optimize

app.py

@@ -211,7 +211,7 @@ def optimize_stage_1(image_block: Image.Image, preprocess_chk: bool, seed_slider
 
     # stage 1
     # subprocess.run(f'python lgm/infer.py big --resume {ckpt_path} --test_path tmp_data/{img_hash}_rgba.png', shell=True)
-    process_lgm(opt, f'tmp_data/{img_hash}_rgba.png', pipe_mvdream, model, rays_embeddings)
+    process_lgm(opt, f'tmp_data/{img_hash}_rgba.png', pipe_mvdream, model, rays_embeddings, seed_slider)
     # return [os.path.join('logs', 'tmp_rgba_model.ply')]
     return os.path.join('vis_data', f'{img_hash}_rgba_static.mp4')
 

lgm/infer_demo.py

@@ -47,7 +47,7 @@ IMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)
 
 
 # process function
-def process(opt: Options, path, pipe, model, rays_embeddings):
+def process(opt: Options, path, pipe, model, rays_embeddings, seed):
     device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
     tan_half_fov = np.tan(0.5 * np.deg2rad(opt.fovy))
     proj_matrix = torch.zeros(4, 4, dtype=torch.float32, device=device)
@@ -72,7 +72,8 @@ def process(opt: Options, path, pipe, model, rays_embeddings):
     if image.shape[-1] == 4:
         image = image[..., :3] * image[..., 3:4] + (1 - image[..., 3:4])
 
-    mv_image = pipe('', image, guidance_scale=5.0, num_inference_steps=30, elevation=0)
+    generator = torch.manual_seed(seed)
+    mv_image = pipe('', image, guidance_scale=5.0, num_inference_steps=30, elevation=0, generator=generator)
     mv_image = np.stack([mv_image[1], mv_image[2], mv_image[3], mv_image[0]], axis=0) # [4, 256, 256, 3], float32
 
     # generate gaussians

main_4d_demo.py

@@ -571,7 +571,7 @@ class GUI:
                     hor = (hor+delta_hor) % 360
 
 
-        imageio.mimwrite(f'vis_data/{opt.save_path}.mp4', image_list, fps=7)
+        imageio.mimwrite(f'vis_data/{self.opt.save_path}.mp4', image_list, fps=7)
 
         if self.gui:
             while True: