various inference scripts for newer models

firered_1536_sweep.py

import os
import torch
from pathlib import Path
from PIL import Image
from diffusers import QwenImageEditPlusPipeline

def load_pipeline(model_path: str) -> QwenImageEditPlusPipeline:
    """Load FireRed image edit pipeline."""
    pipe = QwenImageEditPlusPipeline.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        cache_dir="./models",
    )
    pipe.to("cuda")
    return pipe

def create_grid(images, rows, cols):
    """Helper to create a single large image grid."""
    w, h = images[0].size
    grid = Image.new('RGB', size=(cols * w, rows * h))
    for i, img in enumerate(images):
        grid.paste(img, box=(i % cols * w, i // cols * h))
    return grid

def main():
    # --- Configuration ---
    MODEL_ID = "FireRedTeam/FireRed-Image-Edit-1.0"
    INPUT_FILE_DRESS = "together_raw_1024.png"
    INPUT_FILE_WOMAN = "woman_upscaled.png"
    PROMPT = "Dress hero shot. A close-up detail editorial photograph focusing on the waist and skirt section of the reference dress worn by the woman in a minimalist studio. Background is seamless off-white #F5F5F5. The shot highlights the fabric drape and stitching details. Soft studio light. Kodak film 400 grain."
    
    RESOLUTION = 1536
    CFG_VALUES = [1.0, 3.0]
    SEEDS = [10, 20, 30, 40, 50, 60, 70, 80] # 8 distinct seeds
    
    # Setup Output
    output_dir = Path("final_shots")
    output_dir.mkdir(exist_ok=True)

    # Load Pipeline
    pipeline = load_pipeline(MODEL_ID)

    # Load Inputs
    dress_image = Image.open(INPUT_FILE_DRESS).convert("RGB")
    woman_image = Image.open(INPUT_FILE_WOMAN).convert("RGB")

    for cfg in CFG_VALUES:
        cfg_images = []
        print(f"\n--- Starting Sweep for CFG: {cfg} ---")
        
        for seed in SEEDS:
            file_name = f"cfg{cfg}_seed{seed}.png"
            save_path = output_dir / file_name
            
            print(f"Generating: {file_name}")
            
            inputs = {
                "image": [dress_image, woman_image],
                "prompt": PROMPT,
                "generator": torch.Generator(device="cuda").manual_seed(seed),
                "height": RESOLUTION+512,
                "width": RESOLUTION,
                "true_cfg_scale": cfg,
                "negative_prompt": "",
                "num_inference_steps": 40,
                "num_images_per_prompt": 1,
            }

            with torch.inference_mode():
                result = pipeline(**inputs)
                img = result.images[0]
            
            # Save individual image
            img.save(save_path)
            cfg_images.append(img)

        # Create and save the grid for this CFG (2 rows of 4 images)
        grid_img = create_grid(cfg_images, rows=2, cols=4)
        grid_path = output_dir / f"GRID_cfg{cfg}_all_seeds.jpg"
        grid_img.save(grid_path, quality=95)
        print(f"Grid saved for CFG {cfg} at: {grid_path}")

    print(f"\nAll tasks complete. Check the '{output_dir}' folder.")

if __name__ == "__main__":
    main()

firered_1536_sweeps_higher_res_input.py

import os
import torch
from pathlib import Path
from PIL import Image
from diffusers import QwenImageEditPlusPipeline

def load_pipeline(model_path: str) -> QwenImageEditPlusPipeline:
    """Load FireRed image edit pipeline."""
    pipe = QwenImageEditPlusPipeline.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        cache_dir="./models",
    )
    pipe.to("cuda")
    return pipe

def create_grid(images, rows, cols):
    """Helper to create a single large image grid."""
    w, h = images[0].size
    grid = Image.new('RGB', size=(cols * w, rows * h))
    for i, img in enumerate(images):
        grid.paste(img, box=(i % cols * w, i // cols * h))
    return grid

def main():
    # --- Configuration ---
    MODEL_ID = "FireRedTeam/FireRed-Image-Edit-1.0"
    INPUT_FILE_DRESS = "together_raw_1024.png"
    INPUT_FILE_WOMAN = "woman_upscaled.png"
    PROMPT = "Dress the model in the one-piece dress from Image 1. Match the same style and texture as in Image 1. The woman is in a minimalist studio. Her hair is fastened up. Background is seamless off-white #F5F5F5. The sharp focus is specifically on the fabric texture, neckline, and details of the reference dress. Soft studio light. Kodak film 400 grain."
    
    RESOLUTION = 1536
    CFG_VALUES = [1.0, 3.0]
    SEEDS = [10, 20, 30, 40, 50, 60, 70, 80] # 8 distinct seeds
    
    # Setup Output
    output_dir = Path("final_shots_2")
    output_dir.mkdir(exist_ok=True)

    # Load Pipeline
    pipeline = load_pipeline(MODEL_ID)

    # Load Inputs
    dress_image = Image.open(INPUT_FILE_DRESS).convert("RGB")
    woman_image = Image.open(INPUT_FILE_WOMAN).convert("RGB")

    for cfg in CFG_VALUES:
        cfg_images = []
        print(f"\n--- Starting Sweep for CFG: {cfg} ---")
        
        for seed in SEEDS:
            file_name = f"cfg{cfg}_seed{seed}.png"
            save_path = output_dir / file_name
            
            print(f"Generating: {file_name}")
            
            inputs = {
                "image": [dress_image, woman_image],
                "prompt": PROMPT,
                "generator": torch.Generator(device="cuda").manual_seed(seed),
                "height": RESOLUTION+512,
                "width": RESOLUTION,
                "true_cfg_scale": cfg,
                "negative_prompt": "",
                "num_inference_steps": 40,
                "num_images_per_prompt": 1,
            }

            with torch.inference_mode():
                result = pipeline(**inputs)
                img = result.images[0]
            
            # Save individual image
            img.save(save_path)
            cfg_images.append(img)

        # Create and save the grid for this CFG (2 rows of 4 images)
        grid_img = create_grid(cfg_images, rows=2, cols=4)
        grid_path = output_dir / f"GRID_cfg{cfg}_all_seeds.jpg"
        grid_img.save(grid_path, quality=95)
        print(f"Grid saved for CFG {cfg} at: {grid_path}")

    print(f"\nAll tasks complete. Check the '{output_dir}' folder.")

if __name__ == "__main__":
    main()

firered_sweeps.py

import argparse
import os
from pathlib import Path

import torch
from PIL import Image
from diffusers import QwenImageEditPlusPipeline

def load_pipeline(model_path: str) -> QwenImageEditPlusPipeline:
    """Load FireRed image edit pipeline."""
    pipe = QwenImageEditPlusPipeline.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        cache_dir="./models",
    )
    pipe.to("cuda")
    return pipe

def main():
    # --- Configuration ---
    MODEL_ID = "FireRedTeam/FireRed-Image-Edit-1.0"
    INPUT_FILE_DRESS = "dress.png"
    INPUT_FILE_WOMAN = "woman.png"
    PROMPT = "Dress hero shot. A close-up detail editorial photograph focusing on the waist and skirt section of the reference dress worn by the woman in a minimalist studio. Background is seamless off-white #F5F5F5. The shot highlights the fabric drape and stitching details. Soft studio light. Kodak film 400 grain."
    NEGATIVE_PROMPT = ""
    
    # Sweep Parameters
    RESOLUTIONS = [
        #1024,
        1280, 1536, 1792, 2048]
    CFG_SCALES = [1.0, 2.0, 3.0, 4.0, 5.0]
    
    # Setup Output
    output_dir = Path("sweep_results")
    output_dir.mkdir(exist_ok=True)

    # Load Pipeline
    pipeline = load_pipeline(MODEL_ID)
    print(f"Pipeline loaded onto {pipeline.device}")

    # Start Sweep
    dress_image = Image.open(INPUT_FILE_DRESS).convert("RGB")
    woman_image = Image.open(INPUT_FILE_WOMAN).convert("RGB")

    SEED = 12345
    for res in RESOLUTIONS:
        # Resize image to the target sweep resolution
        for cfg in CFG_SCALES:
            file_name = f"res{res}_cfg{cfg}.png"
            save_path = output_dir / file_name
            
            print(f"Generating: {file_name}")
            
            inputs = {
                "image": [dress_image, woman_image],
                "prompt": PROMPT,
                "generator": torch.Generator(device="cuda").manual_seed(SEED),
                "height": res,
                "width": res,
                "true_cfg_scale": cfg,
                "negative_prompt": NEGATIVE_PROMPT,
                "num_inference_steps": 40,
                "num_images_per_prompt": 1,
            }

            with torch.inference_mode():
                result = pipeline(**inputs)

            # Save result
            result.images[0].save(save_path)

    print(f"\nSweep complete! Results saved in: {output_dir.resolve()}")

if __name__ == "__main__":
    main()

inference_klein_9b_base.py

import os
import torch
from pathlib import Path
from PIL import Image
from diffusers import Flux2KleinPipeline

def load_pipeline(model_path: str) -> Flux2KleinPipeline:
    """Load FireRed image edit pipeline."""
    pipe = Flux2KleinPipeline.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        cache_dir="./models",
    )
    pipe.to("cuda")
    return pipe

def create_grid(images, rows, cols):
    """Helper to create a single large image grid."""
    w, h = images[0].size
    grid = Image.new('RGB', size=(cols * w, rows * h))
    for i, img in enumerate(images):
        grid.paste(img, box=(i % cols * w, i // cols * h))
    return grid

def main():
    # --- Configuration ---
    MODEL_ID = "black-forest-labs/FLUX.2-klein-base-9B"
    INPUT_FILE_ORIGINAL_IMAGE = "417_radiant_see_combined.png"
    INPUT_FILE_PERFUME_BOTTLE = "417_radiant_see.png"
    # PROMPT = "Dress the model in the one-piece dress from Image 1. Match the same style and texture as in Image 1. The woman is in a minimalist studio. Her hair is fastened up. Background is seamless off-white #F5F5F5. The sharp focus is specifically on the fabric texture, neckline, and details of the reference dress. Soft studio light. Kodak film 400 grain."
    PROMPT = "Add proper bokeh for the woman's face and make the perfume bottle clearer. The image features a luxury skincare serum in a sleek, tapered glass bottle with a sophisticated dark amber gradient that fades toward a thick, clear base. The bottle is topped with a tiered, metallic gold dropper cap that adds an elegant, high-end feel. All the text is printed in a matching gold serif font, beginning with the brand logo \"-417\" at the top, followed by the underlined collection name \"RADIANT SEE\" Near the bottom, the product is identified as the \"EXTREME GLOW NOURISHING FACE SERUM\" in clear, all-caps lettering."

    # RESOLUTION = 1536
    CFG_VALUES = [2.5, 3.5]
    SEEDS = [10, 20, 30, 40, 50, 60, 70, 80] # 8 distinct seeds
    
    # Setup Output
    output_dir = Path("klein_perfume_base")
    output_dir.mkdir(exist_ok=True)

    # Load Pipeline
    pipeline = load_pipeline(MODEL_ID)

    # Load Inputs
    input_image = Image.open(INPUT_FILE_ORIGINAL_IMAGE).convert("RGB")
    perfume_image = Image.open(INPUT_FILE_PERFUME_BOTTLE).convert("RGB")

    for cfg in CFG_VALUES:
        cfg_images = []
        print(f"\n--- Starting Sweep for CFG: {cfg} ---")
        
        for seed in SEEDS:
            file_name = f"cfg{cfg}_seed{seed}.png"
            save_path = output_dir / file_name
            
            print(f"Generating: {file_name}")
            w, h = input_image.size
            
            inputs = {
                "image": [input_image, perfume_image],
                "prompt": PROMPT,
                "generator": torch.Generator(device="cuda").manual_seed(seed),
                "height": h,
                "width": w,
                "guidance_scale": cfg,
                # "negative_prompt": "",
                "num_inference_steps": 50,
                "num_images_per_prompt": 1,
            }

            with torch.inference_mode():
                result = pipeline(**inputs)
                img = result.images[0]
            
            # Save individual image
            img.save(save_path)
            cfg_images.append(img)

        # Create and save the grid for this CFG (2 rows of 4 images)
        grid_img = create_grid(cfg_images, rows=2, cols=4)
        grid_path = output_dir / f"GRID_cfg{cfg}_all_seeds.jpg"
        grid_img.save(grid_path, quality=95)
        print(f"Grid saved for CFG {cfg} at: {grid_path}")

    print(f"\nAll tasks complete. Check the '{output_dir}' folder.")

if __name__ == "__main__":
    main()

inference_klein_9b_remake_perfume.py

import os
import torch
from pathlib import Path
from PIL import Image
from diffusers import Flux2KleinPipeline

def load_pipeline(model_path: str) -> Flux2KleinPipeline:
    """Load FireRed image edit pipeline."""
    pipe = Flux2KleinPipeline.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        cache_dir="./models",
    )
    pipe.to("cuda")
    return pipe

def create_grid(images, rows, cols):
    """Helper to create a single large image grid."""
    w, h = images[0].size
    grid = Image.new('RGB', size=(cols * w, rows * h))
    for i, img in enumerate(images):
        grid.paste(img, box=(i % cols * w, i // cols * h))
    return grid

def main():
    # --- Configuration ---
    MODEL_ID = "black-forest-labs/FLUX.2-klein-base-9B"
    INPUT_FILE_ORIGINAL_IMAGE = "face_for_perfume.png"
    INPUT_FILE_PERFUME_BOTTLE = "417_radiant_see.png"
    # PROMPT = "Dress the model in the one-piece dress from Image 1. Match the same style and texture as in Image 1. The woman is in a minimalist studio. Her hair is fastened up. Background is seamless off-white #F5F5F5. The sharp focus is specifically on the fabric texture, neckline, and details of the reference dress. Soft studio light. Kodak film 400 grain."
    PROMPT = "high definition product photography of this woman holding up this perfume bottle, centered on her face with her hand in frame holding it up close. you can clearly read the text on the perfume bottle."

    # RESOLUTION = 1536
    CFG_VALUES = [2.5, 3.5]
    SEEDS = [10, 20, 30, 40, 50, 60, 70, 80] # 8 distinct seeds
    
    # Setup Output
    output_dir = Path("klein_perfume_base_remake")
    output_dir.mkdir(exist_ok=True)

    # Load Pipeline
    pipeline = load_pipeline(MODEL_ID)

    # Load Inputs
    input_image = Image.open(INPUT_FILE_ORIGINAL_IMAGE).convert("RGB")
    perfume_image = Image.open(INPUT_FILE_PERFUME_BOTTLE).convert("RGB")

    for cfg in CFG_VALUES:
        cfg_images = []
        print(f"\n--- Starting Sweep for CFG: {cfg} ---")
        
        for seed in SEEDS:
            file_name = f"cfg{cfg}_seed{seed}.png"
            save_path = output_dir / file_name
            
            print(f"Generating: {file_name}")
            w, h = input_image.size
            
            inputs = {
                "image": [input_image, perfume_image],
                "prompt": PROMPT,
                "generator": torch.Generator(device="cuda").manual_seed(seed),
                "height": h,
                "width": w,
                "guidance_scale": cfg,
                # "negative_prompt": "",
                "num_inference_steps": 50,
                "num_images_per_prompt": 1,
            }

            with torch.inference_mode():
                result = pipeline(**inputs)
                img = result.images[0]
            
            # Save individual image
            img.save(save_path)
            cfg_images.append(img)

        # Create and save the grid for this CFG (2 rows of 4 images)
        grid_img = create_grid(cfg_images, rows=2, cols=4)
        grid_path = output_dir / f"GRID_cfg{cfg}_all_seeds.jpg"
        grid_img.save(grid_path, quality=95)
        print(f"Grid saved for CFG {cfg} at: {grid_path}")

    print(f"\nAll tasks complete. Check the '{output_dir}' folder.")

if __name__ == "__main__":
    main()

inference_klein_9b_turbo.py

import os
import torch
from pathlib import Path
from PIL import Image
from diffusers import Flux2KleinPipeline

def load_pipeline(model_path: str) -> Flux2KleinPipeline:
    """Load FireRed image edit pipeline."""
    pipe = Flux2KleinPipeline.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        cache_dir="./models",
    )
    pipe.to("cuda")
    return pipe

def create_grid(images, rows, cols):
    """Helper to create a single large image grid."""
    w, h = images[0].size
    grid = Image.new('RGB', size=(cols * w, rows * h))
    for i, img in enumerate(images):
        grid.paste(img, box=(i % cols * w, i // cols * h))
    return grid

def main():
    # --- Configuration ---
    MODEL_ID = "black-forest-labs/FLUX.2-klein-9B"
    INPUT_FILE_ORIGINAL_IMAGE = "417_radiant_see_combined.png"
    INPUT_FILE_PERFUME_BOTTLE = "417_radiant_see.png"
    # PROMPT = "Dress the model in the one-piece dress from Image 1. Match the same style and texture as in Image 1. The woman is in a minimalist studio. Her hair is fastened up. Background is seamless off-white #F5F5F5. The sharp focus is specifically on the fabric texture, neckline, and details of the reference dress. Soft studio light. Kodak film 400 grain."
    PROMPT = "Add proper bokeh for the woman's face and make the perfume bottle clearer. The image features a luxury skincare serum in a sleek, tapered glass bottle with a sophisticated dark amber gradient that fades toward a thick, clear base. The bottle is topped with a tiered, metallic gold dropper cap that adds an elegant, high-end feel. All the text is printed in a matching gold serif font, beginning with the brand logo \"-417\" at the top, followed by the underlined collection name \"RADIANT SEE\" Near the bottom, the product is identified as the \"EXTREME GLOW NOURISHING FACE SERUM\" in clear, all-caps lettering."

    # RESOLUTION = 1536
    CFG_VALUES = [1.0]
    SEEDS = [10, 20, 30, 40, 50, 60, 70, 80] # 8 distinct seeds
    
    # Setup Output
    output_dir = Path("klein_perfume")
    output_dir.mkdir(exist_ok=True)

    # Load Pipeline
    pipeline = load_pipeline(MODEL_ID)

    # Load Inputs
    input_image = Image.open(INPUT_FILE_ORIGINAL_IMAGE).convert("RGB")
    perfume_image = Image.open(INPUT_FILE_PERFUME_BOTTLE).convert("RGB")

    for cfg in CFG_VALUES:
        cfg_images = []
        print(f"\n--- Starting Sweep for CFG: {cfg} ---")
        
        for seed in SEEDS:
            file_name = f"cfg{cfg}_seed{seed}.png"
            save_path = output_dir / file_name
            
            print(f"Generating: {file_name}")
            w, h = input_image.size
            
            inputs = {
                "image": [input_image, perfume_image],
                "prompt": PROMPT,
                "generator": torch.Generator(device="cuda").manual_seed(seed),
                "height": h,
                "width": w,
                "guidance_scale": cfg,
                # "negative_prompt": "",
                "num_inference_steps": 4,
                "num_images_per_prompt": 1,
            }

            with torch.inference_mode():
                result = pipeline(**inputs)
                img = result.images[0]
            
            # Save individual image
            img.save(save_path)
            cfg_images.append(img)

        # Create and save the grid for this CFG (2 rows of 4 images)
        grid_img = create_grid(cfg_images, rows=2, cols=4)
        grid_path = output_dir / f"GRID_cfg{cfg}_all_seeds.jpg"
        grid_img.save(grid_path, quality=95)
        print(f"Grid saved for CFG {cfg} at: {grid_path}")

    print(f"\nAll tasks complete. Check the '{output_dir}' folder.")

if __name__ == "__main__":
    main()

inference_klein_9b_vton.py

import os
import torch
from pathlib import Path
from PIL import Image
from diffusers import Flux2KleinPipeline

def load_pipeline(model_path: str) -> Flux2KleinPipeline:
    """Load FireRed image edit pipeline."""
    pipe = Flux2KleinPipeline.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        cache_dir="./models",
    )
    pipe.to("cuda")
    return pipe

def create_grid(images, rows, cols):
    """Helper to create a single large image grid."""
    w, h = images[0].size
    grid = Image.new('RGB', size=(cols * w, rows * h))
    for i, img in enumerate(images):
        grid.paste(img, box=(i % cols * w, i // cols * h))
    return grid

def main():
    # --- Configuration ---
    MODEL_ID = "black-forest-labs/FLUX.2-klein-base-9B"
    INPUT_FILE_FACE = "vton2/vton_face.png"
    INPUT_FILE_PANTSUIT = "vton2/vton_pants.jpeg"
    INPUT_FILE_BG = "vton2/vton_bg.jpeg"
    # PROMPT = "high definition product photography of this woman wearing this pantsuit on this background."
    # PROMPT = "176cm height  woman\nwearing  outfit\nfront full body\nbackground\na bit from the knee low shot"
    PROMPT = "176cm height  woman\nwearing a two-piece loungewear set in a monochromatic black color, consisting of a short-sleeved t-shirt and matching wide-leg pants. Both garments are crafted from a textured fabric featuring an all-over, tonal embossed pattern of the \"DKNY\" logo repeated in horizontal rows. The top has a classic crew neckline and a standard fit, while the pants feature an elasticized waistband and a relaxed, straight-leg cut that drapes down to the ankles.\nfront full body\nbackground\na bit from the knee low shot"

    # RESOLUTION = 1536
    CFG_VALUES = [2.5, 3.5]
    SEEDS = [10, 20, 30, 40] # 8 distinct seeds
    
    # Setup Output
    output_dir = Path("klein_perfume_base_vton")
    output_dir.mkdir(exist_ok=True)

    # Load Pipeline
    pipeline = load_pipeline(MODEL_ID)

    # Load Inputs
    face_image = Image.open(INPUT_FILE_FACE).convert("RGB")
    suit_image = Image.open(INPUT_FILE_PANTSUIT).convert("RGB")
    bg_image = Image.open(INPUT_FILE_BG).convert("RGB")

    for cfg in CFG_VALUES:
        cfg_images = []
        print(f"\n--- Starting Sweep for CFG: {cfg} ---")
        
        for seed in SEEDS:
            file_name = f"cfg{cfg}_seed{seed}.png"
            save_path = output_dir / file_name
            
            print(f"Generating: {file_name}")
            
            inputs = {
                "image": [face_image, suit_image, bg_image],
                "prompt": PROMPT,
                "generator": torch.Generator(device="cuda").manual_seed(seed),
                "height": 2048,
                "width": 1536,
                "guidance_scale": cfg,
                # "negative_prompt": "",
                "num_inference_steps": 50,
                "num_images_per_prompt": 1,
            }

            with torch.inference_mode():
                result = pipeline(**inputs)
                img = result.images[0]
            
            # Save individual image
            img.save(save_path)
            cfg_images.append(img)

        # Create and save the grid for this CFG (2 rows of 4 images)
        grid_img = create_grid(cfg_images, rows=2, cols=2)
        grid_path = output_dir / f"GRID_cfg{cfg}_all_seeds.jpg"
        grid_img.save(grid_path, quality=95)
        print(f"Grid saved for CFG {cfg} at: {grid_path}")

    print(f"\nAll tasks complete. Check the '{output_dir}' folder.")

if __name__ == "__main__":
    main()

qwen_upscaler.py

from diffusers import QwenImageEditPlusPipeline
import torch
from PIL import Image

# Load the pipeline
pipeline = QwenImageEditPlusPipeline.from_pretrained("Qwen/Qwen-Image-Edit-2511", cache_dir="./models")
pipeline.to(torch.bfloat16)
pipeline.to("cuda")

# Load trained LoRA weights for in-scene editing
pipeline.load_lora_weights("valiantcat/Qwen-Image-Edit-2511-Upscale2K", weight_name="qwen_image_edit_2511_upscale.safetensors", cache_dir="./models")

# Load input image
image = Image.open("vton2/vton_pants_512.png").convert("RGB")

# Define in-scene editing prompt
prompt = "Upscale this picture to 4K resolution."

# Generate edited image with enhanced scene understanding
inputs = {
    "image": image,
    "prompt": prompt,
    "generator": torch.manual_seed(12345),
    "true_cfg_scale": 2.5,
    "negative_prompt": " ",
    "num_inference_steps": 50,
    "height": 2048,
    "width": 1024,
}

with torch.inference_mode():
    output = pipeline(**inputs)
    output_image = output.images[0]
    output_image.save("edited_image5.png")