p1atdev commited on
Commit
4c62670
·
verified ·
1 Parent(s): d7db10f

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +141 -0
README.md CHANGED
@@ -132,3 +132,144 @@ You should use `figma` to trigger the image generation.
132
  Weights for this model are available in Safetensors format.
133
 
134
  [Download](/p1atdev/animagine-xl-4.0-opt-pvc-style-lora/tree/main) them in the Files & versions tab.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
132
  Weights for this model are available in Safetensors format.
133
 
134
  [Download](/p1atdev/animagine-xl-4.0-opt-pvc-style-lora/tree/main) them in the Files & versions tab.
135
+
136
+ ## Dataset
137
+
138
+ About 7000 of pvc figure images.
139
+
140
+ ## Training
141
+
142
+ ### Training code
143
+
144
+ https://github.com/p1atdev/vision-ft
145
+
146
+ ### Training settings
147
+
148
+
149
+ <details>
150
+ <summary>`config.yml`</summary>
151
+
152
+ ```yml
153
+ model:
154
+ checkpoint_path: "./models/animagine-xl-4.0-opt.bnb_nf4.safetensors"
155
+ pretrained_model_name_or_path: cagliostrolab/animagine-xl-4.0
156
+
157
+ dtype: bfloat16
158
+
159
+ denoiser:
160
+ attn_implementation: "flash_attention_2"
161
+
162
+ peft:
163
+ config:
164
+ type: lora
165
+ rank: 2
166
+ alpha: 1.0
167
+ dropout: 0.0
168
+
169
+ dtype: bfloat16
170
+
171
+ include_keys:
172
+ - "attn1"
173
+ - "attn2"
174
+ - ".ff."
175
+ exclude_keys: ["text_encoder", "vae"]
176
+
177
+ dataset:
178
+ folder: "data/pvc"
179
+ num_repeats: 4
180
+ batch_size: 4
181
+
182
+ bucket_base_size: 1024
183
+ step: 128
184
+ min_size: 384
185
+ do_upscale: true
186
+
187
+ caption_processors:
188
+ - type: shuffle
189
+ split_separator: ","
190
+
191
+ optimizer:
192
+ name: "schedulefree.RAdamScheduleFree"
193
+ args:
194
+ lr: 0.005
195
+
196
+ scheduler:
197
+ # name: "torch.optim.lr_scheduler.ConstantLR"
198
+ # args: {}
199
+
200
+ tracker:
201
+ project_name: "sdxl-pvc-1"
202
+ loggers:
203
+ - wandb
204
+
205
+ saving:
206
+ strategy:
207
+ per_epochs: 0.25
208
+ per_steps: null
209
+ save_last: true
210
+
211
+ callbacks:
212
+ - type: "safentensors"
213
+ name: "sdxl-pvc"
214
+ save_dir: "./output/sdxl-pvc-2"
215
+
216
+ preview:
217
+ strategy:
218
+ per_epochs: 1
219
+ per_steps: 100
220
+
221
+ callbacks:
222
+ - type: "local"
223
+ save_dir: "./output/sdxl-lora/preview"
224
+
225
+ data:
226
+ path: "./projects/pvc-sdxl/preview.yml"
227
+
228
+ seed: 42
229
+ num_train_epochs: 10
230
+
231
+ trainer:
232
+ # debug_mode: "1step"
233
+
234
+ gradient_checkpointing: true
235
+
236
+ torch_compile: true
237
+ torch_compile_args:
238
+ backend: eager
239
+ mode: default
240
+ fullgraph: false
241
+
242
+ fp32_matmul_precision: "high"
243
+ allow_tf32: true
244
+ ```
245
+
246
+ </details>
247
+
248
+ <details>
249
+ <summary>`preview.yml` </summary>
250
+
251
+ ```yml
252
+ - prompt: |-
253
+ pvc style, 1girl, aqua eyes, baseball cap, blonde hair, closed mouth, earrings,
254
+ green background, hat, hoop earrings, jewelry, looking at viewer,
255
+ shirt, short hair, simple background, solo, upper body, yellow shirt,
256
+ masterpiece, high score, great score, absurdres
257
+ negative_prompt: |-
258
+ lowres, bad anatomy, bad hands, text, error, missing finger, extra digits,
259
+ fewer digits, cropped, worst quality, low quality, low score, bad score,
260
+ average score, signature, watermark, username, blurry
261
+ height: 1024
262
+ width: 1024
263
+ cfg_scale: 5.0
264
+ num_steps: 20
265
+
266
+ seed: 0
267
+ ```
268
+
269
+ </details>
270
+
271
+ The training was stopped in the 9th epoch, because of my room was too hot due to this training.
272
+
273
+ ### Device and Hours
274
+
275
+ RTX 4070 Ti Super (VRAM 16GB) for 27 hours.