Update fluxcombined.py

fluxcombined.py

@@ -874,7 +874,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         # initialize the random noise for denoising
         latents = random_latents.clone().detach()
 
-        self.vae = self.vae.to(torch.float32)
+        # self.vae = self.vae.to(torch.float32)
 
         # 9. Denoising loop
         self.transformer.eval()
@@ -959,7 +959,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         else:
             latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
             latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
-            image = self.vae.decode(latents.to(torch.float32), return_dict=False)[0]
+            image = self.vae.decode(latents, return_dict=False)[0]
             image = self.image_processor.postprocess(image, output_type=output_type)
 
         # Offload all models
@@ -973,7 +973,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
     def get_diff_image(self, latents):
         latents = self._unpack_latents(latents, 1024, 1024, self.vae_scale_factor)
         latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
-        image = self.vae.decode(latents.to(torch.float32), return_dict=False)[0]
+        image = self.vae.decode(latents, return_dict=False)[0]
         image = self.image_processor.postprocess(image, output_type="pt")
         return image
     
@@ -983,7 +983,6 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         img = img.resize((512, 512))
         return custom_image_processor(img).unsqueeze(0).to(device)
     
-    
     @torch.no_grad()
     def edit(
         self,
@@ -1019,283 +1018,6 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         mask_image=None,
         source_steps=1,
     ):
-
-        height = height or self.default_sample_size * self.vae_scale_factor
-        width = width or self.default_sample_size * self.vae_scale_factor
-
-        # 1. Check inputs. Raise error if not correct
-        self.check_inputs(
-            prompt,
-            prompt_2,
-            height,
-            width,
-            # negative_prompt=negative_prompt,
-            # negative_prompt_2=negative_prompt_2,
-            prompt_embeds=prompt_embeds,
-            # negative_prompt_embeds=negative_prompt_embeds,
-            pooled_prompt_embeds=pooled_prompt_embeds,
-            # negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
-            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
-            max_sequence_length=max_sequence_length,
-        )
-
-        self._guidance_scale = guidance_scale
-        self._joint_attention_kwargs = joint_attention_kwargs
-        self._interrupt = False
-
-        # 2. Define call parameters
-        if prompt is not None and isinstance(prompt, str):
-            batch_size = 1
-        elif prompt is not None and isinstance(prompt, list):
-            batch_size = len(prompt)
-        else:
-            batch_size = prompt_embeds.shape[0]
-
-        device = self._execution_device
-
-        lora_scale = (
-            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
-        )
-        do_true_cfg = true_cfg > 1 and negative_prompt is not None
-        (
-            prompt_embeds,
-            pooled_prompt_embeds,
-            text_ids,
-            negative_prompt_embeds,
-            negative_pooled_prompt_embeds,
-        ) = self.encode_prompt_edit(
-            prompt=prompt,
-            prompt_2=prompt_2,
-            negative_prompt=negative_prompt,
-            negative_prompt_2=negative_prompt_2,
-            prompt_embeds=prompt_embeds,
-            pooled_prompt_embeds=pooled_prompt_embeds,
-            negative_prompt_embeds=negative_prompt_embeds,
-            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
-            device=device,
-            num_images_per_prompt=num_images_per_prompt,
-            max_sequence_length=max_sequence_length,
-            lora_scale=lora_scale,
-            do_true_cfg=do_true_cfg,
-        )
-        # text_ids = text_ids.repeat(batch_size, 1, 1)
-
-        if do_true_cfg:
-            # Concatenate embeddings
-            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
-            pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
-
-        # 4. Prepare latent variables
-        num_channels_latents = self.transformer.config.in_channels // 4
-        random_latents, latent_image_ids = self.prepare_latents(
-            batch_size * num_images_per_prompt,
-            num_channels_latents,
-            height,
-            width,
-            prompt_embeds.dtype,
-            device,
-            generator,
-            latents,
-        )
-        # latent_image_ids = latent_image_ids.repeat(batch_size, 1, 1)
-
-        # 5. Prepare timesteps
-        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
-        image_seq_len = random_latents.shape[1]
-        mu = calculate_shift(
-            image_seq_len,
-            self.scheduler.config.base_image_seq_len,
-            self.scheduler.config.max_image_seq_len,
-            self.scheduler.config.base_shift,
-            self.scheduler.config.max_shift,
-        )
-        timesteps, num_inference_steps = retrieve_timesteps(
-            self.scheduler,
-            num_inference_steps,
-            device,
-            timesteps,
-            sigmas,
-            mu=mu,
-        )
-        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
-        self._num_timesteps = len(timesteps)
-
-        # 4. Preprocess image
-        image = self.image_processor.preprocess(input_image)
-        image = image.to(device=device, dtype=self.transformer.dtype)
-        latents = retrieve_latents(self.vae.encode(image), generator=generator) * self.vae.config.scaling_factor
-
-
-        # Convert PIL image to tensor
-        if mask_image:
-            from torchvision import transforms as TF
-
-            h, w = latents.shape[2], latents.shape[3]
-            mask = TF.ToTensor()(mask_image).to(device=device, dtype=self.transformer.dtype)
-            mask = TF.Resize((h, w), interpolation=TF.InterpolationMode.NEAREST)(mask)
-            mask = (mask > 0.5).float()
-            mask = mask.squeeze(0)#.squeeze(0)  # Remove the added dimensions
-        else:
-            mask = torch.ones_like(latents).to(device=device)
-
-        print(mask.shape, latents.shape)
-
-        bool_mask = mask.unsqueeze(0).unsqueeze(0).expand_as(latents)
-        mask=(1-bool_mask*1.0).to(latents.dtype)
-        
-        masked_latents = (latents * mask).clone().detach() # apply the mask and get gt_latents
-        masked_latents = self._pack_latents(masked_latents, batch_size, num_channels_latents, 2 * (int(height) // self.vae_scale_factor), 2 * (int(width) // self.vae_scale_factor))
-
-        source_latents = (latents).clone().detach() # apply the mask and get gt_latents
-        source_latents = self._pack_latents(source_latents, batch_size, num_channels_latents, 2 * (int(height) // self.vae_scale_factor), 2 * (int(width) // self.vae_scale_factor))
-
-        mask = self._pack_latents(mask, batch_size, num_channels_latents, 2 * (int(height) // self.vae_scale_factor), 2 * (int(width) // self.vae_scale_factor))
-
-        # initialize the random noise for denoising
-        latents = random_latents.clone().detach()
-
-        self.vae = self.vae.to(torch.float32)
-
-        # 9. Denoising loop
-        self.transformer.eval()
-        self.vae.eval()
-
-        # 6. Denoising loop
-        with self.progress_bar(total=num_inference_steps) as progress_bar:
-            for i, t in enumerate(timesteps):
-                if self.interrupt:
-                    continue
-
-                latent_model_input = torch.cat([latents] * 2) if do_true_cfg else latents
-
-                # handle guidance
-                if self.transformer.config.guidance_embeds:
-                    guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
-                    guidance = guidance.expand(latent_model_input.shape[0])
-                else:
-                    guidance = None
-
-                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
-                timestep = t.expand(latent_model_input.shape[0]).to(latent_model_input.dtype)
-
-                noise_pred = self.transformer(
-                    hidden_states=latent_model_input,
-                    timestep=timestep / 1000,
-                    guidance=guidance,
-                    pooled_projections=pooled_prompt_embeds,
-                    encoder_hidden_states=prompt_embeds,
-                    txt_ids=text_ids,
-                    img_ids=latent_image_ids,
-                    joint_attention_kwargs=self.joint_attention_kwargs,
-                    return_dict=False,
-                )[0]
-
-                if do_true_cfg:
-                    neg_noise_pred, noise_pred = noise_pred.chunk(2)
-                    # noise_pred = neg_noise_pred + true_cfg * (noise_pred - neg_noise_pred)
-                    noise_pred = noise_pred + (1-mask)*(noise_pred - neg_noise_pred) * true_cfg
-                # else:
-                #     neg_noise_pred, noise_pred = noise_pred.chunk(2)
-
-                # perform CG
-                if i < max_steps:
-                    opt_latents = latents.detach().clone()
-                    with torch.enable_grad():
-                        opt_latents = opt_latents.detach().requires_grad_()
-                        opt_latents = torch.autograd.Variable(opt_latents, requires_grad=True)
-                        # optimizer = torch.optim.Adam([opt_latents], lr=learning_rate)
-
-                        for _ in range(optimization_steps):
-                            latents_p = self.scheduler.step_final(noise_pred, t, opt_latents, return_dict=False)[0]
-                            if i < source_steps:
-                                loss = (1000*torch.nn.functional.mse_loss(latents_p, source_latents, reduction='none')).mean()
-                            else:
-                                loss = (1000*torch.nn.functional.mse_loss(latents_p, masked_latents, reduction='none')*mask).mean()
-
-                            grad = torch.autograd.grad(loss, opt_latents)[0]
-                            # grad = torch.clamp(grad, -0.5, 0.5)
-                            opt_latents = opt_latents - learning_rate * grad
-
-                    latents = opt_latents.detach().clone()
-
-
-                # compute the previous noisy sample x_t -> x_t-1
-                latents_dtype = latents.dtype
-                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
-
-                if latents.dtype != latents_dtype:
-                    if torch.backends.mps.is_available():
-                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
-                        latents = latents.to(latents_dtype)
-
-                if callback_on_step_end is not None:
-                    callback_kwargs = {}
-                    for k in callback_on_step_end_tensor_inputs:
-                        callback_kwargs[k] = locals()[k]
-                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
-
-                    latents = callback_outputs.pop("latents", latents)
-                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
-
-                # call the callback, if provided
-                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
-                    progress_bar.update()
-
-                if XLA_AVAILABLE:
-                    xm.mark_step()
-
-        if output_type == "latent":
-            image = latents
-
-        else:
-            latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
-            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
-            image = self.vae.decode(latents.to(torch.float32), return_dict=False)[0]
-            image = self.image_processor.postprocess(image, output_type=output_type)
-
-        # Offload all models
-        self.maybe_free_model_hooks()
-
-        if not return_dict:
-            return (image,)
-
-        return FluxPipelineOutput(images=image)
-    
-    @torch.no_grad()
-    def edit2(
-        self,
-        prompt: Union[str, List[str]] = None,
-        prompt_2: Optional[Union[str, List[str]]] = None,
-        negative_prompt: Union[str, List[str]] = None,  #
-        negative_prompt_2: Optional[Union[str, List[str]]] = None,
-        true_cfg: float = 1.0,  #
-        height: Optional[int] = None,
-        width: Optional[int] = None,
-        num_inference_steps: int = 28,
-        timesteps: List[int] = None,
-        guidance_scale: float = 3.5,
-        num_images_per_prompt: Optional[int] = 1,
-        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
-        latents: Optional[torch.FloatTensor] = None,
-        prompt_embeds: Optional[torch.FloatTensor] = None,
-        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
-        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
-        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
-        output_type: Optional[str] = "pil",
-        return_dict: bool = True,
-        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
-        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
-        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
-        max_sequence_length: int = 512,
-        optimization_steps: int = 3,
-        learning_rate: float = 0.8,
-        max_steps: int = 5,
-        input_image = None,
-        save_masked_image = False,
-        output_path="",
-        mask_image=None,
-        source_steps=1,
-    ):
         r"""
         Function invoked when calling the pipeline for generation.
 
@@ -1498,7 +1220,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         # initialize the random noise for denoising
         latents = random_latents.clone().detach()
 
-        self.vae = self.vae.to(torch.float32)
+        # self.vae = self.vae.to(torch.float32)
 
         # 9. Denoising loop
         self.transformer.eval()
@@ -1594,7 +1316,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         else:
             latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
             latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
-            image = self.vae.decode(latents.to(torch.float32), return_dict=False)[0]
+            image = self.vae.decode(latents, return_dict=False)[0]
             image = self.image_processor.postprocess(image, output_type=output_type)
 
         # Offload all models