nightly

eval_alchemist.py

@@ -15,8 +15,8 @@ DTYPE = torch.float16
 IMAGE_FOLDER = "/workspace/alchemist" #wget https://huggingface.co/datasets/AiArtLab/alchemist/resolve/main/alchemist.zip
 MIN_SIZE = 1280
 CROP_SIZE = 512
-BATCH_SIZE = 5
-MAX_IMAGES = 100
+BATCH_SIZE = 1
+MAX_IMAGES = 1000
 NUM_WORKERS = 4
 NUM_SAMPLES_TO_SAVE = 10  # Сколько примеров сохранить (0 - не сохранять)
 SAMPLES_FOLDER = "vaetest"
@@ -26,11 +26,11 @@ VAE_LIST = [
 
 #    ("stable-diffusion-v1-5/stable-diffusion-v1-5", AutoencoderKL, "stable-diffusion-v1-5/stable-diffusion-v1-5", "vae"),
 #    ("cross-attention/asymmetric-autoencoder-kl-x-1-5", AsymmetricAutoencoderKL, "cross-attention/asymmetric-autoencoder-kl-x-1-5", None),
-    ("madebyollin/sdxl-vae-fp16", AutoencoderKL, "madebyollin/sdxl-vae-fp16-fix", None),
+#    ("madebyollin/sdxl-vae-fp16", AutoencoderKL, "madebyollin/sdxl-vae-fp16-fix", None),
 #    ("AiArtLab/sdxs", AutoencoderKL, "AiArtLab/sdxs", "vae"),
-    ("AiArtLab/sdxl_vae", AutoencoderKL, "AiArtLab/sdxl_vae", None),
+#    ("vae", AutoencoderKL, "/workspace/sdxl_vae/vae", None),
 #    ("AiArtLab/sdxl_vae_asym", AsymmetricAutoencoderKL, "AiArtLab/sdxl_vae", "asymmetric_vae"),
-    ("AiArtLab/sdxl_vae_asym_new", AsymmetricAutoencoderKL, "AiArtLab/sdxl_vae", "asymmetric_vae_new"),
+    ("vae_nightly", AutoencoderKL, "/workspace/sdxl_vae/vae_nightly", None),
 #    ("KBlueLeaf/EQ-SDXL-VAE", AutoencoderKL, "KBlueLeaf/EQ-SDXL-VAE", None),
 #    ("FLUX.1-schnell-vae", AutoencoderKL, "black-forest-labs/FLUX.1-schnell", "vae"),
 ]

train_sdxl_vae.py

@@ -33,7 +33,7 @@ sample_interval_share = 10
 use_wandb          = True
 save_model         = True
 use_decay          = True
-asymmetric         = True
+asymmetric         = False
 optimizer_type     = "adam8bit"
 dtype              = torch.float32
 # model_resolution — то, что подавается в VAE (низкое разрешение)

vae_nightly/config.json

@@ -0,0 +1,38 @@
+{
+  "_class_name": "AutoencoderKL",
+  "_diffusers_version": "0.35.0.dev0",
+  "_name_or_path": "vae",
+  "act_fn": "silu",
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    512
+  ],
+  "down_block_types": [
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D"
+  ],
+  "force_upcast": false,
+  "in_channels": 3,
+  "latent_channels": 4,
+  "latents_mean": null,
+  "latents_std": null,
+  "layers_per_block": 2,
+  "mid_block_add_attention": true,
+  "norm_num_groups": 32,
+  "out_channels": 3,
+  "sample_size": 512,
+  "scaling_factor": 0.13025,
+  "shift_factor": null,
+  "up_block_types": [
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D"
+  ],
+  "use_post_quant_conv": true,
+  "use_quant_conv": true
+}

vae_nightly/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7e5c30a8580ae711f0401a71e37d2ab96d2cac03392f0469f0c17ed333e3ae3
+size 334643268