Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Add progress to vae decode tiled

pythongosssssc8c9926eeb0b
diff --git a/comfy/sd.py b/comfy/sd.py
index 92dbb931d..2aadefadc 100644
--- a/comfy/sd.py
+++ b/comfy/sd.py
@@ -1,6 +1,7 @@
import torch
import contextlib
import copy
+from tqdm.auto import tqdm
import sd1_clip
import sd2_clip
@@ -437,11 +438,16 @@ class VAE:
self.device = device
def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16):
+ it_1 = -(samples.shape[2] // -(tile_y * 2 - overlap)) * -(samples.shape[3] // -(tile_x // 2 - overlap))
+ it_2 = -(samples.shape[2] // -(tile_y // 2 - overlap)) * -(samples.shape[3] // -(tile_x * 2 - overlap))
+ it_3 = -(samples.shape[2] // -(tile_y - overlap)) * -(samples.shape[3] // -(tile_x - overlap))
+ pbar = tqdm(total=samples.shape[0] * (it_1 + it_2 + it_3))
+
decode_fn = lambda a: (self.first_stage_model.decode(1. / self.scale_factor * a.to(self.device)) + 1.0)
output = torch.clamp((
- (utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = 8) +
- utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = 8) +
- utils.tiled_scale(samples, decode_fn, tile_x, tile_y, overlap, upscale_amount = 8))
+ (utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = 8, pbar = pbar) +
+ utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = 8, pbar = pbar) +
+ utils.tiled_scale(samples, decode_fn, tile_x, tile_y, overlap, upscale_amount = 8, pbar = pbar))
/ 3.0) / 2.0, min=0.0, max=1.0)
return output
diff --git a/comfy/utils.py b/comfy/utils.py
index 68f93403c..c7c6a08c5 100644
--- a/comfy/utils.py
+++ b/comfy/utils.py
@@ -63,7 +63,7 @@ def common_upscale(samples, width, height, upscale_method, crop):
return torch.nn.functional.interpolate(s, size=(height, width), mode=upscale_method)
@torch.inference_mode()
-def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3):
+def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3, pbar = None):
output = torch.empty((samples.shape[0], out_channels, round(samples.shape[2] * upscale_amount), round(samples.shape[3] * upscale_amount)), device="cpu")
for b in range(samples.shape[0]):
s = samples[b:b+1]
@@ -83,6 +83,8 @@ def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_am
mask[:,:,:,mask.shape[3]- 1 - t: mask.shape[3]- t] *= ((1.0/feather) * (t + 1))
out[:,:,round(y*upscale_amount):round((y+tile_y)*upscale_amount),round(x*upscale_amount):round((x+tile_x)*upscale_amount)] += ps * mask
out_div[:,:,round(y*upscale_amount):round((y+tile_y)*upscale_amount),round(x*upscale_amount):round((x+tile_x)*upscale_amount)] += mask
+ if pbar is not None:
+ pbar.update(1)
output[b:b+1] = out/out_div
return output