Fred Bliss
4a091e9298
- Fixed generateTokens to properly handle different region types (bounding_box, polygon, object_reference) - Added checkbox UI to control includeObjectRef for boxes/polygons - Object references now correctly handle 2 coordinates instead of expecting 4 - Updated Python code to respect includeObjectRef flag from JavaScript - Fixed coordinate format issues (x1,y1,x2,y2 for boxes, x,y pairs for polygons) - Restored updateBasePromptWithTokens function for editable prompts - Default behavior keeps object_ref (checkbox checked) for backward compatibility - Without object_ref, generates pure spatial tokens like <|box_start|>0.1,0.2,0.3,0.4<|box_end|>
86 lines
3.2 KiB
Python
86 lines
3.2 KiB
Python
#!/usr/bin/env python3
|
|
"""
|
|
Debug script to understand why optimize_resolution=True causes issues with empty latents
|
|
"""
|
|
|
|
import math
|
|
|
|
# Qwen resolutions from the node
|
|
QWEN_RESOLUTIONS = [
|
|
(256, 256), (256, 512), (256, 768), (256, 1024), (256, 1280), (256, 1536), (256, 1792),
|
|
(512, 256), (512, 512), (512, 768), (512, 1024), (512, 1280), (512, 1536), (512, 1792),
|
|
(768, 256), (768, 512), (768, 768), (768, 1024), (768, 1280), (768, 1536),
|
|
(1024, 256), (1024, 512), (1024, 768), (1024, 1024), (1024, 1280), (1024, 1536),
|
|
(1280, 256), (1280, 512), (1280, 768), (1280, 1024), (1280, 1280),
|
|
(1536, 256), (1536, 512), (1536, 768), (1536, 1024),
|
|
(1792, 256), (1792, 512)
|
|
]
|
|
|
|
def get_optimal_resolution(width: int, height: int) -> tuple:
|
|
"""Find the nearest Qwen-supported resolution"""
|
|
target_pixels = width * height
|
|
|
|
# Find closest resolution by total pixels
|
|
best_res = min(
|
|
QWEN_RESOLUTIONS,
|
|
key=lambda r: abs(r[0] * r[1] - target_pixels)
|
|
)
|
|
|
|
return best_res
|
|
|
|
def get_standard_resolution(width: int, height: int) -> tuple:
|
|
"""Scale to approximately 1M pixels (1024x1024)"""
|
|
total = int(1024 * 1024)
|
|
scale_by = math.sqrt(total / (width * height))
|
|
new_width = round(width * scale_by)
|
|
new_height = round(height * scale_by)
|
|
return new_width, new_height
|
|
|
|
# Test with common image sizes
|
|
test_sizes = [
|
|
(512, 512), # Square small
|
|
(1024, 1024), # Square medium
|
|
(2048, 2048), # Square large
|
|
(1920, 1080), # 16:9 landscape
|
|
(1080, 1920), # 9:16 portrait
|
|
(768, 1024), # 3:4 portrait
|
|
(1024, 768), # 4:3 landscape
|
|
]
|
|
|
|
print("Resolution comparison for different input sizes:")
|
|
print("=" * 80)
|
|
|
|
for orig_w, orig_h in test_sizes:
|
|
orig_pixels = orig_w * orig_h
|
|
|
|
# Optimal Qwen resolution
|
|
opt_w, opt_h = get_optimal_resolution(orig_w, orig_h)
|
|
opt_pixels = opt_w * opt_h
|
|
opt_ratio = opt_pixels / orig_pixels
|
|
|
|
# Standard 1M pixel resolution
|
|
std_w, std_h = get_standard_resolution(orig_w, orig_h)
|
|
std_pixels = std_w * std_h
|
|
std_ratio = std_pixels / orig_pixels
|
|
|
|
print(f"\nOriginal: {orig_w}x{orig_h} ({orig_pixels:,} pixels)")
|
|
print(f" Optimal (Qwen): {opt_w}x{opt_h} ({opt_pixels:,} pixels) - {opt_ratio:.2f}x scale")
|
|
print(f" Standard (1M): {std_w}x{std_h} ({std_pixels:,} pixels) - {std_ratio:.2f}x scale")
|
|
|
|
# Calculate latent dimensions (assuming 8x downscale for VAE)
|
|
opt_latent_h, opt_latent_w = opt_h // 8, opt_w // 8
|
|
std_latent_h, std_latent_w = std_h // 8, std_w // 8
|
|
|
|
print(f" Optimal latent: {opt_latent_w}x{opt_latent_h}")
|
|
print(f" Standard latent: {std_latent_w}x{std_latent_h}")
|
|
|
|
# Check if dimensions differ significantly
|
|
if abs(opt_pixels - std_pixels) > 100000:
|
|
print(f" ⚠️ Large pixel difference: {abs(opt_pixels - std_pixels):,}")
|
|
|
|
print("\n" + "=" * 80)
|
|
print("\nKey observations:")
|
|
print("1. Optimal resolution snaps to predefined Qwen sizes")
|
|
print("2. Standard resolution maintains aspect ratio while targeting 1M pixels")
|
|
print("3. Different resolutions create different-sized reference latents")
|
|
print("4. Larger reference latents may have stronger influence on generation") |