines.
 
-This model will not be capable of as many concepts as SDXL, and some subjects will simply look very bad.
+For Terminus Otaku (this model), the training data is exclusively anime/celshading/3D renders and other hand-drawn or synthetic art styles.
 
-The objective of this model was to use v-prediction and min-SNR gamma loss to efficiently train a full zero-terminal SNR model on a single A100-80G.
+The objective of this model was to continue the use of v-prediction objective and min-SNR gamma loss to adapt Terminus Gamma v2's outputs to a more artistic style.
 
 
-- **Fine-tuned from:** ptx0/terminus-xl-gamma-v1
+- **Fine-tuned from:** ptx0/terminus-xl-gamma-v2
 - **Developed by:** pseudoterminal X (@bghira)
 - **Funded by:** pseudoterminal X (@bghira)
 - **Model type:** Latent Diffusion
@@ -32,13 +32,17 @@ The objective of this model was to use v-prediction and min-SNR gamma loss to ef
 
 ### Direct Use
 
-Terminus XL Gamma can be used for generating high-quality images given text prompts. It should particularly excel at inpainting tasks, where a zero-terminal SNR noise schedule allows it to more effectively retain contrast.
+Terminus XL Otaku can be used for generating high-quality images given text prompts.
+
+It should particularly excel at inpainting tasks for animated subject matter, where a zero-terminal SNR noise schedule allows it to more effectively retain contrast.
 
 The model can be utilized in creative industries such as art, advertising, and entertainment to create visually appealing content.
 
 ### Downstream Use
 
-Terminus XL Gamma can be fine-tuned for specific tasks such as image super-resolution, style transfer, and more.
+Terminus XL Otaku can be fine-tuned for specific tasks such as image super-resolution, style transfer, and more.
+
+However, it's recommended that the v1 preview not be used for fine-tuning until it is fully released, as any structural issues will hopefully be resolved by then.
 
 ### Out-of-Scope Use
 
@@ -58,27 +62,27 @@ Users should be cautious of potential biases in the generated images and thoroug
 
 This model's success largely depended on a somewhat small collection of very high quality data samples.
 
-* LAION-HD, filtered down to EXIF samples without watermarks. Luminance value of samples capped to 100 (.5).
-* Midjourney 5.2 dataset `ptx0/mj-general` with zero filtration.
+* Indiscriminate use of NijiJourney outputs.
+* Midjourney 5.2 outputs that mention anime styles in their tags.
+* Niji and MJ Showcase images that were re-captioned using CogVLM.
+* Anchor data of real human subjects in a small (10%) ratio to the animated material, to retain coherence.
 
 ### Training Procedure
 
 #### Preprocessing
 
-Most of the existing process for terminus-xl-gamma-v1 was followed, with the exception of training extensively on cropped images using SDXL's crop coordinates to improve fine details.
-
-No images were upsampled during this training session. Images were downsampled using LANCZOS instead of BICUBIC filters to attain higher image fidelity and maintain more image context for the model to learn from.
+This model is (so far) trained exclusively on cropped images using SDXL's crop coordinates to improve fine details.
 
-Only high-quality photos were used in this training session, greatly improving the realism qualities.
+No images were upsampled or downsampled during this training session. Instead, random crops (or unaltered 1024px square images) were used in lieu.
 
-~770,000 images were used for this training run.
+~50,000 images were used for this training run with continuous collection throughout the process, making it difficult to ascertain how many exact images were used.
 
 #### Training Hyperparameters
 
 - **Training regime:** bf16 mixed precision
-- **Learning rate:** \(4 \times 10^{-7}\) to \(8 \times 10^{-7}\), cosine schedule
-- **Epochs:** 60
-- **Batch size:** 24 * 15 = 360
+- **Learning rate:** \(1 \times 10^{-7}\) to \(1 \times 10^{-6}\), cosine schedule
+- **Epochs:** 11
+- **Batch size:** 12 * 8 = 96
 
 #### Speeds, Sizes, Times