update

app.py

@@ -236,7 +236,7 @@ def main(args):
                                     eval_data=None,
                                     **args)
     
-    @spaces.GPU(duration=15)
+    @spaces.GPU(duration=30)
     def reconstruct_and_export(*args, **kwargs):
         return training_loop_class.eval_i23d_and_export(*args, **kwargs)
 
@@ -355,7 +355,7 @@ def main(args):
         gr.Markdown(
             """
             ## Comments:
-            1. The sampling time varies since ODE-based sampling method (dopri5 by default) has adaptive internal step, and reducing sampling steps may not reduce the overal sampling time. Sampling steps=250 is the emperical value that works well in most cases.
+            1. Each sample shall take within 20 seconds on Gradio ZeroSpace, and 6~7 seconds locally (A100).
             2. The 3D viewer shows a colored .glb mesh extracted from volumetric tri-plane, and may differ slightly with the volume rendering result.
             3. If you find your result unsatisfying, tune the CFG scale and change the random seed. Usually slightly increase the CFG value can lead to better performance.
             3. Known limitations include:

nsr/lsgm/flow_matching_trainer.py

@@ -750,14 +750,14 @@ class FlowMatchingEngine(TrainLoop3DDiffusionLSGM_crossattn):
 
         def sample_and_save(batch_c):
 
-            with th.cuda.amp.autocast(dtype=self.dtype,
-                                        enabled=self.mp_trainer.use_amp):
+            # with th.cuda.amp.autocast(dtype=self.dtype,
+            #                             enabled=self.mp_trainer.use_amp):
 
-                c, uc = self.conditioner.get_unconditional_conditioning(
-                    batch_c,
-                    force_uc_zero_embeddings=ucg_keys
-                    if len(self.conditioner.embedders) > 0 else [],
-                )
+            c, uc = self.conditioner.get_unconditional_conditioning(
+                batch_c,
+                force_uc_zero_embeddings=ucg_keys
+                if len(self.conditioner.embedders) > 0 else [],
+            )
 
             for k in c:
                 if isinstance(c[k], th.Tensor):

nsr/train_util_diffusion.py

@@ -205,8 +205,8 @@ class TrainLoopDiffusionWithRec(TrainLoop):
             rec_model(latent=ddpm_latent,
                       behaviour='decode_after_vae_no_render'))
 
-        # if export_mesh:
-        if False:
+        if export_mesh:
+        # if False:
             import mcubes
             import trimesh
             dump_path = f'{logger.get_dir()}/mesh/'
@@ -374,7 +374,7 @@ class TrainLoopDiffusionWithRec(TrainLoop):
         print('logged video to: ', f'{vid_dump_path}')
 
         del vis, pred_vis, micro, pred,
-        mesh_dump_path = 'logs/LSGM/inference/Objaverse/i23d/dit-L2/gradio_app/mesh/cfg=4.0_sample-0-rotatex.obj'
+        # mesh_dump_path = 'logs/LSGM/inference/Objaverse/i23d/dit-L2/gradio_app/mesh/cfg=4.0_sample-0-rotatex.obj'
 
         return vid_dump_path, mesh_dump_path