import os
import gc
import random
import gradio as gr
import numpy as np
import spaces
import torch
from typing import Iterable
from gradio.themes import Soft
from gradio.themes.utils import colors, fonts, sizes
# ── Theme ──────────────────────────────────────────────────────────────────────
colors.steel_blue = colors.Color(
name="steel_blue",
c50="#EBF3F8", c100="#D3E5F0", c200="#A8CCE1", c300="#7DB3D2",
c400="#529AC3", c500="#4682B4", c600="#3E72A0", c700="#36638C",
c800="#2E5378", c900="#264364", c950="#1E3450",
)
class SteelBlueTheme(Soft):
def __init__(
self,
*,
primary_hue: colors.Color | str = colors.gray,
secondary_hue: colors.Color | str = colors.steel_blue,
neutral_hue: colors.Color | str = colors.slate,
text_size: sizes.Size | str = sizes.text_lg,
font: fonts.Font | str | Iterable[fonts.Font | str] = (
fonts.GoogleFont("Outfit"), "Arial", "sans-serif",
),
font_mono: fonts.Font | str | Iterable[fonts.Font | str] = (
fonts.GoogleFont("IBM Plex Mono"), "ui-monospace", "monospace",
),
):
super().__init__(
primary_hue=primary_hue, secondary_hue=secondary_hue,
neutral_hue=neutral_hue, text_size=text_size, font=font, font_mono=font_mono,
)
super().set(
body_background_fill="linear-gradient(135deg, *primary_200, *primary_100)",
body_background_fill_dark="linear-gradient(135deg, *primary_900, *primary_800)",
button_primary_text_color="white",
button_primary_background_fill="linear-gradient(90deg, *secondary_500, *secondary_600)",
button_primary_background_fill_hover="linear-gradient(90deg, *secondary_600, *secondary_700)",
slider_color="*secondary_500",
block_title_text_weight="600",
block_border_width="3px",
block_shadow="*shadow_drop_lg",
)
steel_blue_theme = SteelBlueTheme()
# ── Device / dtype ─────────────────────────────────────────────────────────────
# ZeroGPU exposes a CUDA emulation device at import time. Models MUST be
# placed on "cuda" at module level (HF ZeroGPU docs). Real GPU (RTX Pro 6000
# Blackwell) is only attached inside @spaces.GPU.
dtype = torch.bfloat16
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("CUDA available:", torch.cuda.is_available())
print("Using device:", device)
# ── Model loading (local qwenimage package) ────────────────────────────────────
from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
from qwenimage.transformer_qwenimage import (
QwenImageTransformer2DModel,
QwenDoubleStreamAttnProcessor2_0,
)
pipe = QwenImageEditPlusPipeline.from_pretrained(
"Qwen/Qwen-Image-Edit-2509",
transformer=QwenImageTransformer2DModel.from_pretrained(
"prithivMLmods/Qwen-Image-Edit-Rapid-AIO-V4",
torch_dtype=dtype,
device_map="cuda",
),
torch_dtype=dtype,
).to(device)
# VAE tiling/slicing — bound peak VRAM. APIs differ across diffusers versions.
try:
pipe.vae.enable_tiling(tile_sample_min_width=256, tile_sample_min_height=256)
except TypeError:
try:
pipe.vae.enable_tiling()
except Exception as e:
print(f"VAE tiling not enabled: {e}")
try:
pipe.vae.enable_slicing()
except Exception as e:
print(f"VAE slicing not enabled: {e}")
# ZeroGPU is NVIDIA RTX Pro 6000 Blackwell (sm_120). The vllm-flash-attn3
# kernel has no image for sm_120, so FA3 at import (under CUDA emulation)
# later crashes inference with: "CUDA error: no kernel image is available".
# Always use PyTorch SDPA on this Space.
try:
pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessor2_0())
print("Using SDPA attention processor (Blackwell-safe).")
except Exception as e:
print(f"Warning: Could not set SDPA processor: {e}")
# ── LoRA catalog ──────────────────────────────────────────────────────────
LORA_REPO = "wiikoo/Qwen-lora-nsfw"
LORA_CONFIGS = {
"CockQwen_v3": "loras/CockQwen-v3.safetensors",
"Eva_Qwen_V3": "loras/Eva_Qwen_V3.safetensors",
"Facial_Cumshots_V1": "loras/Facial_Cumshots_For_Qwen_Image_V1.safetensors",
"HearmemanAI_V3_Breasts": "loras/HearmemanAI_V3_Rank64_BreastsLoRA_Epoch60.safetensors",
"HearmemanAI_V4_Breasts": "loras/HearmemanAI_V4_Rank128_BreastsLoRA_Epoch80.safetensors",
"InniePussy": "loras/InniePussy.safetensors",
"JTT2_5": "loras/[QWEN] JTT2_5.safetensors",
"LumiNude01a": "loras/LumiNude01a_CE_QWEN_AIT3k.safetensors",
"MEXX_QWEN_TG300": "loras/MEXX_QWEN_TG300_23.safetensors",
"Meta4": "loras/Meta4.safetensors",
"MysticXXX": "loras/Qwen-MysticXXX-v1.safetensors",
"Nsfw_Body_V10": "loras/Qwen_Nsfw_Body_V10-4K.safetensors",
"Nsfw_Body_V14": "loras/Qwen_Nsfw_Body_V14-10K.safetensors",
"OilySkin_V2": "loras/Oily Skin QWEN V2-GMR.safetensors",
"PillowHump_2509": "loras/PillowHump_2509.safetensors",
"PutItHere_V2": "loras/Put it here_Qwen edit_V2.0.safetensors",
"PutItHere_V01": "loras/put it here_QwenEdit_V0.1.safetensors",
"Qwen4Play_v2": "loras/Qwen4Play_v2.safetensors",
"QwenHentai_v3": "loras/QwenImageHentaiPIV_v3.1.safetensors",
"Qwen_Helm": "loras/Qwen-Image-Helm_v0.1.safetensors",
"Qwen_NSFW_Beta1": "loras/Qwen-NSFW.safetensors",
"Qwen_NSFW_Beta2": "loras/Qwen-NSFW-Beta2.safetensors",
"Qwen_NSFW_Beta4": "loras/Qwen-NSFW-Beta4.safetensors",
"Qwen_NSFW_Beta5": "loras/Qwen-NSFW-Beta5.safetensors",
"Qwen_Real_Nud3s": "loras/Qwen_Real_Nud3s.safetensors",
"Qwen_Real_PS": "loras/Qwen-Real PS_v1_83K.safetensors",
"QwenSnofs_v1": "loras/qwen_snofs.safetensors",
"QwenSnofs_v1_1": "loras/QwenSnofs1_1.safetensors",
"Real_Breast_Nipples": "loras/Real Breast Nipples-QWEN-[rbn]-GMR.safetensors",
"SendDudes": "loras/[QWEN] SendDudes.safetensors",
"SendNudesLite": "loras/SendNudesLite (Qwen).safetensors",
"SendNudesPro_Beta": "loras/[QWEN] Send Nudes Pro - Beta v1.safetensors",
"Ultimate_Breast_Nipples": "loras/Ultimate Realistic Breast NIPPLES-QWEN-[rab]-GMR.safetensors",
"ass_up_QWEN": "loras/ass_up_QWEN.safetensors",
"barbell_nipples_QWEN": "loras/QWEN_jtn_barbell.safetensors",
"bfs_v2_face": "loras-sfw/face_swap_5500_qwen_image_edit_2509_v1.safetensors",
"bfs_v2_focus_face": "loras-sfw/bfs_v2_000005000.safetensors",
"bfs_v2_head": "loras-sfw/bfs_v2_head_000007000.safetensors",
"big_nipples_QWEN": "loras/big_nipples_QWEN.safetensors",
"bumpynipples": "loras/bumpynipples1.safetensors",
"cmslt_cum_on_her": "loras/cmslt_2509_2.safetensors",
"consistence_edit_v1": "loras-2/consistence_edit_v1.safetensors",
"consistence_edit_v2": "loras2/consistence_edit_v2.safetensors",
"d33p7hroa7": "loras/d33p7hroa7_qwen.safetensors",
"d1ck_p3n1s_V1_1": "loras/qwen-image_d!ck_P3N1S_LoRA_V1.1.safetensors",
"goblin_anal_v1": "loras/goblin_anal_v1_qwen.safetensors",
"horseshoe_nipple_rings": "loras/horseshoe_nipple_rings_QWEN.safetensors",
"jib_nudity_fixer": "loras/jib_qwen_fix_000002750.safetensors",
"jillin": "loras/jillin1.safetensors",
"male_nude": "loras/lora_nudenan_v1.safetensors",
"milk_juggs": "loras/milk_juggs_QWEN.safetensors",
"n00d_b": "loras/n00d-b-qwen.safetensors",
"nsfw_adv_v1": "loras/qwen-image_nsfw_adv_v1.0.safetensors",
"p0ssy_lora_v1": "loras/p0ssy_lora_v1.safetensors",
"p3nis": "loras/p3nis.safetensors",
"qwen_MCNL": "loras/qwen_MCNL_v1.0.safetensors",
"qwen_PENISLORA": "loras/qwen-PENISLORA.safetensors",
"qwen_hand_grab": "loras/qwen_hand_grab_6000s.safetensors",
"qwen_uncensor": "loras/qwen_uncensor_000014928.safetensors",
"reclining_nude": "loras/reclining_nude_v1_000003500.safetensors",
"remove_clothing": "loras/qwen_image_edit_remove-clothing_v1.0.safetensors",
"royal_treatment_V3": "loras/royal+treatment+V3.safetensors",
"sabi_character": "loras-2/sabi_character_v1.safetensors",
"snapchat_selfie": "loras/qwen_image_snapchat.safetensors",
"uka_qwen": "loras/uka_1_qwen.safetensors",
"ultimate_realistic_breast":"loras/ultimate realistic breast.safetensors",
}
LORA_TRIGGER_WORDS = {
"Qwen4Play_v2": "d0gg13, c0wg1rl, r3v3rs3_c0wg1rl, m15510n4ry, bl0wj0b, penis",
"qwen_MCNL": "nsfw, cum_on_face, blowjob, cowgirlout, creamp1e, penis, l1ck, missionary, nipples, reversecowgirlpov, vagina",
"remove_clothing": "remove her clothing",
"Qwen_Real_Nud3s": "nud3",
"HearmemanAI_V4_Breasts": "large breasts, hard nipples, erect nipples",
"HearmemanAI_V3_Breasts": "large breasts, hard nipples, erect nipples",
"Ultimate_Breast_Nipples": "rab",
"ass_up_QWEN": "ass up showing pussy and anus",
"PillowHump_2509": "Pillow, Humping",
"InniePussy": "Innie pussy, Clean shaven, Vertical slit",
"p0ssy_lora_v1": "Nude",
"CockQwen_v3": "Erect Penis",
"p3nis": "holding a p3nis",
"qwen_PENISLORA": "PENISLORA",
"Facial_Cumshots_V1": "cum",
"bfs_v2_head": "head swap, transfer head from image 1 to image 2",
"bfs_v2_face": "keep the face consistent, preserve facial identity",
"bfs_v2_focus_face": "head swap from Image 1 to Image 2",
"goblin_anal_v1": "anal penetration, spread ass",
"d33p7hroa7": "deepthroat, penis deep in mouth",
"QwenHentai_v3": "nsfw, anime style, explicit",
"Eva_Qwen_V3": "Eva_gothic, in a kneeling position",
"JTT2_5": "massive breasts, large breasts, medium breasts, small breasts",
"MEXX_QWEN_TG300": "nsfw, female body",
"OilySkin_V2": "oilski",
"barbell_nipples_QWEN": "barbell nipple piercings",
"Qwen_Helm": "nsfw, anime style",
"MysticXXX": "nsfw",
"Qwen_NSFW_Beta1": "nsfw",
"Qwen_NSFW_Beta2": "nsfw",
"Qwen_NSFW_Beta4": "nsfw",
"Qwen_NSFW_Beta5": "nsfw",
"QwenSnofs_v1": "sex, missionary, cum, cowgirl, reverse cowgirl, selfie, snapchat selfie, prone position, spooning position, undressing",
"QwenSnofs_v1_1": "nsfw, nude, sex, blowjob, cum, selfie",
"Nsfw_Body_V10": "Hourglass figure, Hairless pussy, Hairly pussy",
"Nsfw_Body_V14": "SSS Waistline, Hairless pussy, Hairly pussy",
"SendNudesLite": "nude",
"SendNudesPro_Beta": "flat chest, small breasts, medium breasts, large breasts, massive breasts, big nipples",
"SendDudes": "Penis",
"cmslt_cum_on_her": "Put cum on her",
"horseshoe_nipple_rings": "horseshoe-ring nipple piercings, circular-barbell nipple piercings",
"jib_nudity_fixer": "nude, nipples, vagina",
"jillin": "masturbating",
"male_nude": "nudeman",
"n00d_b": "nude, art photography",
"nsfw_adv_v1": "nsfw",
"d1ck_p3n1s_V1_1": "P3N1S, penis",
"qwen_uncensor": "nsfw, cum_on_face, blowjob, cowgirlout, creamp1e, penis, l1ck, missionary, nipples, reversecowgirlpov, vagina",
"royal_treatment_V3": "lick ass, blowjob",
"snapchat_selfie": "selfie, snapchat",
"ultimate_realistic_breast":"urb, realistic breast",
}
# Tracks which adapter names have been successfully loaded into the pipeline.
# We keep this across @spaces.GPU calls because the LoRA weights stay in RAM
# (only active GPU state is reset by ZeroGPU). This avoids re-downloading
# and re-registering the same adapters on every generation.
LOADED_ADAPTERS: set[str] = set()
# ── Helpers ────────────────────────────────────────────────────────────────────
def append_triggers(current_prompt: str, lora_name: str) -> str:
"""Append a LoRA's trigger words to the prompt (no duplicates)."""
if lora_name == "None":
return current_prompt
triggers = LORA_TRIGGER_WORDS.get(lora_name, "")
if not triggers:
return current_prompt
existing = {w.strip().lower() for w in current_prompt.replace(",", " ").split()}
new_words = [w.strip() for w in triggers.split(",")
if w.strip().lower() not in existing and w.strip()]
if not new_words:
return current_prompt
sep = ", " if current_prompt.strip() else ""
return current_prompt.rstrip(", ") + sep + ", ".join(new_words)
def _resolve_lora(name: str):
"""Return (repo_id, weight_name) for a catalog entry."""
cfg = LORA_CONFIGS.get(name)
if cfg is None:
return None
if isinstance(cfg, (tuple, list)) and len(cfg) == 2:
return cfg[0], cfg[1]
return LORA_REPO, cfg
def load_and_apply_stack(extra_adapters: list[str], extra_weights: list[float]):
"""Lazy-load any unseen adapters (only once per app lifetime), then
activate exactly the requested stack for this inference.
"""
if not extra_adapters:
pipe.disable_lora()
return [], []
loaded, weights_out = [], []
for name, weight in zip(extra_adapters, extra_weights):
resolved = _resolve_lora(name)
if resolved is None:
continue
repo_id, weight_name = resolved
if name not in LOADED_ADAPTERS:
try:
print(f"--- Loading adapter: {name} ({repo_id}/{weight_name}) ---")
pipe.load_lora_weights(
repo_id,
weight_name=weight_name,
adapter_name=name,
)
LOADED_ADAPTERS.add(name)
except Exception as e:
msg = str(e).lower()
if "already exists" in msg or "already loaded" in msg:
print(f"Adapter '{name}' already registered — reusing.")
LOADED_ADAPTERS.add(name)
else:
print(f"WARNING: Failed to load LoRA '{name}': {e}")
continue
loaded.append(name)
weights_out.append(weight)
if loaded:
pipe.enable_lora()
pipe.set_adapters(loaded, adapter_weights=weights_out)
else:
pipe.disable_lora()
return loaded, weights_out
def clear_lora_stack():
"""Reset all 6 LoRA dropdowns to 'None' and sliders to 0.75."""
updates = []
for _ in range(6):
updates.append(gr.update(value="None"))
updates.append(gr.update(value=0.75))
return updates
def _empty_cuda():
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
# ── Inference ──────────────────────────────────────────────────────────────────
MAX_SEED = np.iinfo(np.int32).max
DEFAULT_NEGATIVE_PROMPT = (
"worst quality, low quality, bad anatomy, bad hands, text, error, "
"missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, "
"signature, watermark, username, blurry"
)
def gpu_duration(*_args, **_kwargs):
# Rapid-AIO is 4-step; budget extra time for first LoRA download.
return 120
@spaces.GPU(duration=gpu_duration, size="xlarge")
def infer(
input_image,
prompt,
seed,
randomize_seed,
guidance_scale,
steps,
negative_prompt,
*lora_params,
progress=gr.Progress(track_tqdm=True),
):
# Qwen-Image-Edit-2509 (~58GB bf16) does not fit the default ZeroGPU
# `large` slice (48GB). `xlarge` is a full RTX Pro 6000 (96GB).
_empty_cuda()
try:
print("GPU:", torch.cuda.get_device_name(0), "cap:", torch.cuda.get_device_capability(0))
except Exception:
pass
pipe.disable_lora()
if input_image is None:
raise gr.Error("Please upload an image.")
image = input_image.convert("RGB")
w, h = image.size
ratio = max(w, h) / max(min(w, h), 1)
if ratio > 4.0:
raise gr.Error(
f"Image aspect ratio too extreme ({w}x{h}, ratio {ratio:.1f}:1). "
"Please use an image with aspect ratio ≤ 4:1."
)
extra_adapters, extra_weights = [], []
for i in range(0, len(lora_params), 2):
name, strength = lora_params[i], lora_params[i + 1]
if name != "None" and float(strength) > 0.05:
extra_adapters.append(name)
extra_weights.append(float(strength))
load_and_apply_stack(extra_adapters, extra_weights)
if randomize_seed:
seed = random.randint(0, MAX_SEED)
# Always seed on the real CUDA device attached inside @spaces.GPU —
# do not reuse the import-time `device` object.
cuda_device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
generator = torch.Generator(device=cuda_device).manual_seed(int(seed))
try:
# Pipeline's __call__ is already decorated with @torch.no_grad().
# Do NOT use torch.inference_mode() here — it breaks LoRA in-place
# weight scaling (scale_lora_layers / unscale_lora_layers).
result = pipe(
image=image,
prompt=prompt,
negative_prompt=negative_prompt if guidance_scale > 1.0 else None,
num_inference_steps=int(steps),
generator=generator,
true_cfg_scale=float(guidance_scale),
).images[0]
return result, seed
except torch.cuda.OutOfMemoryError:
_empty_cuda()
raise gr.Error(
"GPU out of memory. Try reducing inference steps or using fewer LoRAs."
)
except RuntimeError as e:
msg = str(e)
if "no kernel image" in msg.lower():
raise gr.Error(
"Attention kernel is not compatible with this GPU. "
"The Space should be using SDPA; restart the Space after updating app.py."
)
if "CUDA" in msg or "out of memory" in msg.lower():
_empty_cuda()
raise gr.Error(f"GPU error: {e}")
raise gr.Error(f"Inference failed: {e}")
finally:
pipe.disable_lora()
_empty_cuda()
# ── UI ─────────────────────────────────────────────────────────────────────────
css = """
#col-container { margin: 0 auto; max-width: 980px; }
#main-title h1 { font-size: 2.25em !important; letter-spacing: -0.02em; }
.gr-button { transition: all 0.1s ease; }
.gr-button:hover { transform: translateY(-1px); box-shadow: 0 4px 12px rgba(0,0,0,0.15); }
/* Better image preview */
.contain-preview .image-frame img {
object-fit: contain !important;
max-height: min(78vh, 920px) !important;
width: auto !important;
max-width: 100% !important;
margin: 0 auto;
border-radius: 12px;
box-shadow: 0 10px 30px rgba(0,0,0,0.1);
}
/* LoRA rows alignment */
.gr-row { align-items: center; }
"""
LORA_NAMES = ["None"] + sorted(LORA_CONFIGS.keys())
with gr.Blocks(css=css, theme=steel_blue_theme) as demo:
with gr.Column(elem_id="col-container"):
gr.Markdown("# **Qwen-Image-Edit • 2509**", elem_id="main-title")
gr.Markdown(
"**Fast local LoRA-powered image editing** powered by `Qwen-Image-Edit-2509` + `Qwen-Image-Edit-Rapid-AIO-V4` + 50+ community LoRAs.
"
"Upload an image, describe the edit, optionally stack LoRAs, and hit **Edit Image**."
)
with gr.Accordion("💡 Quick Tips", open=False):
gr.Markdown(
"- **LoRAs auto-add trigger words** to your prompt when selected.
"
"- Keep total LoRA strength under ~2.5 to avoid artifacts.
"
"- Use **Randomize Seed** for variations of the same edit.
"
"- For best results, start with 4-8 steps and CFG 1.0–2.0."
)
with gr.Row(equal_height=False):
with gr.Column():
input_image = gr.Image(
label="Input Image",
type="pil",
elem_classes=["contain-preview"],
)
prompt = gr.Textbox(
label="Edit Prompt",
placeholder="e.g. change clothing...",
lines=3,
)
run_button = gr.Button("✨ Edit Image", variant="primary", size="lg")
with gr.Column():
with gr.Column():
output_image = gr.Image(
label="Output",
interactive=False,
format="png",
elem_classes=["contain-preview"],
)
with gr.Row():
use_as_input_btn = gr.Button("↻ Use as Input", size="sm", variant="secondary")
clear_output_btn = gr.Button("✕ Clear", size="sm", variant="secondary")
with gr.Accordion("➕ Extra LoRAs (optional)", open=False):
with gr.Row():
gr.Markdown("**Stack up to 6 LoRAs.** Select a LoRA to auto-fill trigger words in the prompt.")
clear_btn = gr.Button("🧹 Clear All", size="sm", variant="secondary", scale=0)
lora_stack = []
for i in range(6):
with gr.Row():
dd = gr.Dropdown(
choices=LORA_NAMES,
value="None",
label=f"LoRA {i + 1}",
scale=3,
interactive=True,
)
sl = gr.Slider(
0.0, 1.5, value=0.75,
step=0.05, label="Strength", scale=2,
)
lora_stack.extend([dd, sl])
clear_btn.click(
fn=clear_lora_stack,
inputs=None,
outputs=lora_stack,
)
with gr.Accordion("⚙️ Advanced", open=False):
negative_prompt = gr.Textbox(
label="Negative Prompt",
value=DEFAULT_NEGATIVE_PROMPT,
lines=2,
placeholder="Customize what to avoid...",
)
seed = gr.Slider(
label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0
)
randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
guidance_scale = gr.Slider(
label="CFG Scale", minimum=1.0, maximum=5.0, step=0.1, value=1.0
)
steps = gr.Slider(
label="Steps", minimum=1, maximum=30, step=1, value=4
)
run_button.click(
fn=infer,
inputs=[input_image, prompt, seed, randomize_seed, guidance_scale, steps, negative_prompt]
+ lora_stack,
outputs=[output_image, seed],
)
# Auto-fill trigger words when a LoRA is selected
for i in range(0, len(lora_stack), 2):
lora_stack[i].change(
fn=append_triggers,
inputs=[prompt, lora_stack[i]],
outputs=[prompt],
)
use_as_input_btn.click(
fn=lambda img: img,
inputs=[output_image],
outputs=[input_image],
)
clear_output_btn.click(
fn=lambda: gr.update(value=None),
inputs=None,
outputs=[output_image],
)
if __name__ == "__main__":
demo.queue(max_size=30).launch(
mcp_server=True, ssr_mode=False, show_error=True
)