wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-01-25 13:50:15 +08:00
Code Issues Actions 15 Packages Projects Releases Wiki Activity

Add support for live previews on separate stream

Browse Source
This commit is contained in:
drhead 2025-05-23 18:12:10 -04:00 committed by GitHub
parent 0b50d4c0db
commit e7cc72cee0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 36 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

50
latent_preview.py
View File

@ -6,18 +6,29 @@ import comfy.model_management
import folder_paths import folder_paths
import comfy.utils import comfy.utils
import logging import logging
from contextlib import nullcontext
import threading
MAX_PREVIEW_RESOLUTION = args.preview_size MAX_PREVIEW_RESOLUTION = args.preview_size
def preview_to_image(latent_image): if args.preview_stream:
latents_ubyte = (((latent_image + 1.0) / 2.0).clamp(0, 1) # change scale from -1..1 to 0..1 preview_stream = torch.cuda.Stream()
.mul(0xFF) # to 0..255 preview_context = torch.cuda.stream(preview_stream)
) else:
if comfy.model_management.directml_enabled: preview_context = nullcontext()
latents_ubyte = latents_ubyte.to(dtype=torch.uint8)
latents_ubyte = latents_ubyte.to(device="cpu", dtype=torch.uint8, non_blocking=comfy.model_management.device_supports_non_blocking(latent_image.device))
return Image.fromarray(latents_ubyte.numpy()) def preview_to_image(preview_image: torch.Tensor):
# no reason why any of this has to happen on GPU, also non-blocking transfers to cpu aren't safe ever
# but we don't care about it blocking because the main stream is fine
preview_image = preview_image.cpu()
preview_image.add_(1.0)
preview_image.div_(2.0)
preview_image.clamp_(0, 1) # change scale from -1..1 to 0..1 and clamp
preview_image.mul_(255.) # change to uint8 range
preview_image.round_() # default behavior when casting is truncate which is wrong for image processing
return Image.fromarray(preview_image.to(dtype=torch.uint8).numpy())
class LatentPreviewer: class LatentPreviewer:
def decode_latent_to_preview(self, x0): def decode_latent_to_preview(self, x0):
@ -97,12 +108,23 @@ def prepare_callback(model, steps, x0_output_dict=None):
pbar = comfy.utils.ProgressBar(steps) pbar = comfy.utils.ProgressBar(steps)
def callback(step, x0, x, total_steps): def callback(step, x0, x, total_steps):
if x0_output_dict is not None: @torch.inference_mode
x0_output_dict["x0"] = x0 def worker():
if x0_output_dict is not None:
x0_output_dict["x0"] = x0
preview_bytes = None
if previewer:
with preview_context:
preview_bytes = previewer.decode_latent_to_preview_image(preview_format, x0)
pbar.update_absolute(step + 1, total_steps, preview_bytes)
if args.preview_stream:
# must wait for default stream to catch up else we will decode a garbage tensor
# the default stream will not, under any circumstances, stop because of this
preview_stream.wait_stream(torch.cuda.default_stream())
threading.Thread(target=worker, daemon=True).start()
else: worker() # no point in threading this off if there's no separate stream
preview_bytes = None
if previewer:
preview_bytes = previewer.decode_latent_to_preview_image(preview_format, x0)
pbar.update_absolute(step + 1, total_steps, preview_bytes)
return callback return callback