init

app.py

@@ -9,14 +9,11 @@ from pipeline_freescale import StableDiffusionXLPipeline
 from free_lunch_utils import register_free_upblock2d, register_free_crossattn_upblock2d
 
 @spaces.GPU(duration=120)
-def infer_gpu_part(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_scale, resolutions_list, fast_mode, cosine_scale, disable_freeu):
+def infer_gpu_part(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_scale, resolutions_list, fast_mode, cosine_scale):
     pipe = pipe.to("cuda") 
     generator = torch.Generator(device='cuda')
     generator = generator.manual_seed(seed)
-    if not disable_freeu:
-        register_free_upblock2d(pipe, b1=1.1, b2=1.2, s1=0.6, s2=0.4)
-        register_free_crossattn_upblock2d(pipe, b1=1.1, b2=1.2, s1=0.6, s2=0.4)
-    resul = pipe(prompt, negative_prompt=negative_prompt, generator=generator,
+    result = pipe(prompt, negative_prompt=negative_prompt, generator=generator,
                 num_inference_steps=ddim_steps, guidance_scale=guidance_scale,
                 resolutions_list=resolutions_list, fast_mode=fast_mode, cosine_scale=cosine_scale,
                 )
@@ -44,9 +41,12 @@ def infer(prompt, output_size, ddim_steps, guidance_scale, cosine_scale, seed, o
 
     model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
     pipe = StableDiffusionXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16)
+    if not disable_freeu:
+        register_free_upblock2d(pipe, b1=1.1, b2=1.2, s1=0.6, s2=0.4)
+        register_free_crossattn_upblock2d(pipe, b1=1.1, b2=1.2, s1=0.6, s2=0.4)
 
     print('GPU starts')
-    result = infer_gpu_part(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_scale, resolutions_list, fast_mode, cosine_scale, disable_freeu)
+    result = infer_gpu_part(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_scale, resolutions_list, fast_mode, cosine_scale)
     print('GPU ends')
 
     image = result.images[0]

free_lunch_utils.py

@@ -30,7 +30,7 @@ def Fourier_filter(x, threshold, scale):
     x_freq = fft.fftshift(x_freq, dim=(-2, -1))
     
     B, C, H, W = x_freq.shape
-    mask = torch.ones((B, C, H, W)).cuda() 
+    mask = torch.ones((B, C, H, W)).to(x.device)
 
     crow, ccol = H // 2, W //2
     mask[..., crow - threshold:crow + threshold, ccol - threshold:ccol + threshold] = scale