DanielBartolicandClaude Opus 4.5
c462bb152b
- Detect flow models and use calculate_denoised method - Add latent_format processing - Better error handling for 16-channel latents - Support for QwenImage 2512 Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
422 lines
16 KiB
Python
422 lines
16 KiB
Python
"""
|
|
Save Intermediate Steps Node
|
|
Saves preview images at each sampling step for streaming to frontend
|
|
"""
|
|
|
|
import torch
|
|
import os
|
|
import json
|
|
import folder_paths
|
|
import comfy.model_management
|
|
from PIL import Image
|
|
import numpy as np
|
|
from typing import Optional
|
|
import time
|
|
import base64
|
|
from io import BytesIO
|
|
|
|
|
|
def get_previewer(model):
|
|
"""Get previewer for latent-to-image conversion"""
|
|
try:
|
|
import latent_preview
|
|
device = comfy.model_management.get_torch_device()
|
|
latent_format = model.model.latent_format
|
|
previewer = latent_preview.get_previewer(device, latent_format)
|
|
return previewer
|
|
except Exception as e:
|
|
print(f"[SaveIntermediates] Warning: Could not load previewer: {e}")
|
|
return None
|
|
|
|
|
|
class IntermediateStreamer:
|
|
"""Handles saving/streaming intermediate latents as images"""
|
|
|
|
def __init__(self, output_dir: str, prefix: str, format: str, previewer,
|
|
latent_format, job_id: str = "", save_base64: bool = False, total_steps: int = 20):
|
|
self.output_dir = output_dir
|
|
self.prefix = prefix
|
|
self.format = format.lower()
|
|
self.previewer = previewer
|
|
self.latent_format = latent_format
|
|
self.job_id = job_id
|
|
self.save_base64 = save_base64
|
|
self.total_steps = total_steps
|
|
self.step_count = 0
|
|
self.progress_file = os.path.join(output_dir, "progress.json")
|
|
|
|
os.makedirs(output_dir, exist_ok=True)
|
|
self._update_progress(0, None, "starting")
|
|
|
|
def _update_progress(self, step: int, image_path: Optional[str], status: str, base64_data: str = None):
|
|
"""Update progress.json for frontend polling"""
|
|
progress = {
|
|
"job_id": self.job_id,
|
|
"status": status,
|
|
"current_step": step,
|
|
"total_steps": self.total_steps,
|
|
"progress_percent": int((step / max(self.total_steps, 1)) * 100),
|
|
"latest_image": image_path,
|
|
"timestamp": time.time(),
|
|
}
|
|
if base64_data and self.save_base64:
|
|
progress["image_base64"] = base64_data
|
|
|
|
try:
|
|
with open(self.progress_file, 'w') as f:
|
|
json.dump(progress, f)
|
|
except Exception as e:
|
|
print(f"[SaveIntermediates] Error writing progress: {e}")
|
|
|
|
def save_step(self, latent: torch.Tensor, step: int):
|
|
"""Save a single step's latent as an image"""
|
|
try:
|
|
if self.previewer is None:
|
|
print(f"[SaveIntermediates] No previewer available")
|
|
return None
|
|
|
|
# Ensure proper shape: [B, C, H, W]
|
|
if latent.dim() == 3:
|
|
latent = latent.unsqueeze(0)
|
|
elif latent.dim() == 5:
|
|
# Video latent [B, F, C, H, W] - take first frame
|
|
latent = latent[:, 0, :, :, :]
|
|
|
|
# Take first batch item, ensure float32, keep on same device
|
|
latent_sample = latent[0:1].clone().to(torch.float32)
|
|
|
|
# Process through latent format if available
|
|
if self.latent_format is not None:
|
|
try:
|
|
latent_sample = self.latent_format.process_out(latent_sample)
|
|
except:
|
|
pass
|
|
|
|
# Decode using previewer
|
|
preview_result = self.previewer.decode_latent_to_preview_image(
|
|
"JPEG" if self.format == "jpeg" else "PNG",
|
|
latent_sample
|
|
)
|
|
|
|
if preview_result and len(preview_result) > 1:
|
|
img = preview_result[1]
|
|
|
|
filename = f"{self.prefix}_{step:04d}.{self.format}"
|
|
filepath = os.path.join(self.output_dir, filename)
|
|
img.save(filepath, quality=85 if self.format == "jpeg" else None)
|
|
|
|
latest_path = os.path.join(self.output_dir, f"latest.{self.format}")
|
|
img.save(latest_path, quality=85 if self.format == "jpeg" else None)
|
|
|
|
base64_data = None
|
|
if self.save_base64:
|
|
buffer = BytesIO()
|
|
img.save(buffer, format=self.format.upper(), quality=85)
|
|
base64_data = base64.b64encode(buffer.getvalue()).decode('utf-8')
|
|
|
|
self.step_count += 1
|
|
self._update_progress(step, filename, "generating", base64_data)
|
|
return filepath
|
|
|
|
except Exception as e:
|
|
print(f"[SaveIntermediates] Error saving step {step}: {e}")
|
|
|
|
return None
|
|
|
|
def finalize(self):
|
|
"""Mark generation as complete"""
|
|
self._update_progress(self.total_steps, f"{self.prefix}_final.{self.format}", "completed")
|
|
|
|
|
|
class SaveIntermediateSteps:
|
|
"""
|
|
Saves intermediate sampling steps for streaming to frontend.
|
|
Works with Flux, SD3.5, SDXL, SD1.5, etc.
|
|
"""
|
|
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"model": ("MODEL",),
|
|
"steps": ("INT", {"default": 20, "min": 1, "max": 1000}),
|
|
},
|
|
"optional": {
|
|
"job_id": ("STRING", {"default": ""}),
|
|
"output_folder": ("STRING", {"default": "progress"}),
|
|
"filename_prefix": ("STRING", {"default": "step"}),
|
|
"format": (["jpeg", "png"], {"default": "jpeg"}),
|
|
"include_base64": ("BOOLEAN", {"default": False}),
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("MODEL",)
|
|
RETURN_NAMES = ("model",)
|
|
FUNCTION = "wrap_model"
|
|
CATEGORY = "sampling/streaming"
|
|
|
|
def wrap_model(self, model, steps: int, job_id: str = "", output_folder: str = "progress",
|
|
filename_prefix: str = "step", format: str = "jpeg", include_base64: bool = False):
|
|
|
|
base_output = folder_paths.get_output_directory()
|
|
if job_id:
|
|
full_output_dir = os.path.join(base_output, output_folder, job_id)
|
|
else:
|
|
timestamp = time.strftime("%Y%m%d_%H%M%S")
|
|
full_output_dir = os.path.join(base_output, output_folder, timestamp)
|
|
|
|
previewer = get_previewer(model)
|
|
latent_format = model.model.latent_format if hasattr(model.model, 'latent_format') else None
|
|
|
|
streamer = IntermediateStreamer(
|
|
output_dir=full_output_dir,
|
|
prefix=filename_prefix,
|
|
format=format,
|
|
previewer=previewer,
|
|
latent_format=latent_format,
|
|
job_id=job_id,
|
|
save_base64=include_base64,
|
|
total_steps=steps
|
|
)
|
|
|
|
# Clone model
|
|
wrapped_model = model.clone()
|
|
|
|
# Check model type
|
|
model_sampling = wrapped_model.model.model_sampling if hasattr(wrapped_model.model, 'model_sampling') else None
|
|
is_flow_model = hasattr(model_sampling, 'calculate_denoised') if model_sampling else False
|
|
|
|
# Get original apply_model
|
|
original_apply_model = wrapped_model.model.apply_model
|
|
step_counter = [0]
|
|
|
|
def patched_apply_model(x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
|
|
# Call original
|
|
model_output = original_apply_model(x, t, c_concat, c_crossattn, control, transformer_options, **kwargs)
|
|
|
|
try:
|
|
# Calculate denoised based on model type
|
|
if model_sampling is not None and is_flow_model:
|
|
# For Flux/Flow models - use the model's own denoising calculation
|
|
try:
|
|
sigma = model_sampling.sigma(t)
|
|
denoised = model_sampling.calculate_denoised(sigma, model_output, x)
|
|
except:
|
|
# Fallback for flow matching: denoised = x - (1-sigma) * v
|
|
sigma = model_sampling.sigma(t)
|
|
if sigma.dim() == 0:
|
|
sigma = sigma.unsqueeze(0)
|
|
while sigma.dim() < x.dim():
|
|
sigma = sigma.unsqueeze(-1)
|
|
# For rectified flow: x0 = x - sigma * v
|
|
denoised = x - sigma * model_output
|
|
elif model_sampling is not None:
|
|
# For eps-prediction models (SD1.5, SDXL, etc.)
|
|
sigma = model_sampling.sigma(t)
|
|
if sigma.dim() == 0:
|
|
sigma = sigma.unsqueeze(0)
|
|
while sigma.dim() < x.dim():
|
|
sigma = sigma.unsqueeze(-1)
|
|
# x0 = x - sigma * eps
|
|
denoised = x - sigma * model_output
|
|
else:
|
|
# Fallback: use x directly
|
|
denoised = x
|
|
|
|
# Save the preview
|
|
if denoised.dim() == 4:
|
|
streamer.save_step(denoised, step_counter[0])
|
|
elif denoised.dim() == 5:
|
|
streamer.save_step(denoised[:, 0], step_counter[0])
|
|
|
|
step_counter[0] += 1
|
|
|
|
except Exception as e:
|
|
print(f"[SaveIntermediates] Error at step {step_counter[0]}: {e}")
|
|
step_counter[0] += 1
|
|
|
|
return model_output
|
|
|
|
wrapped_model.model.apply_model = patched_apply_model
|
|
wrapped_model._intermediate_streamer = streamer
|
|
|
|
print(f"[SaveIntermediates] Streaming to: {full_output_dir}")
|
|
print(f"[SaveIntermediates] Poll progress at: {full_output_dir}/progress.json")
|
|
print(f"[SaveIntermediates] Flow model detected: {is_flow_model}")
|
|
|
|
return (wrapped_model,)
|
|
|
|
|
|
class SaveIntermediateStepsAdvanced:
|
|
"""
|
|
Advanced version with step filtering.
|
|
"""
|
|
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"model": ("MODEL",),
|
|
"steps": ("INT", {"default": 20, "min": 1, "max": 1000}),
|
|
"save_every_n": ("INT", {"default": 1, "min": 1, "max": 100}),
|
|
},
|
|
"optional": {
|
|
"job_id": ("STRING", {"default": ""}),
|
|
"output_folder": ("STRING", {"default": "progress"}),
|
|
"filename_prefix": ("STRING", {"default": "step"}),
|
|
"format": (["jpeg", "png"], {"default": "jpeg"}),
|
|
"include_base64": ("BOOLEAN", {"default": False}),
|
|
"start_step": ("INT", {"default": 0, "min": 0, "max": 1000}),
|
|
"end_step": ("INT", {"default": 999, "min": 0, "max": 1000}),
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("MODEL",)
|
|
RETURN_NAMES = ("model",)
|
|
FUNCTION = "wrap_model"
|
|
CATEGORY = "sampling/streaming"
|
|
|
|
def wrap_model(self, model, steps: int, save_every_n: int, job_id: str = "",
|
|
output_folder: str = "progress", filename_prefix: str = "step",
|
|
format: str = "jpeg", include_base64: bool = False,
|
|
start_step: int = 0, end_step: int = 999):
|
|
|
|
base_output = folder_paths.get_output_directory()
|
|
if job_id:
|
|
full_output_dir = os.path.join(base_output, output_folder, job_id)
|
|
else:
|
|
timestamp = time.strftime("%Y%m%d_%H%M%S")
|
|
full_output_dir = os.path.join(base_output, output_folder, timestamp)
|
|
|
|
previewer = get_previewer(model)
|
|
latent_format = model.model.latent_format if hasattr(model.model, 'latent_format') else None
|
|
|
|
streamer = IntermediateStreamer(
|
|
output_dir=full_output_dir,
|
|
prefix=filename_prefix,
|
|
format=format,
|
|
previewer=previewer,
|
|
latent_format=latent_format,
|
|
job_id=job_id,
|
|
save_base64=include_base64,
|
|
total_steps=steps
|
|
)
|
|
|
|
wrapped_model = model.clone()
|
|
model_sampling = wrapped_model.model.model_sampling if hasattr(wrapped_model.model, 'model_sampling') else None
|
|
is_flow_model = hasattr(model_sampling, 'calculate_denoised') if model_sampling else False
|
|
original_apply_model = wrapped_model.model.apply_model
|
|
step_counter = [0]
|
|
|
|
def patched_apply_model(x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
|
|
model_output = original_apply_model(x, t, c_concat, c_crossattn, control, transformer_options, **kwargs)
|
|
|
|
current_step = step_counter[0]
|
|
should_save = (
|
|
current_step >= start_step and
|
|
current_step <= end_step and
|
|
(current_step - start_step) % save_every_n == 0
|
|
)
|
|
|
|
if should_save:
|
|
try:
|
|
if model_sampling is not None and is_flow_model:
|
|
try:
|
|
sigma = model_sampling.sigma(t)
|
|
denoised = model_sampling.calculate_denoised(sigma, model_output, x)
|
|
except:
|
|
sigma = model_sampling.sigma(t)
|
|
if sigma.dim() == 0:
|
|
sigma = sigma.unsqueeze(0)
|
|
while sigma.dim() < x.dim():
|
|
sigma = sigma.unsqueeze(-1)
|
|
denoised = x - sigma * model_output
|
|
elif model_sampling is not None:
|
|
sigma = model_sampling.sigma(t)
|
|
if sigma.dim() == 0:
|
|
sigma = sigma.unsqueeze(0)
|
|
while sigma.dim() < x.dim():
|
|
sigma = sigma.unsqueeze(-1)
|
|
denoised = x - sigma * model_output
|
|
else:
|
|
denoised = x
|
|
|
|
if denoised.dim() == 4:
|
|
streamer.save_step(denoised, current_step)
|
|
elif denoised.dim() == 5:
|
|
streamer.save_step(denoised[:, 0], current_step)
|
|
|
|
except Exception as e:
|
|
print(f"[SaveIntermediates] Error at step {current_step}: {e}")
|
|
|
|
step_counter[0] += 1
|
|
return model_output
|
|
|
|
wrapped_model.model.apply_model = patched_apply_model
|
|
wrapped_model._intermediate_streamer = streamer
|
|
|
|
print(f"[SaveIntermediates] Streaming to: {full_output_dir}")
|
|
|
|
return (wrapped_model,)
|
|
|
|
|
|
class GetIntermediateImages:
|
|
"""
|
|
Collects all intermediate images after sampling completes.
|
|
Connect after your sampler to get all steps as IMAGE batch.
|
|
"""
|
|
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"latent": ("LATENT",),
|
|
},
|
|
"optional": {
|
|
"job_id": ("STRING", {"default": ""}),
|
|
"output_folder": ("STRING", {"default": "progress"}),
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("IMAGE",)
|
|
RETURN_NAMES = ("images",)
|
|
FUNCTION = "collect_images"
|
|
CATEGORY = "sampling/streaming"
|
|
|
|
def collect_images(self, latent, job_id: str = "", output_folder: str = "progress"):
|
|
"""Collect all saved intermediate images as a batch"""
|
|
|
|
base_output = folder_paths.get_output_directory()
|
|
|
|
if job_id:
|
|
search_dir = os.path.join(base_output, output_folder, job_id)
|
|
else:
|
|
progress_dir = os.path.join(base_output, output_folder)
|
|
if os.path.exists(progress_dir):
|
|
folders = sorted([f for f in os.listdir(progress_dir) if os.path.isdir(os.path.join(progress_dir, f))])
|
|
if folders:
|
|
search_dir = os.path.join(progress_dir, folders[-1])
|
|
else:
|
|
return (torch.zeros(1, 64, 64, 3),)
|
|
else:
|
|
return (torch.zeros(1, 64, 64, 3),)
|
|
|
|
images = []
|
|
if os.path.exists(search_dir):
|
|
files = sorted([f for f in os.listdir(search_dir) if f.startswith("step_") and (f.endswith(".jpeg") or f.endswith(".png"))])
|
|
|
|
for filename in files:
|
|
filepath = os.path.join(search_dir, filename)
|
|
try:
|
|
img = Image.open(filepath).convert("RGB")
|
|
img_array = np.array(img).astype(np.float32) / 255.0
|
|
images.append(torch.from_numpy(img_array))
|
|
except Exception as e:
|
|
print(f"[GetIntermediateImages] Error loading {filename}: {e}")
|
|
|
|
if images:
|
|
batch = torch.stack(images, dim=0)
|
|
return (batch,)
|
|
else:
|
|
return (torch.zeros(1, 64, 64, 3),)
|