improve vae loading

CHANGELOG.md

@@ -19,6 +19,8 @@
   - enable fp16 vae decode when using optimized vae  
     this pretty much doubles performance of decode step (delay after generate is done)  
   - sd-xl: loading vae now applies to both base and refiner and saves a bit of vram  
+  - vae: enable loading of pure-safetensors vae files without config  
+    also enable *automatic* selection to work with diffusers  
 - diffusers: future-proof  
   requires `diffusers==0.19.dev`, not yet released, but can be installed manually  
   - sd-xl: denoising_start/denoising_end

modules/sd_vae.py

@@ -76,6 +76,12 @@ def refresh_vae_list():
                 os.path.join(shared.opts.vae_dir, '**/*.safetensors'),
             ]
     elif shared.backend == shared.Backend.DIFFUSERS:
+        if sd_models.model_path is not None and os.path.isdir(sd_models.model_path):
+            vae_paths += [os.path.join(sd_models.model_path, 'VAE', '**/*.vae.safetensors')]
+        if shared.opts.ckpt_dir is not None and os.path.isdir(shared.opts.ckpt_dir):
+            vae_paths += [os.path.join(shared.opts.ckpt_dir, '**/*.vae.safetensors')]
+        if shared.opts.vae_dir is not None and os.path.isdir(shared.opts.vae_dir):
+            vae_paths += [os.path.join(shared.opts.vae_dir, '**/*.safetensors')]
         vae_paths += [
             os.path.join(sd_models.model_path, 'VAE', '**/*.json'),
             os.path.join(shared.opts.vae_dir, '**/*.json'),
@@ -88,7 +94,10 @@ def refresh_vae_list():
         if shared.backend == shared.Backend.ORIGINAL:
             vae_dict[name] = filepath
         else:
+            if filepath.endswith(".json"):
                 vae_dict[name] = os.path.dirname(filepath)
+            else:
+                vae_dict[name] = filepath
     shared.log.info(f"Available VAEs: {vae_path} {len(vae_dict)}")
 
 
@@ -176,6 +185,11 @@ def load_vae_diffusers(_model, vae_file=None, vae_source="from unknown source"):
     shared.log.debug(f'Diffusers VAE load config: {diffusers_load_config}')
     try:
         import diffusers
+        print('HERE', vae_file, diffusers_load_config)
+        if os.path.isfile(vae_file):
+            vae = diffusers.AutoencoderKL.from_single_file(vae_file)
+            vae = vae.to(devices.dtype_vae)
+        else:
             vae = diffusers.AutoencoderKL.from_pretrained(vae_file, **diffusers_load_config)
         # shared.log.debug(f'Diffusers VAE config: {vae.config}')
         return vae

pyproject.toml

@@ -52,6 +52,7 @@ ignore = [
   "RUF005", # Consider concatenation
   "RUF012", # Mutable class attributes
   "RUF013", # Implict optional
+  "RUF015", # Prefer `next`
 ]
 
 [tool.ruff.flake8-bugbear]