linoyts's picture
linoyts HF Staff
change #edit default to 3
73ff513 verified
raw
history blame
29.1 kB
import os
import gc
from typing import List, Tuple, Dict
import json
import spaces
import traceback
import torch
import gradio as gr
from PIL import Image
from model.transformer_flux import FluxTransformer2DModelwithSliderConditioning
# from diffusers import FluxTransformer2DModel
from model.sliders_model import SliderProjector, SliderProjector_wo_clip
from model.sliders_pipeline import FluxKontextSliderPipeline
from huggingface_hub import login, snapshot_download
HF_TOKEN = os.getenv("HF_TOKEN")
if HF_TOKEN:
# Auth for this process (does not print or persist the token in your logs)
login(token=HF_TOKEN)
# -----------------------------
# Environment & device
# -----------------------------
# Avoid meta-tensor init from environment leftovers
os.environ.pop("ACCELERATE_INIT_EMPTY_WEIGHTS", None)
# -----------------------------
# Model / pipeline loading
# -----------------------------
def _log(msg): print(msg, flush=True)
# def load_pipeline_single_gpu():
# global PIPELINE
# if PIPELINE is not None:
# _log("[worker] PIPELINE already initialized; skipping.")
# return "warm"
# try:
# os.environ.pop("ACCELERATE_INIT_EMPTY_WEIGHTS", None)
# token = os.environ.get("HF_TOKEN")
# cuda_ok = torch.cuda.is_available()
# _log(f"[worker] cuda available: {cuda_ok}")
# if cuda_ok:
# torch.backends.cudnn.benchmark = True
# # ---------- config ----------
# pretrained = "black-forest-labs/FLUX.1-Kontext-dev"
# trained_models_path = "./model_weights/"
# projector_path = os.path.join(trained_models_path, "slider_projector.pth")
# offload_dir = "/tmp/offload"; os.makedirs(offload_dir, exist_ok=True)
# if not os.path.isdir(trained_models_path):
# return f"error: missing dir {trained_models_path}"
# if not os.path.isfile(projector_path):
# return f"error: missing projector weights at {projector_path}"
# # dtype selection to cut memory
# if cuda_ok and torch.cuda.get_device_capability(0)[0] >= 8:
# dtype = torch.bfloat16
# elif cuda_ok:
# dtype = torch.float16
# else:
# dtype = torch.float32
# max_memory = {"cuda": "80GiB", "cpu": "60GiB"} # tune if needed
# _log("[worker] loading transformer (sharded/offloaded)…")
# transformer = FluxTransformer2DModelwithSliderConditioning.from_pretrained(
# pretrained,
# subfolder="transformer",
# token=token,
# trust_remote_code=True,
# torch_dtype=dtype,
# low_cpu_mem_usage=True,
# # device_map="balanced_low_0",
# offload_folder=offload_dir,
# offload_state_dict=True,
# # max_memory=max_memory,
# )
# weight_dtype = transformer.dtype
# _log(f"[worker] transformer loaded, dtype={weight_dtype}")
# _log("[worker] building slider projector…")
# slider_projector = SliderProjector(out_dim=6144, pe_dim=2, n_layers=4, is_clip_input=True)
# slider_projector.eval()
# _log("[worker] loading projector weights…")
# state_dict = torch.load(projector_path, map_location="cpu", weights_only=True)
# slider_projector.load_state_dict(state_dict, strict=True)
# _log("[worker] assembling pipeline (sharded/offloaded)…")
# pipe = FluxKontextSliderPipeline.from_pretrained(
# pretrained,
# token=token,
# trust_remote_code=True,
# transformer=transformer,
# slider_projector=slider_projector,
# torch_dtype=weight_dtype,
# low_cpu_mem_usage=True,
# # device_map="balanced_low_0",
# offload_folder=offload_dir,
# offload_state_dict=True,
# # max_memory=max_memory,
# )
# _log("[worker] pipeline assembled.")
# _log(f"[worker] loading LoRA from: {trained_models_path}")
# pipe.load_lora_weights(trained_models_path)
# _log("[worker] LoRA loaded.")
# # DO NOT pipe.to("cuda") here; keep auto device_map to avoid OOM
# PIPELINE = pipe
# if cuda_ok:
# free, total = torch.cuda.mem_get_info()
# _log(f"[worker] VRAM free/total: {free/1e9:.2f}/{total/1e9:.2f} GB")
# _log("[worker] PIPELINE ready.")
# return "ok"
# except Exception:
# _log("[worker] init exception:\n" + traceback.format_exc())
# return "error"
# -----------------------------
# Loading the pipeline without any function so that it will be called directly in the inference
# -----------------------------
os.environ.pop("ACCELERATE_INIT_EMPTY_WEIGHTS", None)
token = os.environ.get("HF_TOKEN")
cuda_ok = torch.cuda.is_available()
_log(f"[worker] cuda available: {cuda_ok}")
if cuda_ok:
torch.backends.cudnn.benchmark = True
# ---------- config ----------
pretrained = "black-forest-labs/FLUX.1-Kontext-dev"
trained_models_path = "./model_weights/"
projector_path = os.path.join(trained_models_path, "slider_projector.pth")
offload_dir = "/tmp/offload"; os.makedirs(offload_dir, exist_ok=True)
# dtype selection to cut memory
if cuda_ok and torch.cuda.get_device_capability(0)[0] >= 8:
dtype = torch.bfloat16
elif cuda_ok:
dtype = torch.float16
else:
dtype = torch.float32
max_memory = {"cuda": "80GiB", "cpu": "60GiB"} # tune if needed
_log("[worker] loading transformer (sharded/offloaded)…")
transformer = FluxTransformer2DModelwithSliderConditioning.from_pretrained(
pretrained,
subfolder="transformer",
token=token,
trust_remote_code=True,
# torch_dtype=dtype,
# low_cpu_mem_usage=True,
# device_map="balanced_low_0",
# offload_folder=offload_dir,
# offload_state_dict=True,
# max_memory=max_memory,
)
weight_dtype = transformer.dtype
_log(f"[worker] transformer loaded, dtype={weight_dtype}")
_log("[worker] building slider projector…")
slider_projector = SliderProjector(out_dim=6144, pe_dim=2, n_layers=4, is_clip_input=True)
slider_projector.eval()
_log("[worker] loading projector weights…")
state_dict = torch.load(projector_path, map_location="cpu", weights_only=True)
slider_projector.load_state_dict(state_dict, strict=True)
_log("[worker] assembling pipeline (sharded/offloaded)…")
PIPELINE = FluxKontextSliderPipeline.from_pretrained(
pretrained,
token=token,
trust_remote_code=True,
transformer=transformer,
slider_projector=slider_projector,
torch_dtype=weight_dtype,
# low_cpu_mem_usage=True,
# device_map="balanced_low_0",
# offload_folder=offload_dir,
# offload_state_dict=True,
# max_memory=max_memory,
)
_log("[worker] pipeline assembled.")
_log(f"[worker] loading LoRA from: {trained_models_path}")
PIPELINE.load_lora_weights(trained_models_path)
_log("[worker] LoRA loaded.")
# moving the pipeline to GPU
PIPELINE.to('cuda')
# -----------------------------
# Sample Images & Precomputed Results
# -----------------------------
def create_sample_entry(name, image_filename, prompt, result_folder, num_results=5, result_pattern="image_{i}.png", precomputed_base="./sample_images/precomputed"):
"""
Helper function to create a sample entry with subfolder organization.
Args:
name: Display name in dropdown
image_filename: Filename in ./sample_images/
prompt: Editing instruction
result_folder: Subfolder name in precomputed directory
num_results: Number of precomputed results (default 5)
result_pattern: Filename pattern, {i} will be replaced with 0,1,2,3,4 (default "image_{i}.png")
precomputed_base: Base path for precomputed results (default "./sample_images/precomputed")
"""
return {
"name": name,
"image_path": f"./sample_images/{image_filename}",
"prompt": prompt,
"precomputed_results": [f"{precomputed_base}/{result_folder}/{result_pattern.format(i=i)}" for i in range(num_results)]
}
def load_samples_from_config(config_file="sample_config.json"):
"""Load sample data from a JSON configuration file."""
if os.path.exists(config_file):
try:
with open(config_file, 'r') as f:
return json.load(f)
except Exception as e:
print(f"Error loading sample config: {e}")
return []
def discover_samples_automatically(sample_dir="./sample_images", precomputed_dir="./sample_images/precomputed"):
"""Automatically discover samples based on directory structure with subfolders."""
discovered_samples = []
if not os.path.exists(sample_dir) or not os.path.exists(precomputed_dir):
return discovered_samples
# Look for subfolders in precomputed directory
for subfolder in os.listdir(precomputed_dir):
subfolder_path = os.path.join(precomputed_dir, subfolder)
if os.path.isdir(subfolder_path):
# Look for sequential result files in subfolder
precomputed_files = []
for i in range(0, 15): # Check for up to 15 results starting from 0
# Try different patterns
for pattern in [f"image_{i}.png", f"image_{i}.jpg", f"{i}.jpg", f"{i}.png", f"result_{i}.jpg", f"output_{i}.png"]:
result_path = os.path.join(subfolder_path, pattern)
if os.path.exists(result_path):
precomputed_files.append(result_path)
break
else:
# If no file with this index found, stop looking (but continue if we found at least one)
if i == 0 and not precomputed_files:
continue # Keep trying from index 0
elif not precomputed_files:
break # No files found at all
else:
break # Found some files but this index is missing, stop here
if precomputed_files:
# Try to find corresponding source image
img_path = None
# Common naming patterns for source images
base_name = subfolder.split('_')[0] # e.g., "portrait" from "portrait_smile"
for ext in ['.jpg', '.jpeg', '.png']:
candidate = os.path.join(sample_dir, f"{base_name}{ext}")
if os.path.exists(candidate):
img_path = candidate
break
if img_path:
sample = {
"name": f"{subfolder.replace('_', ' ').title()} - Auto-discovered",
"image_path": img_path,
"prompt": f"Edit: {subfolder.replace('_', ' ')}", # Default prompt
"precomputed_results": precomputed_files
}
discovered_samples.append(sample)
return discovered_samples
# Main sample data - using your actual folder structure
SAMPLE_DATA = [
create_sample_entry("Stylization", "aesthetic_model2_vangogh.png", "Transform the image into a Van Gogh Style painting", "aesthetic_model2_vangogh", 11),
create_sample_entry("Weather Change", "enfield3_winter_snow.png", "Transform the scene into winter season with heavy snowfall", "enfield3_winter_snow", 11),
create_sample_entry("Illumination Change", "light_lamp_blue_side.png", "Turn on the lamp with blue lighting", "light_lamp_blue_side", 11),
create_sample_entry("Appearance Change", "jackson_fluffy.png", "Transform his jacket into a blue fluffy fur jacket", "jackson_fluffy", 11),
create_sample_entry("Scene Edit", "venice1_grow_ivy.png", "Grow ivy on the walls of the buildings on the side", "venice1_grow_ivy", 11)
]
# Add more samples using the helper function
# Modify these examples or add your own:
ADDITIONAL_SAMPLES = [
# Add your own samples here following your folder structure:
#
# For your structure (./sample_images/precomputed/folder_name/image_0.png, image_1.png, etc.):
# create_sample_entry("Display Name", "your_image.png", "editing prompt", "folder_name", 12),
#
# Examples based on your pattern:
# create_sample_entry("New Sample", "new_image.png", "apply some effect", "new_folder", 12),
# create_sample_entry("Another Edit", "source.png", "different editing instruction", "another_folder", 10),
# Note:
# - Images should be in ./sample_images/
# - Precomputed results should be in ./sample_images/precomputed/folder_name/
# - Default pattern is image_0.png, image_1.png, etc.
# - Adjust the number (12) to match how many results you have
]
# Extend the main sample data with additional samples
SAMPLE_DATA.extend(ADDITIONAL_SAMPLES)
# Optional: Auto-discover additional samples from directories
# Uncomment to automatically find additional samples beyond the manual ones above:
# AUTO_DISCOVERED = discover_samples_automatically()
# if AUTO_DISCOVERED:
# print(f"Auto-discovered {len(AUTO_DISCOVERED)} additional samples:")
# for sample in AUTO_DISCOVERED:
# print(f" - {sample['name']}")
# SAMPLE_DATA.extend(AUTO_DISCOVERED)
# Optional: Load samples from external JSON config
# CONFIG_SAMPLES = load_samples_from_config("sample_config.json")
# SAMPLE_DATA.extend(CONFIG_SAMPLES)
def load_sample_image(image_path: str) -> Image.Image:
"""Load a sample image, with fallback to a placeholder if file doesn't exist."""
try:
if os.path.exists(image_path):
return Image.open(image_path)
else:
# Create a placeholder image if sample doesn't exist
placeholder = Image.new('RGB', (512, 512), color=(200, 200, 200))
return placeholder
except Exception as e:
print(f"Error loading sample image {image_path}: {e}")
# Return a placeholder image
placeholder = Image.new('RGB', (512, 512), color=(200, 200, 200))
return placeholder
def load_precomputed_results(result_paths: List[str]) -> List[Image.Image]:
"""Load precomputed result images, with fallbacks for missing files."""
results = []
for path in result_paths:
try:
if os.path.exists(path):
results.append(Image.open(path))
else:
# Create placeholder result
placeholder = Image.new('RGB', (512, 512), color=(150, 150, 150))
results.append(placeholder)
except Exception as e:
print(f"Error loading precomputed result {path}: {e}")
placeholder = Image.new('RGB', (512, 512), color=(150, 150, 150))
results.append(placeholder)
return results
# -----------------------------
# Helpers
# -----------------------------
def resize_image(img: Image.Image, target: int = 512) -> Image.Image:
"""Resize shortest side to target, then center-crop to target x target."""
w, h = img.size
try:
resample = Image.Resampling.BICUBIC # PIL >= 10
except Exception:
resample = Image.BICUBIC
if h > w:
new_w, new_h = target, int(target * h / w)
elif h < w:
new_w, new_h = int(target * w / h), target
else:
new_w, new_h = target, target
# resizing the image to a fixed lower dimension size of 512
img = img.resize((new_w, new_h), resample)
return img
# -----------------------------
# Inference functions
# -----------------------------
@spaces.GPU(duration=220)
@torch.no_grad()
def generate_image_stack_edits(text_prompt, n_edits, input_image):
"""
Compute n_edits images on a single GPU for slider values in (0,1],
return (list_of_images, first_image) so the UI shows immediately.
"""
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
# pipelien will be loaded already in the global context and will be called here
if not input_image or not text_prompt or text_prompt.startswith("Please select"):
return [], None
n = int(n_edits) if n_edits is not None else 1
n = max(1, n)
slider_values = [(i + 1) / float(n) for i in range(n)] # (0,1] inclusive
img = resize_image(input_image, 512)
pe, ppe, _ = PIPELINE.encode_prompt(prompt=text_prompt, prompt_2=text_prompt)
results: List[Image.Image] = []
gen_base = 64 # deterministic seed base
# not using batching for now just a simple forward loop
# batch_size = 2
# n_batches = n // batch_size
# batched_slider_values = [[slider_values[i*batch_size: (i+1)*batch_size]] for i in range(n_batches)]
# print(f"batched_slider_values: {batched_slider_values}")
for i, sv in enumerate(slider_values):
gen = torch.Generator(device=DEVICE if DEVICE != "cpu" else "cpu").manual_seed(gen_base + i)
with torch.no_grad():
# replicating based on the number of examples in the batch size
out = PIPELINE(
image=img,
height=img.height,
width=img.width,
num_inference_steps=28,
prompt_embeds=pe,
pooled_prompt_embeds=ppe,
generator=gen,
text_condn=False,
modulation_condn=True,
slider_value=torch.tensor(sv, device=DEVICE if DEVICE != "cpu" else "cpu").reshape(1, 1),
is_clip_input=True,
)
results.append(out.images[0])
if DEVICE.startswith("cuda"):
torch.cuda.empty_cache()
gc.collect()
first = results[0] if results else None
return results, first
@spaces.GPU(duration=80)
def generate_single_image(text_prompt, slider_value, input_image):
if not input_image or not text_prompt or text_prompt.startswith("Please select"):
return None
img = resize_image(input_image, 512)
sv = float(slider_value)
pe, ppe = _encode_prompt(text_prompt)
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
gen = torch.Generator(device=DEVICE if DEVICE != "cpu" else "cpu").manual_seed(64)
with torch.no_grad():
out = PIPELINE(
image=img,
height=img.height,
width=img.width,
num_inference_steps=28,
prompt_embeds=pe,
pooled_prompt_embeds=ppe,
generator=gen,
text_condn=False,
modulation_condn=True,
slider_value=torch.tensor(sv, device=DEVICE if DEVICE != "cpu" else "cpu").reshape(1, 1),
is_clip_input=True,
)
result = out.images[0]
if DEVICE.startswith("cuda"):
torch.cuda.empty_cache()
gc.collect()
return result
# -----------------------------
# Sample Loading Functions
# -----------------------------
def get_sample_by_name(sample_name: str):
"""Get sample data by name."""
for sample in SAMPLE_DATA:
if sample["name"] == sample_name:
return sample
return None
def load_sample_to_main_interface(sample_name: str):
"""Load selected sample to main interface with precomputed results."""
if not sample_name:
return (
None,
"Please select a sample above to see the editing instruction",
[],
None,
gr.update(minimum=0, maximum=0, step=1, value=0, label="Edit Strength Level")
)
sample = get_sample_by_name(sample_name)
if not sample:
return (
None,
"Sample not found",
[],
None,
gr.update(minimum=0, maximum=0, step=1, value=0, label="Edit Strength Level")
)
# Load sample image
sample_image = load_sample_image(sample["image_path"])
prompt = sample["prompt"]
# Load precomputed results
precomputed_images = load_precomputed_results(sample["precomputed_results"])
first_result = precomputed_images[0] if precomputed_images else None
# Update slider range for precomputed results
n_results = len(precomputed_images)
slider_update = gr.update(
minimum=0,
maximum=max(0, n_results-1),
step=1,
value=0,
label=f"Edit Strength Level (0-{n_results-1}) - Precomputed"
)
return sample_image, prompt, precomputed_images, first_result, slider_update
# -----------------------------
# Helpers
# -----------------------------
def update_slider_range(n_edits):
"""Update the slider range based on number of edits."""
return gr.update(
minimum=0,
maximum=max(0, int(n_edits)-1),
step=1,
value=0,
label=f"Edit Strength Level (0-{int(n_edits)-1})"
)
def display_selected_image(slider_index: int, images_list: List[Image.Image]) -> Image.Image:
"""
Display the image corresponding to the slider index from the generated images list.
Args:
slider_index: Current slider position (0-based index)
images_list: List of generated/precomputed images
Returns:
Selected image or None if invalid index/empty list
"""
if not images_list or len(images_list) == 0:
return None
# Clamp index to valid range
idx = max(0, min(int(slider_index), len(images_list) - 1))
return images_list[idx]
# -----------------------------
# Gradio UI
# -----------------------------
# Add new helper function for user uploads
def process_user_upload(uploaded_image, user_prompt, n_edits_val):
"""Handle user uploaded images and custom prompts."""
if uploaded_image is None:
return None, [], None, gr.update(minimum=0, maximum=0, step=1, value=0, label="Edit Strength Level")
# Resize uploaded image
processed_image = resize_image(uploaded_image, 512)
# Generate edits
generated_list, first_result = generate_image_stack_edits(user_prompt, n_edits_val, processed_image)
# Update slider range
slider_update = gr.update(
minimum=0,
maximum=max(0, len(generated_list)),
step=1,
value=0,
label=f"Edit Strength Level (0-{len(generated_list)-1})"
)
return processed_image, generated_list, first_result, slider_update
with gr.Blocks() as demo:
gr.Markdown("# Kontinuous Kontext - Continuous Strength Control for Instruction-based Image Editing")
# Add description section
gr.Markdown("""
## About
### Kontinuous Kontext allows you to edit a given image with a freeform text instruction and a slider strength value.
### The slider strength enables precise control for the extent of the applied edit and generates smooth transitions between different editing levels.
### You can either:
1. Choose from our sample images with predefined edit instructions
2. Upload your own image and specify custom editing instructions
Checkout the [paper](https://arxiv.org/pdf/2510.08532v1) and the [project page](https://snap-research.github.io/kontinuouskontext) for more details.
""")
# Add custom CSS for tabs
gr.Markdown("""
<style>
.tabs.svelte-710i53 {
margin-top: 2em !important;
margin-bottom: 2em !important;
}
.tabs.svelte-710i53 button {
font-size: 1.2em !important;
padding: 0.5em 2em !important;
min-width: 200px !important;
}
#sample_image, #sample_output, #upload_image, #upload_output {
min-height: 512px !important;
max-height: 512px !important;
}
</style>
""")
with gr.Tabs() as tabs:
# Common style parameters for images
IMAGE_WIDTH = 512
IMAGE_HEIGHT = 512
with gr.TabItem("📸 Examples") as tab1: # Added emoji and changed tab name
with gr.Row(equal_height=True):
with gr.Column(scale=1):
sample_dropdown = gr.Dropdown(
choices=[sample["name"] for sample in SAMPLE_DATA],
label="Select Sample Image & Prompt",
value=None
)
sample_text = gr.Textbox(lines=1, show_label=False, placeholder="Please select a sample above", interactive=False)
sample_image = gr.Image(
type="pil",
label="Source Image",
width=IMAGE_WIDTH,
height=IMAGE_HEIGHT,
interactive=False,
elem_id="sample_image"
)
with gr.Column(scale=1):
with gr.Row():
sample_slider = gr.Slider(
minimum=0,
maximum=1,
step=0.1,
value=0,
label="Edit Strength",
scale=1,
min_width=100
)
sample_output = gr.Image(
type="pil",
label="Edited Output",
width=IMAGE_WIDTH,
height=IMAGE_HEIGHT,
elem_id="sample_output"
)
with gr.TabItem("⬆️ Upload Your Image") as tab2: # Added emoji and changed tab name
with gr.Row(equal_height=True):
with gr.Column(scale=1):
upload_text = gr.Textbox(lines=1, label="Enter Editing Prompt", placeholder="Describe the edit you want...")
upload_n_edits = gr.Number(value=3, minimum=1, maximum=6, step=1, label="Number of Edits", precision=0)
upload_image = gr.Image(
type="pil",
label="Upload Image",
width=IMAGE_WIDTH,
height=IMAGE_HEIGHT,
elem_id="upload_image"
)
upload_button = gr.Button("Generate Edits") # Kept consistent with sample tab
with gr.Column(scale=1):
with gr.Row():
upload_slider = gr.Slider(
minimum=0,
maximum=1,
step=0.1,
value=0,
label="Edit Strength Level",
scale=1,
min_width=100
)
upload_output = gr.Image(
type="pil",
label="Edited Output",
width=IMAGE_WIDTH,
height=IMAGE_HEIGHT,
elem_id="upload_output"
)
# States for both tabs
sample_generated_images = gr.State([])
upload_generated_images = gr.State([])
# Sample tab logic
sample_dropdown.change(
load_sample_to_main_interface,
inputs=[sample_dropdown],
outputs=[sample_image, sample_text, sample_generated_images, sample_output, sample_slider]
)
# sample_button.click(
# generate_image_stack_edits,
# inputs=[sample_text, sample_n_edits, sample_image],
# outputs=[sample_generated_images, sample_output],
# ).then(
# update_slider_range,
# inputs=[sample_n_edits],
# outputs=[sample_slider],
# )
sample_slider.change(
display_selected_image,
inputs=[sample_slider, sample_generated_images],
outputs=[sample_output],
)
# Upload tab logic - Remove duplicate click handler and combine the logic
upload_button.click(
generate_image_stack_edits, # Generate images first
inputs=[upload_text, upload_n_edits, upload_image],
outputs=[upload_generated_images, upload_output],
).then(
update_slider_range, # Then update slider range
inputs=[upload_n_edits],
outputs=[upload_slider],
)
# Update slider when n_edits changes
upload_n_edits.change(
update_slider_range,
inputs=[upload_n_edits],
outputs=[upload_slider],
)
upload_slider.change(
display_selected_image,
inputs=[upload_slider, upload_generated_images],
outputs=[upload_output],
)
# Add citation section at the bottom
gr.Markdown("""
---
### If you find this work useful, please cite:
```bibtex
@article{kontinuous_kontext_2025,
title={Kontinuous Kontext: Continuous Strength Control for Instruction-based Image Editing},
author={R Parihar, O Patashnik, D Ostashev, R Venkatesh Babu, D Cohen-Or, and J Wang},
journal={Arxiv},
year={2025}
}
```
""")
if __name__ == "__main__":
demo.launch(server_name="0.0.0.0", server_port=7860, share=False, debug=True)