app.py

import gradio as gr
import numpy as np
import random
import spaces
import torch
from diffusers import SanaSprintPipeline
import peft
from peft.tuners.lora.layer import Linear as LoraLinear
import types

dtype = torch.bfloat16
device = "cuda" if torch.cuda.is_available() else "cpu"
adapter_name = "hypernoise_adapter"

# Load the pipeline and adapter
pipe = SanaSprintPipeline.from_pretrained(
    "Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers",
    torch_dtype=dtype,
).to(device, dtype)

pipe.transformer = peft.PeftModel.from_pretrained(
    pipe.transformer,
    "lucaeyring/HyperNoise_Sana_Sprint_0.6B",
    adapter_name=adapter_name,
    dtype=dtype,
).to(device, dtype)

# Define the custom forward function for LoRA
def scaled_base_lora_forward(self, x, *args, **kwargs):
    if self.disable_adapters: 
        return self.base_layer(x, *args, **kwargs)
    return self.lora_B[adapter_name](self.lora_A[adapter_name](x)) * self.scaling[adapter_name]

# Apply the custom forward to proj_out module
for name, module in pipe.transformer.base_model.model.named_modules():
    if name == "proj_out" and isinstance(module, LoraLinear):
        module.forward = types.MethodType(scaled_base_lora_forward, module)
        break

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024  # Sana Sprint is optimized for 1024px

@spaces.GPU()
def infer(prompt, seed=42, randomize_seed=False, width=1024, height=1024, 
          num_inference_steps=4, guidance_scale=4.5, progress=gr.Progress(track_tqdm=True)):
    
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    
    # Set random seed for reproducibility
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    
    # Calculate latent dimensions based on image size
    # Sana uses 32x downsampling factor
    latent_height = height // 32
    latent_width = width // 32
    
    with torch.inference_mode():
        # Encode the prompt
        prompt_embeds, prompt_attention_mask = pipe.encode_prompt(
            [prompt], 
            device=device
        )
        
        # Generate initial random latents with correct dimensions
        init_latents = torch.randn(
            [1, 32, latent_height, latent_width], 
            device=device, 
            dtype=dtype
        )
        
        # Apply HyperNoise modulation with adapter enabled (single forward pass)
        pipe.transformer.enable_adapter_layers()
        modulated_latents = pipe.transformer(
            hidden_states=init_latents,
            encoder_hidden_states=prompt_embeds,
            encoder_attention_mask=prompt_attention_mask,
            guidance=torch.tensor([guidance_scale], device=device, dtype=dtype) * 0.1,
            timestep=torch.tensor([1.0], device=device, dtype=dtype),
        ).sample + init_latents
        
        # Generate final image with adapter disabled
        pipe.transformer.disable_adapter_layers()
        
        # For SCM scheduler, we need to handle the timesteps carefully
        # The pipeline expects intermediate_timesteps only when num_inference_steps=2
        # For other values, we use the workaround from the original code
        if num_inference_steps == 2:
            # Use the default pipeline behavior for 2 steps
            image = pipe(
                latents=modulated_latents,
                prompt_embeds=prompt_embeds,
                prompt_attention_mask=prompt_attention_mask,
                num_inference_steps=num_inference_steps,
            ).images[0]
        else:
            # For num_inference_steps != 2, we need to work around the restriction
            # by directly calling the denoising loop
            pipe.scheduler.set_timesteps(
                num_inference_steps, 
                device=device,
                timesteps=torch.linspace(1.57080, 0, num_inference_steps + 1, device=device)
            )
            
            # Run the denoising loop manually
            latents = modulated_latents
            for i, t in enumerate(pipe.scheduler.timesteps[:-1]):
                # Expand timestep to match batch dimension
                timestep = t.expand(latents.shape[0])
                
                # Predict noise
                noise_pred = pipe.transformer(
                    hidden_states=latents,
                    encoder_hidden_states=prompt_embeds,
                    encoder_attention_mask=prompt_attention_mask,
                    timestep=timestep,
                    guidance=torch.tensor([0.0], device=device, dtype=dtype),  # No guidance for denoising
                    return_dict=False,
                )[0]
                
                # Compute previous noisy sample
                latents = pipe.scheduler.step(
                    noise_pred,
                    t,
                    latents,
                    return_dict=False
                )[0]
            
            # Decode latents to image
            latents = pipe._unpack_latents(latents, height, width, pipe.vae_scale_factor)
            latents = (latents / pipe.vae.scaling_factor) + pipe.vae.shift_factor
            image = pipe.vae.decode(latents, return_dict=False)[0]
            image = pipe.image_processor.postprocess(image, output_type="pil")[0]
    
    return image, seed

examples = [
    "A smiling slice of pizza doing yoga on a mountain top",
    "A fluffy cat wearing a wizard hat casting spells",
    "A robot painting a self-portrait in Van Gogh style",
    "A tiny dragon sleeping in a teacup",
    "An astronaut riding a unicorn through a rainbow",
]

css = """
#col-container {
    margin: 0 auto;
    max-width: 520px;
}
"""

with gr.Blocks(css=css) as demo:
    
    with gr.Column(elem_id="col-container"):
        gr.Markdown("""# HyperNoise Sana Sprint 0.6B
        
Fast text-to-image generation with HyperNoise adapter for Sana Sprint model.
        
[[Sana Sprint Model](https://huggingface.co/Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers)] 
[[HyperNoise Adapter](https://huggingface.co/lucaeyring/HyperNoise_Sana_Sprint_0.6B)]
        """)
        
        with gr.Row():
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt",
                container=False,
            )
            
            run_button = gr.Button("Run", scale=0)
        
        result = gr.Image(label="Result", show_label=False)
        
        with gr.Accordion("Advanced Settings", open=False):
            
            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=42,
            )
            
            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
            
            with gr.Row():
                width = gr.Slider(
                    label="Width",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=64,
                    value=1024,
                )
                
                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=64,
                    value=1024,
                )
            
            with gr.Row():
                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=20,
                    step=1,
                    value=4,
                )
                
                guidance_scale = gr.Slider(
                    label="Guidance Scale",
                    minimum=1.0,
                    maximum=10.0,
                    step=0.5,
                    value=4.5,
                )
        
        gr.Examples(
            examples=examples,
            fn=infer,
            inputs=[prompt],
            outputs=[result, seed],
            cache_examples="lazy"
        )

    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn=infer,
        inputs=[prompt, seed, randomize_seed, width, height, num_inference_steps, guidance_scale],
        outputs=[result, seed]
    )

demo.launch()