Merge pull request #12311 from AUTOMATIC1111/efficient-vae-methods

Add TAESD(or more) options for all the VAE encode/decode operation

modules/generation_parameters_copypaste.py

@@ -307,6 +307,12 @@ def parse_generation_parameters(x: str):
     if "Schedule rho" not in res:
         res["Schedule rho"] = 0
 
+    if "VAE Encoder" not in res:
+        res["VAE Encoder"] = "Full"
+
+    if "VAE Decoder" not in res:
+        res["VAE Decoder"] = "Full"
+
     return res
 
 
@@ -332,6 +338,8 @@ def parse_generation_parameters(x: str):
     ('RNG', 'randn_source'),
     ('NGMS', 's_min_uncond'),
     ('Pad conds', 'pad_cond_uncond'),
+    ('VAE Encoder', 'sd_vae_encode_method'),
+    ('VAE Decoder', 'sd_vae_decode_method'),
 ]
 
 

modules/processing.py

@@ -16,6 +16,7 @@
 import modules.sd_hijack
 from modules import devices, prompt_parser, masking, sd_samplers, lowvram, generation_parameters_copypaste, extra_networks, sd_vae_approx, scripts, sd_samplers_common, sd_unet, errors
 from modules.sd_hijack import model_hijack
+from modules.sd_samplers_common import images_tensor_to_samples, decode_first_stage, approximation_indexes
 from modules.shared import opts, cmd_opts, state
 import modules.shared as shared
 import modules.paths as paths
@@ -30,7 +31,6 @@
 from einops import repeat, rearrange
 from blendmodes.blend import blendLayers, BlendType
 
-decode_first_stage = sd_samplers_common.decode_first_stage
 
 # some of those options should not be changed at all because they would break the model, so I removed them from options.
 opt_C = 4
@@ -84,7 +84,7 @@ def txt2img_image_conditioning(sd_model, x, width, height):
 
         # The "masked-image" in this case will just be all zeros since the entire image is masked.
         image_conditioning = torch.zeros(x.shape[0], 3, height, width, device=x.device)
-        image_conditioning = sd_model.get_first_stage_encoding(sd_model.encode_first_stage(image_conditioning))
+        image_conditioning = images_tensor_to_samples(image_conditioning, approximation_indexes.get(opts.sd_vae_encode_method))
 
         # Add the fake full 1s mask to the first dimension.
         image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
@@ -203,7 +203,7 @@ def depth2img_image_conditioning(self, source_image):
         midas_in = torch.from_numpy(transformed["midas_in"][None, ...]).to(device=shared.device)
         midas_in = repeat(midas_in, "1 ... -> n ...", n=self.batch_size)
 
-        conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(source_image))
+        conditioning_image = images_tensor_to_samples(source_image*0.5+0.5, approximation_indexes.get(opts.sd_vae_encode_method))
         conditioning = torch.nn.functional.interpolate(
             self.sd_model.depth_model(midas_in),
             size=conditioning_image.shape[2:],
@@ -216,7 +216,7 @@ def depth2img_image_conditioning(self, source_image):
         return conditioning
 
     def edit_image_conditioning(self, source_image):
-        conditioning_image = self.sd_model.encode_first_stage(source_image).mode()
+        conditioning_image = images_tensor_to_samples(source_image*0.5+0.5, approximation_indexes.get(opts.sd_vae_encode_method))
 
         return conditioning_image
 
@@ -795,6 +795,7 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed:
             if getattr(samples_ddim, 'already_decoded', False):
                 x_samples_ddim = samples_ddim
             else:
+                p.extra_generation_params['VAE Decoder'] = opts.sd_vae_decode_method
                 x_samples_ddim = decode_latent_batch(p.sd_model, samples_ddim, target_device=devices.cpu, check_for_nans=True)
 
             x_samples_ddim = torch.stack(x_samples_ddim).float()
@@ -1135,11 +1136,10 @@ def save_intermediate(image, index):
                 batch_images.append(image)
 
             decoded_samples = torch.from_numpy(np.array(batch_images))
-            decoded_samples = decoded_samples.to(shared.device)
-            decoded_samples = 2. * decoded_samples - 1.
             decoded_samples = decoded_samples.to(shared.device, dtype=devices.dtype_vae)
 
-            samples = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(decoded_samples))
+            self.extra_generation_params['VAE Encoder'] = opts.sd_vae_encode_method
+            samples = images_tensor_to_samples(decoded_samples, approximation_indexes.get(opts.sd_vae_encode_method))
 
             image_conditioning = self.img2img_image_conditioning(decoded_samples, samples)
 
@@ -1374,10 +1374,9 @@ def init(self, all_prompts, all_seeds, all_subseeds):
             raise RuntimeError(f"bad number of images passed: {len(imgs)}; expecting {self.batch_size} or less")
 
         image = torch.from_numpy(batch_images)
-        image = 2. * image - 1.
         image = image.to(shared.device, dtype=devices.dtype_vae)
-
-        self.init_latent = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(image))
+        self.extra_generation_params['VAE Encoder'] = opts.sd_vae_encode_method
+        self.init_latent = images_tensor_to_samples(image, approximation_indexes.get(opts.sd_vae_encode_method), self.sd_model)
         devices.torch_gc()
 
         if self.resize_mode == 3:

modules/sd_samplers_common.py

@@ -23,19 +23,29 @@ def setup_img2img_steps(p, steps=None):
 approximation_indexes = {"Full": 0, "Approx NN": 1, "Approx cheap": 2, "TAESD": 3}
 
 
-def single_sample_to_image(sample, approximation=None):
+def samples_to_images_tensor(sample, approximation=None, model=None):
+    '''latents -> images [-1, 1]'''
     if approximation is None:
         approximation = approximation_indexes.get(opts.show_progress_type, 0)
 
     if approximation == 2:
-        x_sample = sd_vae_approx.cheap_approximation(sample) * 0.5 + 0.5
+        x_sample = sd_vae_approx.cheap_approximation(sample)
     elif approximation == 1:
-        x_sample = sd_vae_approx.model()(sample.to(devices.device, devices.dtype).unsqueeze(0))[0].detach() * 0.5 + 0.5
+        x_sample = sd_vae_approx.model()(sample.to(devices.device, devices.dtype)).detach()
     elif approximation == 3:
         x_sample = sample * 1.5
-        x_sample = sd_vae_taesd.model()(x_sample.to(devices.device, devices.dtype).unsqueeze(0))[0].detach()
+        x_sample = sd_vae_taesd.decoder_model()(x_sample.to(devices.device, devices.dtype)).detach()
+        x_sample = x_sample * 2 - 1
     else:
-        x_sample = decode_first_stage(shared.sd_model, sample.unsqueeze(0))[0] * 0.5 + 0.5
+        if model is None:
+            model = shared.sd_model
+        x_sample = model.decode_first_stage(sample)
+
+    return x_sample
+
+
+def single_sample_to_image(sample, approximation=None):
+    x_sample = samples_to_images_tensor(sample.unsqueeze(0), approximation)[0] * 0.5 + 0.5
 
     x_sample = torch.clamp(x_sample, min=0.0, max=1.0)
     x_sample = 255. * np.moveaxis(x_sample.cpu().numpy(), 0, 2)
@@ -45,9 +55,9 @@ def single_sample_to_image(sample, approximation=None):
 
 
 def decode_first_stage(model, x):
-    x = model.decode_first_stage(x.to(devices.dtype_vae))
-
-    return x
+    x = x.to(devices.dtype_vae)
+    approx_index = approximation_indexes.get(opts.sd_vae_decode_method, 0)
+    return samples_to_images_tensor(x, approx_index, model)
 
 
 def sample_to_image(samples, index=0, approximation=None):
@@ -58,6 +68,24 @@ def samples_to_image_grid(samples, approximation=None):
     return images.image_grid([single_sample_to_image(sample, approximation) for sample in samples])
 
 
+def images_tensor_to_samples(image, approximation=None, model=None):
+    '''image[0, 1] -> latent'''
+    if approximation is None:
+        approximation = approximation_indexes.get(opts.sd_vae_encode_method, 0)
+
+    if approximation == 3:
+        image = image.to(devices.device, devices.dtype)
+        x_latent = sd_vae_taesd.encoder_model()(image)
+    else:
+        if model is None:
+            model = shared.sd_model
+        image = image.to(shared.device, dtype=devices.dtype_vae)
+        image = image * 2 - 1
+        x_latent = model.get_first_stage_encoding(model.encode_first_stage(image))
+
+    return x_latent
+
+
 def store_latent(decoded):
     state.current_latent = decoded
 

modules/sd_vae_approx.py

@@ -81,6 +81,6 @@ def cheap_approximation(sample):
 
     coefs = torch.tensor(coeffs).to(sample.device)
 
-    x_sample = torch.einsum("lxy,lr -> rxy", sample, coefs)
+    x_sample = torch.einsum("...lxy,lr -> ...rxy", sample, coefs)
 
     return x_sample

modules/sd_vae_taesd.py

@@ -44,7 +44,17 @@ def decoder():
     )
 
 
-class TAESD(nn.Module):
+def encoder():
+    return nn.Sequential(
+        conv(3, 64), Block(64, 64),
+        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
+        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
+        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
+        conv(64, 4),
+    )
+
+
+class TAESDDecoder(nn.Module):
     latent_magnitude = 3
     latent_shift = 0.5
 
@@ -55,21 +65,28 @@ def __init__(self, decoder_path="taesd_decoder.pth"):
         self.decoder.load_state_dict(
             torch.load(decoder_path, map_location='cpu' if devices.device.type != 'cuda' else None))
 
-    @staticmethod
-    def unscale_latents(x):
-        """[0, 1] -> raw latents"""
-        return x.sub(TAESD.latent_shift).mul(2 * TAESD.latent_magnitude)
+
+class TAESDEncoder(nn.Module):
+    latent_magnitude = 3
+    latent_shift = 0.5
+
+    def __init__(self, encoder_path="taesd_encoder.pth"):
+        """Initialize pretrained TAESD on the given device from the given checkpoints."""
+        super().__init__()
+        self.encoder = encoder()
+        self.encoder.load_state_dict(
+            torch.load(encoder_path, map_location='cpu' if devices.device.type != 'cuda' else None))
 
 
 def download_model(model_path, model_url):
     if not os.path.exists(model_path):
         os.makedirs(os.path.dirname(model_path), exist_ok=True)
 
-        print(f'Downloading TAESD decoder to: {model_path}')
+        print(f'Downloading TAESD model to: {model_path}')
         torch.hub.download_url_to_file(model_url, model_path)
 
 
-def model():
+def decoder_model():
     model_name = "taesdxl_decoder.pth" if getattr(shared.sd_model, 'is_sdxl', False) else "taesd_decoder.pth"
     loaded_model = sd_vae_taesd_models.get(model_name)
 
@@ -78,11 +95,30 @@ def model():
         download_model(model_path, 'https://github.com/madebyollin/taesd/raw/main/' + model_name)
 
         if os.path.exists(model_path):
-            loaded_model = TAESD(model_path)
+            loaded_model = TAESDDecoder(model_path)
             loaded_model.eval()
             loaded_model.to(devices.device, devices.dtype)
             sd_vae_taesd_models[model_name] = loaded_model
         else:
             raise FileNotFoundError('TAESD model not found')
 
     return loaded_model.decoder
+
+
+def encoder_model():
+    model_name = "taesdxl_encoder.pth" if getattr(shared.sd_model, 'is_sdxl', False) else "taesd_encoder.pth"
+    loaded_model = sd_vae_taesd_models.get(model_name)
+
+    if loaded_model is None:
+        model_path = os.path.join(paths_internal.models_path, "VAE-taesd", model_name)
+        download_model(model_path, 'https://github.com/madebyollin/taesd/raw/main/' + model_name)
+
+        if os.path.exists(model_path):
+            loaded_model = TAESDEncoder(model_path)
+            loaded_model.eval()
+            loaded_model.to(devices.device, devices.dtype)
+            sd_vae_taesd_models[model_name] = loaded_model
+        else:
+            raise FileNotFoundError('TAESD model not found')
+
+    return loaded_model.encoder

modules/shared.py

@@ -435,6 +435,8 @@ def list_samplers():
     "upcast_attn": OptionInfo(False, "Upcast cross attention layer to float32"),
     "auto_vae_precision": OptionInfo(True, "Automaticlly revert VAE to 32-bit floats").info("triggers when a tensor with NaNs is produced in VAE; disabling the option in this case will result in a black square image"),
     "randn_source": OptionInfo("GPU", "Random number generator source.", gr.Radio, {"choices": ["GPU", "CPU", "NV"]}).info("changes seeds drastically; use CPU to produce the same picture across different videocard vendors; use NV to produce same picture as on NVidia videocards"),
+    "sd_vae_encode_method": OptionInfo("Full", "VAE type for encode", gr.Radio, {"choices": ["Full", "TAESD"]}).info("method to encode image to latent (use in img2img, hires-fix or inpaint mask)"),
+    "sd_vae_decode_method": OptionInfo("Full", "VAE type for decode", gr.Radio, {"choices": ["Full", "TAESD"]}).info("method to decode latent to image"),
 }))
 
 options_templates.update(options_section(('sdxl', "Stable Diffusion XL"), {