Hello, thank you for open-sourcing the Tar project.
I'm currently trying to reproduce the text-image-to-image (TI→I) task mentioned in your paper, aiming to generate images like "a photo of XX in the style of
".
My implementation is similar to this:
prompt = "a photo of panda in the style of <image>"
image = Image.open(image_path).convert('RGB')
image = to_tensor(image).unsqueeze(0).to(self.device)
image_code = self.input_visual_tokenizer(image)['encoded']
image_text = "".join([f"<I{x}>" for x in image_code[0].cpu().tolist()])
# Prepare prompt
messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": f"{image_text}\n{prompt}"}
]
input_text = self.tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True)
input_text += f"<im_start><S{self.config.scale}>"
# Generate tokens
inputs = self.tokenizer(input_text, return_tensors="pt")
gen_ids = self.model.generate(
inputs.input_ids.to(self.device),
max_new_tokens=self.config.seq_len,
do_sample=True,
temperature=self.config.temperature,
top_p=self.config.top_p,
top_k=self.config.top_k)
# Process generated tokens
gen_text = self.tokenizer.batch_decode(gen_ids)[0]
gen_code = [int(x) for x in re.findall(r'<I(\d+)>', gen_text)]
gen_code = gen_code[:self.config.seq_len] + [0] * max(0, self.config.seq_len - len(gen_code))
gen_code = torch.tensor(gen_code).unsqueeze(0).to(self.device)
gen_tensor = self.visual_tokenizer.decode_from_encoder_indices(
gen_code,
{'cfg_scale': self.config.cfg_scale}
)
gen_image = Image.fromarray(gen_tensor[0].numpy())
However, the generated image is almost identical to the input image, rather than a new image in the specified style. Could you please provide some guidance on the recommended inference code for handling subject-driven TI→I tasks with the Tar model?
Any code examples or detailed instructions would be greatly appreciated.
Thank you again for your hard work and for sharing this project with us!
Hello, thank you for open-sourcing the Tar project.
I'm currently trying to reproduce the text-image-to-image (TI→I) task mentioned in your paper, aiming to generate images like "a photo of XX in the style of
".
My implementation is similar to this:
However, the generated image is almost identical to the input image, rather than a new image in the specified style. Could you please provide some guidance on the recommended inference code for handling subject-driven TI→I tasks with the Tar model?
Any code examples or detailed instructions would be greatly appreciated.
Thank you again for your hard work and for sharing this project with us!