set_timesteps() got an unexpected keyword argument 'offset' about prompt-to-prompt

[kornellewy]

!pip install diffusers==0.8.1
!pip install transformers==4.24.0

[SapirW]

I'm getting this too, diffusers==0.6.0 solved the issue

[MiaoQiaowei]

thx, i will try it!

[JoyHuYY1412]

Another solution is to replace

def run_and_display(prompts, controller, latent=None, run_baseline=False, generator=None):
    if run_baseline:
        print("w.o. prompt-to-prompt")
        images, latent = run_and_display(prompts, EmptyControl(), latent=latent, run_baseline=False, generator=generator)
        print("with prompt-to-prompt")
    images, x_t = ptp_utils.text2image_ldm_stable(ldm_stable, prompts, controller, latent=latent, num_inference_steps=NUM_DIFFUSION_STEPS, guidance_scale=GUIDANCE_SCALE, generator=generator, low_resource=LOW_RESOURCE)
    ptp_utils.view_images(images)
    return images, x_t

with

@torch.no_grad()
def text2image_ldm_stable(
    model,
    prompt:  List[str],
    controller,
    num_inference_steps: int = 50,
    guidance_scale: Optional[float] = 7.5,
    generator: Optional[torch.Generator] = None,
    latent: Optional[torch.FloatTensor] = None,
    uncond_embeddings=None,
    start_time=50,
    return_type='image'
):
    batch_size = len(prompt)
    ptp_utils.register_attention_control(model, controller)
    height = width = 512
    
    text_input = model.tokenizer(
        prompt,
        padding="max_length",
        max_length=model.tokenizer.model_max_length,
        truncation=True,
        return_tensors="pt",
    )
    text_embeddings = model.text_encoder(text_input.input_ids.to(model.device))[0]
    max_length = text_input.input_ids.shape[-1]
    if uncond_embeddings is None:
        uncond_input = model.tokenizer(
            [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
        )
        uncond_embeddings_ = model.text_encoder(uncond_input.input_ids.to(model.device))[0]
    else:
        uncond_embeddings_ = None

    latent, latents = ptp_utils.init_latent(latent, model, height, width, generator, batch_size)
    model.scheduler.set_timesteps(num_inference_steps)
    for i, t in enumerate(tqdm(model.scheduler.timesteps[-start_time:])):
        if uncond_embeddings_ is None:
            context = torch.cat([uncond_embeddings[i].expand(*text_embeddings.shape), text_embeddings])
        else:
            context = torch.cat([uncond_embeddings_, text_embeddings])
        latents = ptp_utils.diffusion_step(model, controller, latents, context, t, guidance_scale, low_resource=False)
        
    if return_type == 'image':
        image = ptp_utils.latent2image(model.vae, latents)
    else:
        image = latents
    return image, latent



def run_and_display(prompts, controller, latent=None, run_baseline=False, generator=None, uncond_embeddings=None, verbose=True):
    if run_baseline:
        print("w.o. prompt-to-prompt")
        images, latent = run_and_display(prompts, EmptyControl(), latent=latent, run_baseline=False, generator=generator)
        print("with prompt-to-prompt")
    images, x_t = text2image_ldm_stable(ldm_stable, prompts, controller, latent=latent, num_inference_steps=NUM_DIFFUSION_STEPS, guidance_scale=GUIDANCE_SCALE, generator=generator, uncond_embeddings=uncond_embeddings)
    if verbose:
        ptp_utils.view_images(images)
    return images, x_t