1st commit

.gitattributes

@@ -32,3 +32,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,52 @@
 ---
 license: unknown
 ---
+# The White Queen Dreambooth Model
+
+The tweo models included were trained to compare a novel approach that I call "multi-resolution" dreambooth
+
+I have combined instance images with two resolution:
+
+- 512x512
+- 640x448
+
+One model was produced using the multi-subject feature of this dreambooth fork: https://github.com/ShivamShrirao/diffusers 
+
+That fork was slightly modified to allow it to handle non square resolutions and to allow various subject instances to use different sizes.
+
+The other one was produced by creating two seperate models (one for each different instance resolution) and them merging them according to the number of instance images in each.
+
+Based on preliminary testing it appear that the multi-subject approach provides the best results.
+
+Here is the shivam configuration used for the multi-subject training:
+
+```
+  # multi-subject for different ratio processing
+
+  accelerate launch --num_cpu_threads_per_process 6 train_dreambooth.py `
+    --pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" `
+    --pretrained_vae_name_or_path="stabilityai/sd-vae-ft-mse" `
+    --with_prior_preservation --prior_loss_weight=1 `
+    --seed=494481440 `
+    --resolution=448 `
+    --train_batch_size=4 `
+    --train_text_encoder `
+    --mixed_precision="fp16" `
+    --revision="fp16" `
+    --use_8bit_adam `
+    --gradient_checkpointing `
+    --gradient_accumulation_steps=1 `
+    --learning_rate=1e-6 `
+    --lr_scheduler="constant" `
+    --lr_warmup_steps=0 `
+    --num_class_images=256 `
+    --sample_batch_size=4 `
+    --max_train_steps=1500 `
+    --save_interval=400 `
+    --save_sample_prompt="twq woman" `
+    --concepts_list="concepts_list_twq.json" `
+    --output_dir="./twq/output-multi" `
+    --n_save_sample=4 `
+    --save_min_steps=400 `
+    --pad_tokens
+```

twq_woman-shiv-1200-multires-merge.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:259d11d6c8122253de519b598c1f1b16dec2e45d0d07924308e7660cb56babd2
+size 2132856622

twq_woman-shiv-1200-multires.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d4f02ebe3c940e535c8ae248472daeceec3815e5f88782a060247286298d239
+size 2132856622