@torch.inference_mode()
    def predict(self, dreams: List[Data], entry_time: int):
        # handle timeout
        if time.time() - entry_time > INFERENCE_REQUEST_TIMEOUT:
            raise TimeoutException()

        # sets up the inference settings
        height = width = IMAGE_SIZE
        num_inference_steps = 50 if dreams[0].high_quality else 25

        prompts = [dream.prompt for dream in dreams]

        # GPU inference
        if torch.cuda.is_available():
            with autocast("cuda"):
                torch.cuda.empty_cache()
                pil_results = self._model(
                    prompts,
                    height=height,
                    width=width,
                    num_inference_steps=num_inference_steps,
                )
            # apply filter
            nsfw_content = self._safety_checker(pil_results)
            for i, nsfw in enumerate(nsfw_content):
                if nsfw:
                    pil_results[i] = Image.open("assets/nsfw-warning.png")
        else:
            time.sleep(3)
            pil_results = [Image.fromarray(np.random.randint(0, 255, (height, width, 3), dtype="uint8"))] * len(prompts)

        # return the model
        results = []
        for image in pil_results:
            buffered = BytesIO()
            image.save(buffered, format="PNG")
            img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
            # make sure pil_results is a single item array or it'll rewrite image
            results.append({"image": f"data:image/png;base64,{img_str}"})

        return results