Refactor UI structure and import spaces module

app.py

@@ -11,20 +11,9 @@ import spaces
 import gradio as gr
 from huggingface_hub import ModelCard
 import torch
-import numpy as np
 from pydantic import BaseModel
 from PIL import Image
 from diffusers import (
-    FluxPipeline,
-    FluxImg2ImgPipeline,
-    FluxInpaintPipeline,
-    FluxControlNetPipeline,
-    StableDiffusionXLPipeline,
-    StableDiffusionXLImg2ImgPipeline,
-    StableDiffusionXLInpaintPipeline,
-    StableDiffusionXLControlNetPipeline,
-    StableDiffusionXLControlNetImg2ImgPipeline,
-    StableDiffusionXLControlNetInpaintPipeline,
     AutoPipelineForText2Image,
     AutoPipelineForImage2Image,
     AutoPipelineForInpainting,
@@ -32,22 +21,12 @@ from diffusers import (
     AutoencoderKL,
     FluxControlNetModel,
     FluxMultiControlNetModel,
-    ControlNetModel,
 )
-from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
 from huggingface_hub import hf_hub_download
-from transformers import CLIPFeatureExtractor
-from photomaker import FaceAnalysis2
 from diffusers.schedulers import *
 from huggingface_hub import hf_hub_download
-from safetensors.torch import load_file
 from controlnet_aux.processor import Processor
-from photomaker import (
-    PhotoMakerStableDiffusionXLPipeline,
-    PhotoMakerStableDiffusionXLControlNetPipeline,
-    analyze_faces
-)
-from sd_embed.embedding_funcs import get_weighted_text_embeddings_sdxl, get_weighted_text_embeddings_flux1
+from sd_embed.embedding_funcs import get_weighted_text_embeddings_flux1
 
 
 # Initialize System
@@ -64,11 +43,6 @@ def load_sd():
             "repo_id": "black-forest-labs/FLUX.1-dev",
             "loader": "flux",
             "compute_type": torch.bfloat16,
-        },
-        {
-            "repo_id": "SG161222/RealVisXL_V4.0",
-            "loader": "xl",
-            "compute_type": torch.float16,
         }
     ]
 
@@ -76,96 +50,39 @@ def load_sd():
         try:
             model["pipeline"] = AutoPipelineForText2Image.from_pretrained(
                 model['repo_id'],
+                vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=torch.bfloat16).to(device),
                 torch_dtype = model['compute_type'],
                 safety_checker = None,
                 variant = "fp16"
             ).to(device)
-            model["pipeline"].enable_model_cpu_offload()
         except:
             model["pipeline"] = AutoPipelineForText2Image.from_pretrained(
                 model['repo_id'],
+                vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=torch.bfloat16).to(device),
                 torch_dtype = model['compute_type'],
                 safety_checker = None
             ).to(device)
-            model["pipeline"].enable_model_cpu_offload() 
+        
+        model["pipeline"].enable_model_cpu_offload()
 
 
     # VAE n Refiner
+    flux_vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=torch.bfloat16).to(device)
     sdxl_vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to(device)
     refiner = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-1.0", vae=sdxl_vae, torch_dtype=torch.float16, use_safetensors=True, variant="fp16").to(device)
     refiner.enable_model_cpu_offload()
 
 
-    # Safety Checker
-    safety_checker = StableDiffusionSafetyChecker.from_pretrained("CompVis/stable-diffusion-safety-checker").to(device)
-    feature_extractor = CLIPFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32", from_pt=True)
-
-
-    # Controlnets
-    controlnet_models = [
-        {
-            "repo_id": "xinsir/controlnet-depth-sdxl-1.0",
-            "name": "depth_xl",
-            "layers": ["depth"],
-            "loader": "xl",
-            "compute_type": torch.float16,
-        },
-        {
-            "repo_id": "xinsir/controlnet-canny-sdxl-1.0",
-            "name": "canny_xl",
-            "layers": ["canny"],
-            "loader": "xl",
-            "compute_type": torch.float16,
-        },
-        {
-            "repo_id": "xinsir/controlnet-openpose-sdxl-1.0",
-            "name": "openpose_xl",
-            "layers": ["pose"],
-            "loader": "xl",
-            "compute_type": torch.float16,
-        },
-        {
-            "repo_id": "xinsir/controlnet-scribble-sdxl-1.0",
-            "name": "scribble_xl",
-            "layers": ["scribble"],
-            "loader": "xl",
-            "compute_type": torch.float16,
-        },
-        {
-            "repo_id": "Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro",
-            "name": "flux1_union_pro",
-            "layers": ["canny_fl", "tile_fl", "depth_fl", "blur_fl", "pose_fl", "gray_fl", "low_quality_fl"],
-            "loader": "flux-multi",
-            "compute_type": torch.bfloat16,
-        }
-    ]
-
-    for controlnet in controlnet_models:
-        if controlnet["loader"] == "xl":
-            controlnet["controlnet"] = ControlNetModel.from_pretrained(
-                controlnet["repo_id"],
-                torch_dtype = controlnet['compute_type']
-            ).to(device)
-        elif controlnet["loader"] == "flux-multi":
-            controlnet["controlnet"] = FluxMultiControlNetModel([FluxControlNetModel.from_pretrained(
-                controlnet["repo_id"],
-                torch_dtype = controlnet['compute_type']
-            ).to(device)])
-        #TODO: Add support for flux only controlnet
-
-
-    # Face Detection (for PhotoMaker)
-    face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
-    face_detector.prepare(ctx_id=0, det_size=(640, 640))
-
-
-    # PhotoMaker V2 (for SDXL only)
-    photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker-V2", filename="photomaker-v2.bin", repo_type="model")
-
-    return device, models, sdxl_vae, refiner, safety_checker, feature_extractor, controlnet_models, face_detector, photomaker_ckpt
+    # ControlNet
+    controlnet = FluxMultiControlNetModel([FluxControlNetModel.from_pretrained(
+        "Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro",
+        torch_dtype=torch.bfloat16
+    ).to(device)])
+    
+    return device, models, flux_vae, sdxl_vae, refiner, controlnet
 
 
-device, models, sdxl_vae, refiner, safety_checker, feature_extractor, controlnet_models, face_detector, photomaker_ckpt = load_sd()
+device, models, flux_vae, sdxl_vae, refiner, controlnet = load_sd()
 
 
 # Models
@@ -178,13 +95,11 @@ class ControlNetReq(BaseModel):
         arbitrary_types_allowed=True
 
 
-class SDReq(BaseModel):
+class FluxReq(BaseModel):
     model: str = ""
     prompt: str = ""
-    negative_prompt: Optional[str] = "black-forest-labs/FLUX.1-dev"
     fast_generation: Optional[bool] = True
     loras: Optional[list] = []
-    embeddings: Optional[list] = []
     resize_mode: Optional[str] = "resize_and_fill" # resize_only, crop_and_resize, resize_and_fill
     scheduler: Optional[str] = "euler_fl"
     height: int = 1024
@@ -196,13 +111,12 @@ class SDReq(BaseModel):
     refiner: bool = False
     vae: bool = True
     controlnet_config: Optional[ControlNetReq] = None
-    photomaker_images: Optional[List[Image.Image]] = None
     
     class Config:
         arbitrary_types_allowed=True
 
 
-class SDImg2ImgReq(SDReq):
+class FluxImg2ImgReq(FluxReq):
     image: Image.Image
     strength: float = 1.0
     
@@ -210,115 +124,76 @@ class SDImg2ImgReq(SDReq):
         arbitrary_types_allowed=True
 
 
-class SDInpaintReq(SDImg2ImgReq):
+class FluxInpaintReq(FluxImg2ImgReq):
     mask_image: Image.Image
     
     class Config:
         arbitrary_types_allowed=True
 
 
-# Helper functions
-def get_controlnet(controlnet_config: ControlNetReq):
+# Helper Functions
+def get_control_mode(controlnet_config: ControlNetReq):
     control_mode = []
-    controlnet = []
+    layers = ["canny", "tile", "depth", "blur", "pose", "gray", "low_quality"]
     
-    for m in controlnet_models:
-        for c in controlnet_config.controlnets:
-            if c in m["layers"]:
-                control_mode.append(m["layers"].index(c))
-                controlnet.append(m["controlnet"])
+    for c in controlnet_config.controlnets:
+        if c in layers:
+            control_mode.append(layers.index(c))
     
-    return controlnet, control_mode
+    return control_mode
 
 
-def get_pipe(request: SDReq | SDImg2ImgReq | SDInpaintReq):
+def get_pipe(request: FluxReq | FluxImg2ImgReq | FluxInpaintReq):
     for m in models:
-        if m["repo_id"] == request.model:
-            pipeline = m['pipeline']
-            controlnet, control_mode = get_controlnet(request.controlnet_config) if request.controlnet_config else (None, None)
-            
+        if m['repo_id'] == request.model:
             pipe_args = {
-                "pipeline": pipeline,
-                "control_mode": control_mode,
+                "pipeline": m['pipeline'],
             }
+            
+            
+            # Set ControlNet config
             if request.controlnet_config:
-                pipe_args["controlnet"] = controlnet
-
-            if not request.photomaker_images:
-                if isinstance(request, SDReq):
-                    pipe_args['pipeline'] = AutoPipelineForText2Image.from_pipe(**pipe_args)
-                elif isinstance(request, SDImg2ImgReq):
-                    pipe_args['pipeline'] = AutoPipelineForImage2Image.from_pipe(**pipe_args)
-                elif isinstance(request, SDInpaintReq):
-                    pipe_args['pipeline'] = AutoPipelineForInpainting.from_pipe(**pipe_args)
-                else:
-                    raise ValueError(f"Unknown request type: {type(request)}")
-            elif isinstance(request, any([PhotoMakerStableDiffusionXLPipeline, PhotoMakerStableDiffusionXLControlNetPipeline])):
-                if request.controlnet_config:
-                    pipe_args['pipeline'] = PhotoMakerStableDiffusionXLControlNetPipeline.from_pipe(**pipe_args)
-                else:
-                    pipe_args['pipeline'] = PhotoMakerStableDiffusionXLPipeline.from_pipe(**pipe_args)
-            else:
-                raise ValueError(f"Invalid request type: {type(request)}")
-        
-    return pipe_args
-
-
-def load_scheduler(pipeline, scheduler):
-    schedulers = {
-        "dpmpp_2m": (DPMSolverMultistepScheduler, {}),
-        "dpmpp_2m_k": (DPMSolverMultistepScheduler, {"use_karras_sigmas": True}),
-        "dpmpp_2m_sde": (DPMSolverMultistepScheduler, {"algorithm_type": "sde-dpmsolver++"}),
-        "dpmpp_2m_sde_k": (DPMSolverMultistepScheduler, {"algorithm_type": "sde-dpmsolver++", "use_karras_sigmas": True}),
-        "dpmpp_sde": (DPMSolverSinglestepScheduler, {}),
-        "dpmpp_sde_k": (DPMSolverSinglestepScheduler, {"use_karras_sigmas": True}),
-        "dpm2": (KDPM2DiscreteScheduler, {}),
-        "dpm2_k": (KDPM2DiscreteScheduler, {"use_karras_sigmas": True}),
-        "dpm2_a": (KDPM2AncestralDiscreteScheduler, {}),
-        "dpm2_a_k": (KDPM2AncestralDiscreteScheduler, {"use_karras_sigmas": True}),
-        "euler": (EulerDiscreteScheduler, {}),
-        "euler_a": (EulerAncestralDiscreteScheduler, {}),
-        "heun": (HeunDiscreteScheduler, {}),
-        "lms": (LMSDiscreteScheduler, {}),
-        "lms_k": (LMSDiscreteScheduler, {"use_karras_sigmas": True}),
-        "deis": (DEISMultistepScheduler, {}),
-        "unipc": (UniPCMultistepScheduler, {}),
-        "fm_euler": (FlowMatchEulerDiscreteScheduler, {}),
-    }
-    scheduler_class, kwargs = schedulers.get(scheduler, (None, {}))
-    
-    if scheduler_class is not None:
-        scheduler = scheduler_class.from_config(pipeline.scheduler.config, **kwargs)
-    else:
-        raise ValueError(f"Unknown scheduler: {scheduler}")
-    
-    return scheduler
-
-
-def load_loras(pipeline, loras, fast_generation):
-    for i, lora in enumerate(loras):
-        pipeline.load_lora_weights(lora['repo_id'], adapter_name=f"lora_{i}")
-    adapter_names = [f"lora_{i}" for i in range(len(loras))]
-    adapter_weights = [lora['weight'] for lora in loras]
-    
-    if fast_generation:
-        hyper_lora = hf_hub_download(
-            "ByteDance/Hyper-SD",
-            "Hyper-FLUX.1-dev-8steps-lora.safetensors" if isinstance(pipeline, FluxPipeline) else "Hyper-SDXL-2steps-lora.safetensors"
-        )
-        hyper_weight = 0.125 if isinstance(pipeline, FluxPipeline) else 1.0
-        pipeline.load_lora_weights(hyper_lora, adapter_name="hyper_lora")
-        adapter_names.append("hyper_lora")
-        adapter_weights.append(hyper_weight)
-    
-    pipeline.set_adapters(adapter_names, adapter_weights)
-
-
-def load_xl_embeddings(pipeline, embeddings):
-    for embedding in embeddings:
-        state_dict = load_file(hf_hub_download(embedding['repo_id']))
-        pipeline.load_textual_inversion(state_dict['clip_g'], token=embedding['token'], text_encoder=pipeline.text_encoder_2, tokenizer=pipeline.tokenizer_2)
-        pipeline.load_textual_inversion(state_dict["clip_l"], token=embedding['token'], text_encoder=pipeline.text_encoder, tokenizer=pipeline.tokenizer)
+                pipe_args["control_mode"] = get_control_mode(request.controlnet_config)
+                pipe_args["controlnet"] = [controlnet]
+            
+            
+            # Choose Pipeline Mode
+            if isinstance(request, FluxReq):
+                pipe_args['pipeline'] = AutoPipelineForText2Image.from_pipe(**pipe_args)
+            elif isinstance(request, FluxImg2ImgReq):
+                pipe_args['pipeline'] = AutoPipelineForImage2Image.from_pipe(**pipe_args)
+            elif isinstance(request, FluxInpaintReq):
+                pipe_args['pipeline'] = AutoPipelineForInpainting.from_pipe(**pipe_args)
+            
+            
+            # Enable or Disable Refiner
+            if request.vae:
+                pipe_args["pipeline"].vae = flux_vae
+            elif not request.vae:
+                pipe_args["pipeline"].vae = None
+            
+            
+            # Set Scheduler
+            pipe_args["pipeline"].scheduler = FlowMatchEulerDiscreteScheduler.from_config(pipe_args["pipeline"].scheduler.config)
+            
+            
+            # Set Loras
+            if request.loras:
+                for i, lora in enumerate(request.loras):
+                    pipe_args["pipeline"].load_lora_weights(request.lora['repo_id'], adapter_name=f"lora_{i}")
+                adapter_names = [f"lora_{i}" for i in range(len(request.loras))]
+                adapter_weights = [lora['weight'] for lora in request.loras]
+                
+                if request.fast_generation:
+                    hyper_lora = hf_hub_download("ByteDance/Hyper-SD", "Hyper-FLUX.1-dev-8steps-lora.safetensors")
+                    hyper_weight = 0.125
+                    pipe_args["pipeline"].load_lora_weights(hyper_lora, adapter_name="hyper_lora")
+                    adapter_names.append("hyper_lora")
+                    adapter_weights.append(hyper_weight)
+                
+                pipe_args["pipeline"].set_adapters(adapter_names, adapter_weights)
+            
+            return pipe_args
 
 
 def resize_images(images: List[Image.Image], height: int, width: int, resize_mode: str):
@@ -333,18 +208,16 @@ def resize_images(images: List[Image.Image], height: int, width: int, resize_mod
     return images
 
 
-def get_controlnet_images(controlnets: List[str], control_images: List[Image.Image], height: int, width: int, resize_mode: str):
+def get_controlnet_images(controlnet_config: ControlNetReq, height: int, width: int, resize_mode: str):
     response_images = []
-    control_images = resize_images(control_images, height, width, resize_mode)
-    for controlnet, image in zip(controlnets, control_images):
-        if controlnet == "canny" or controlnet == "canny_xs" or controlnet == "canny_fl":
+    control_images = resize_images(controlnet_config.control_images, height, width, resize_mode)
+    for controlnet, image in zip(controlnet_config.controlnets, control_images):
+        if controlnet == "canny":
             processor = Processor('canny')
-        elif controlnet == "depth" or controlnet == "depth_xs" or controlnet == "depth_fl":
+        elif controlnet == "depth":
             processor = Processor('depth_midas')
-        elif controlnet == "pose" or controlnet == "pose_fl":
+        elif controlnet == "pose":
             processor = Processor('openpose_full')
-        elif controlnet == "scribble":
-            processor = Processor('scribble')
         else:
             raise ValueError(f"Invalid Controlnet: {controlnet}")
     
@@ -353,72 +226,25 @@ def get_controlnet_images(controlnets: List[str], control_images: List[Image.Ima
     return response_images
 
 
-def check_image_safety(images: List[Image.Image]):
-    safety_checker_input = feature_extractor(images, return_tensors="pt").to("cuda")
-    has_nsfw_concepts = safety_checker(
-        images=[images],
-        clip_input=safety_checker_input.pixel_values.to("cuda"),
-    )
-    
-    return has_nsfw_concepts[1]
-
-
-def get_prompt_attention(pipeline, prompt, negative_prompt):
-    if isinstance(pipeline, (FluxPipeline, FluxImg2ImgPipeline, FluxInpaintPipeline, FluxControlNetPipeline)):
-        prompt_embeds, pooled_prompt_embeds = get_weighted_text_embeddings_flux1(pipeline, prompt)
-        return prompt_embeds, None, pooled_prompt_embeds, None
-    elif isinstance(pipeline, StableDiffusionXLPipeline):
-        prompt_embeds, prompt_neg_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds = get_weighted_text_embeddings_sdxl(pipeline, prompt, negative_prompt)
-        return prompt_embeds, prompt_neg_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
-    else:
-        raise ValueError(f"Invalid pipeline type: {type(pipeline)}")
-
-
-def get_photomaker_images(photomaker_images: List[Image.Image], height: int, width: int, resize_mode: str):
-    image_input_ids = []
-    image_id_embeds = []
-    photomaker_images = resize_images(photomaker_images, height, width, resize_mode)
-    
-    for image in photomaker_images:
-        image_input_ids.append(img)
-        img = np.array(image)[:, :, ::-1]
-        faces = analyze_faces(face_detector, image)
-        if len(faces) > 0:
-            image_id_embeds.append(torch.from_numpy(faces[0]['embeddings']))
-        else:
-            raise ValueError("No face detected in the image")
-    
-    return image_input_ids, image_id_embeds
+def get_prompt_attention(pipeline, prompt):
+    return get_weighted_text_embeddings_flux1(pipeline, prompt)
 
 
-def cleanup(pipeline, loras = None, embeddings = None):
+def cleanup(pipeline, loras = None):
     if loras:
-        pipeline.disable_lora()
         pipeline.unload_lora_weights()
-    if embeddings:
-        pipeline.unload_textual_inversion()
     gc.collect()
     torch.cuda.empty_cache()
 
 
-# Gen function
-def gen_img(
-    request: SDReq | SDImg2ImgReq | SDInpaintReq
-):
-    pipeline_args = get_pipe(request)
-    pipeline = pipeline_args['pipeline']
+# Gen Function
+def gen_img(request: FluxReq | FluxImg2ImgReq | FluxInpaintReq):
+    pipe_args = get_pipe(request)
+    pipeline = pipe_args["pipeline"]
     try:
-        pipeline.scheduler = load_scheduler(pipeline, request.scheduler)
+        positive_prompt_embeds, positive_prompt_pooled = get_prompt_attention(pipeline, request.prompt)
         
-        load_loras(pipeline, request.loras, request.fast_generation)
-        load_xl_embeddings(pipeline, request.embeddings)
-        
-        control_images = get_controlnet_images(request.controlnet_config.controlnets, request.controlnet_config.control_images, request.height, request.width, request.resize_mode) if request.controlnet_config else None
-        photomaker_images, photomaker_id_embeds = get_photomaker_images(request.photomaker_images, request.height, request.width) if request.photomaker_images else (None, None)
-        
-        positive_prompt_embeds, negative_prompt_embeds, positive_prompt_pooled, negative_prompt_pooled = get_prompt_attention(pipeline, request.prompt, request.negative_prompt)
-        
-        # Common args
+        # Common Args
         args = {
             'prompt_embeds': positive_prompt_embeds,
             'pooled_prompt_embeds': positive_prompt_pooled,
@@ -430,54 +256,32 @@ def gen_img(
             'generator': [torch.Generator(device=device).manual_seed(request.seed + i) if not request.seed is any([None, 0, -1]) else torch.Generator(device=device).manual_seed(random.randint(0, 2**32 - 1)) for i in range(request.num_images_per_prompt)],
         }
         
-        if isinstance(pipeline, any([StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline, StableDiffusionXLInpaintPipeline,
-                                     StableDiffusionXLControlNetPipeline, StableDiffusionXLControlNetImg2ImgPipeline, StableDiffusionXLControlNetInpaintPipeline])):
-            args['clip_skip'] = request.clip_skip
-            args['negative_prompt_embeds'] = negative_prompt_embeds
-            args['negative_pooled_prompt_embeds'] = negative_prompt_pooled
-        
-        if isinstance(pipeline, FluxControlNetPipeline) and request.controlnet_config:
-            args['control_mode'] = pipeline_args['control_mode']
-            args['control_image'] = control_images
-            args['controlnet_conditioning_scale'] = request.controlnet_conditioning_scale
-        
-        if not isinstance(pipeline, FluxControlNetPipeline) and request.controlnet_config:
-            args['controlnet_conditioning_scale'] = request.controlnet_conditioning_scale
+        if request.controlnet_config:
+            args['control_mode'] = get_control_mode(request.controlnet_config)
+            args['control_images'] = get_controlnet_images(request.controlnet_config, request.height, request.width, request.resize_mode)
+            args['controlnet_conditioning_scale'] = request.controlnet_config.controlnet_conditioning_scale
         
-            if isinstance(request, SDReq):
-                args['image'] = control_images
-            elif isinstance(request, (SDImg2ImgReq, SDInpaintReq)):
-                args['control_image'] = control_images
-        
-        if request.photomaker_images and isinstance(pipeline, any([PhotoMakerStableDiffusionXLPipeline, PhotoMakerStableDiffusionXLControlNetPipeline])):
-            args['input_id_images'] = photomaker_images
-            args['input_id_embeds'] = photomaker_id_embeds
-            args['start_merge_step'] = 10
-        
-        if isinstance(request, SDImg2ImgReq):
-            args['image'] = resize_images([request.image], request.height, request.width, request.resize_mode)
-            args['strength'] = request.strength
-        elif isinstance(request, SDInpaintReq):
-            args['image'] = resize_images([request.image], request.height, request.width, request.resize_mode)
-            args['mask_image'] = resize_images([request.mask_image], request.height, request.width, request.resize_mode)
+        if isinstance(request, (FluxImg2ImgReq, FluxInpaintReq)):
+            args['image'] = resize_images([request.image], request.height, request.width, request.resize_mode)[0]
             args['strength'] = request.strength
         
+        if isinstance(request, FluxInpaintReq):
+            args['mask_image'] = resize_images([request.mask_image], request.height, request.width, request.resize_mode)[0]
+        
+        # Generate
         images = pipeline(**args).images
         
+        # Refiner
         if request.refiner:
-            images = refiner(
-                prompt=request.prompt,
-                num_inference_steps=40,
-                denoising_start=0.7,
-                image=images.images
-            ).images
+            images = refiner(image=images, prompt=request.prompt, num_inference_steps=40, denoising_start=0.7).images
         
-        cleanup(pipeline, request.loras, request.embeddings)
+        cleanup(pipeline, request.loras)
         
         return images
     except Exception as e:
-        cleanup(pipeline, request.loras, request.embeddings)
-        raise ValueError(f"Error generating image: {e}") from e
+        cleanup(pipeline, request.loras)
+        raise gr.Error(f"Error: {e}")
+
 
 
 # CSS
@@ -730,18 +534,16 @@ def generate_image(
             "vae": vae,
             "controlnet_config": None,
         }
-        base_args = SDReq(**base_args)
-
+        base_args = FluxReq(**base_args)
+        
         if len(enabled_loras) > 0:
             base_args.loras = []
-            for enabled_lora, lora_slider in zip(enabled_loras, [lora_slider_0, lora_slider_1, lora_slider_2, lora_slider_3, lora_slider_4, lora_slider_5]):
-                if enabled_lora.get("repo_id", None):
-                    base_args.loras.append(
-                        {
-                            "repo_id": enabled_lora["repo_id"],
-                            "weight": lora_slider
-                        }
-                    )
+            for enabled_lora, slider in zip(enabled_loras, [lora_slider_0, lora_slider_1, lora_slider_2, lora_slider_3, lora_slider_4, lora_slider_5]):
+                if enabled_lora['repo_id']:
+                    base_args.loras.append({
+                        "repo_id": enabled_lora['repo_id'],
+                        "weight": slider
+                    })
         
         image = None
         mask_image = None
@@ -751,7 +553,7 @@ def generate_image(
             image = img2img_image
             strength = float(img2img_strength)
             
-            base_args = SDImg2ImgReq(
+            base_args = FluxImg2ImgReq(
                 **base_args.__dict__,
                 image=image,
                 strength=strength
@@ -761,7 +563,7 @@ def generate_image(
             mask_image = inpaint_image['layers'][0] if image else None
             strength = float(inpaint_strength)
             
-            base_args = SDInpaintReq(
+            base_args = FluxInpaintReq(
                 **base_args.__dict__,
                 image=image,
                 mask_image=mask_image,
@@ -775,27 +577,23 @@ def generate_image(
             )
             
             if canny_image:
-                base_args.controlnet_config.controlnets.append("canny_fl")
+                base_args.controlnet_config.controlnets.append("canny")
                 base_args.controlnet_config.control_images.append(canny_image)
                 base_args.controlnet_config.controlnet_conditioning_scale.append(float(canny_strength))
             if pose_image:
-                base_args.controlnet_config.controlnets.append("pose_fl")
+                base_args.controlnet_config.controlnets.append("pose")
                 base_args.controlnet_config.control_images.append(pose_image)
                 base_args.controlnet_config.controlnet_conditioning_scale.append(float(pose_strength))
             if depth_image:
-                base_args.controlnet_config.controlnets.append("depth_fl")
+                base_args.controlnet_config.controlnets.append("depth")
                 base_args.controlnet_config.control_images.append(depth_image)
                 base_args.controlnet_config.controlnet_conditioning_scale.append(float(depth_strength))
         else:
-            base_args = SDReq(**base_args.__dict__)
-
-        images = gen_img(base_args)
+            base_args = FluxReq(**base_args.__dict__)
         
-        return (
-            gr.update(
-                value=images,
-                interactive=True
-            )
+        return gr.update(
+            value=gen_img(base_args),
+            interactive=True
         )
 
 

app3.py

@@ -0,0 +1,1018 @@
+# Testing one file gradio app for zero gpu spaces not working as expected.
+# Check here for the issue: https://huggingface.co/spaces/zero-gpu-explorers/README/discussions/106#66e278a396acd45223e0d00b
+
+import os
+import gc
+import json
+import random
+from typing import List, Optional
+
+import spaces
+import gradio as gr
+from huggingface_hub import ModelCard
+import torch
+import numpy as np
+from pydantic import BaseModel
+from PIL import Image
+from diffusers import (
+    FluxPipeline,
+    FluxImg2ImgPipeline,
+    FluxInpaintPipeline,
+    FluxControlNetPipeline,
+    StableDiffusionXLPipeline,
+    StableDiffusionXLImg2ImgPipeline,
+    StableDiffusionXLInpaintPipeline,
+    StableDiffusionXLControlNetPipeline,
+    StableDiffusionXLControlNetImg2ImgPipeline,
+    StableDiffusionXLControlNetInpaintPipeline,
+    AutoPipelineForText2Image,
+    AutoPipelineForImage2Image,
+    AutoPipelineForInpainting,
+    DiffusionPipeline,
+    AutoencoderKL,
+    FluxControlNetModel,
+    FluxMultiControlNetModel,
+    ControlNetModel,
+)
+from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
+from huggingface_hub import hf_hub_download
+from transformers import CLIPFeatureExtractor
+from photomaker import FaceAnalysis2
+from diffusers.schedulers import *
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file
+from controlnet_aux.processor import Processor
+from photomaker import (
+    PhotoMakerStableDiffusionXLPipeline,
+    PhotoMakerStableDiffusionXLControlNetPipeline,
+    analyze_faces
+)
+from sd_embed.embedding_funcs import get_weighted_text_embeddings_sdxl, get_weighted_text_embeddings_flux1
+
+
+# Initialize System
+os.system("pip install --upgrade pip")
+
+
+def load_sd():
+    # device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+
+    # Models
+    models = [
+        {
+            "repo_id": "black-forest-labs/FLUX.1-dev",
+            "loader": "flux",
+            "compute_type": torch.bfloat16,
+        },
+        {
+            "repo_id": "SG161222/RealVisXL_V4.0",
+            "loader": "xl",
+            "compute_type": torch.float16,
+        }
+    ]
+
+    for model in models:
+        try:
+            model["pipeline"] = AutoPipelineForText2Image.from_pretrained(
+                model['repo_id'],
+                torch_dtype = model['compute_type'],
+                safety_checker = None,
+                variant = "fp16"
+            ).to(device)
+            model["pipeline"].enable_model_cpu_offload()
+        except:
+            model["pipeline"] = AutoPipelineForText2Image.from_pretrained(
+                model['repo_id'],
+                torch_dtype = model['compute_type'],
+                safety_checker = None
+            ).to(device)
+            model["pipeline"].enable_model_cpu_offload() 
+
+
+    # VAE n Refiner
+    sdxl_vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to(device)
+    refiner = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-1.0", vae=sdxl_vae, torch_dtype=torch.float16, use_safetensors=True, variant="fp16").to(device)
+    refiner.enable_model_cpu_offload()
+
+
+    # Safety Checker
+    safety_checker = StableDiffusionSafetyChecker.from_pretrained("CompVis/stable-diffusion-safety-checker").to(device)
+    feature_extractor = CLIPFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32", from_pt=True)
+
+
+    # Controlnets
+    controlnet_models = [
+        {
+            "repo_id": "xinsir/controlnet-depth-sdxl-1.0",
+            "name": "depth_xl",
+            "layers": ["depth"],
+            "loader": "xl",
+            "compute_type": torch.float16,
+        },
+        {
+            "repo_id": "xinsir/controlnet-canny-sdxl-1.0",
+            "name": "canny_xl",
+            "layers": ["canny"],
+            "loader": "xl",
+            "compute_type": torch.float16,
+        },
+        {
+            "repo_id": "xinsir/controlnet-openpose-sdxl-1.0",
+            "name": "openpose_xl",
+            "layers": ["pose"],
+            "loader": "xl",
+            "compute_type": torch.float16,
+        },
+        {
+            "repo_id": "xinsir/controlnet-scribble-sdxl-1.0",
+            "name": "scribble_xl",
+            "layers": ["scribble"],
+            "loader": "xl",
+            "compute_type": torch.float16,
+        },
+        {
+            "repo_id": "Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro",
+            "name": "flux1_union_pro",
+            "layers": ["canny_fl", "tile_fl", "depth_fl", "blur_fl", "pose_fl", "gray_fl", "low_quality_fl"],
+            "loader": "flux-multi",
+            "compute_type": torch.bfloat16,
+        }
+    ]
+
+    for controlnet in controlnet_models:
+        if controlnet["loader"] == "xl":
+            controlnet["controlnet"] = ControlNetModel.from_pretrained(
+                controlnet["repo_id"],
+                torch_dtype = controlnet['compute_type']
+            ).to(device)
+        elif controlnet["loader"] == "flux-multi":
+            controlnet["controlnet"] = FluxMultiControlNetModel([FluxControlNetModel.from_pretrained(
+                controlnet["repo_id"],
+                torch_dtype = controlnet['compute_type']
+            ).to(device)])
+        #TODO: Add support for flux only controlnet
+
+
+    # Face Detection (for PhotoMaker)
+    face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
+    face_detector.prepare(ctx_id=0, det_size=(640, 640))
+
+
+    # PhotoMaker V2 (for SDXL only)
+    photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker-V2", filename="photomaker-v2.bin", repo_type="model")
+
+    return device, models, sdxl_vae, refiner, safety_checker, feature_extractor, controlnet_models, face_detector, photomaker_ckpt
+
+
+device, models, sdxl_vae, refiner, safety_checker, feature_extractor, controlnet_models, face_detector, photomaker_ckpt = load_sd()
+
+
+# Models
+class ControlNetReq(BaseModel):
+    controlnets: List[str] # ["canny", "tile", "depth"]
+    control_images: List[Image.Image]
+    controlnet_conditioning_scale: List[float]
+    
+    class Config:
+        arbitrary_types_allowed=True
+
+
+class SDReq(BaseModel):
+    model: str = ""
+    prompt: str = ""
+    negative_prompt: Optional[str] = "black-forest-labs/FLUX.1-dev"
+    fast_generation: Optional[bool] = True
+    loras: Optional[list] = []
+    embeddings: Optional[list] = []
+    resize_mode: Optional[str] = "resize_and_fill" # resize_only, crop_and_resize, resize_and_fill
+    scheduler: Optional[str] = "euler_fl"
+    height: int = 1024
+    width: int = 1024
+    num_images_per_prompt: int = 1
+    num_inference_steps: int = 8
+    guidance_scale: float = 3.5
+    seed: Optional[int] = 0
+    refiner: bool = False
+    vae: bool = True
+    controlnet_config: Optional[ControlNetReq] = None
+    photomaker_images: Optional[List[Image.Image]] = None
+    
+    class Config:
+        arbitrary_types_allowed=True
+
+
+class SDImg2ImgReq(SDReq):
+    image: Image.Image
+    strength: float = 1.0
+    
+    class Config:
+        arbitrary_types_allowed=True
+
+
+class SDInpaintReq(SDImg2ImgReq):
+    mask_image: Image.Image
+    
+    class Config:
+        arbitrary_types_allowed=True
+
+
+# Helper functions
+def get_controlnet(controlnet_config: ControlNetReq):
+    control_mode = []
+    controlnet = []
+    
+    for m in controlnet_models:
+        for c in controlnet_config.controlnets:
+            if c in m["layers"]:
+                control_mode.append(m["layers"].index(c))
+                controlnet.append(m["controlnet"])
+    
+    return controlnet, control_mode
+
+
+def get_pipe(request: SDReq | SDImg2ImgReq | SDInpaintReq):
+    for m in models:
+        if m["repo_id"] == request.model:
+            pipeline = m['pipeline']
+            controlnet, control_mode = get_controlnet(request.controlnet_config) if request.controlnet_config else (None, None)
+            
+            pipe_args = {
+                "pipeline": pipeline,
+                "control_mode": control_mode,
+            }
+            if request.controlnet_config:
+                pipe_args["controlnet"] = controlnet
+
+            if not request.photomaker_images:
+                if isinstance(request, SDReq):
+                    pipe_args['pipeline'] = AutoPipelineForText2Image.from_pipe(**pipe_args)
+                elif isinstance(request, SDImg2ImgReq):
+                    pipe_args['pipeline'] = AutoPipelineForImage2Image.from_pipe(**pipe_args)
+                elif isinstance(request, SDInpaintReq):
+                    pipe_args['pipeline'] = AutoPipelineForInpainting.from_pipe(**pipe_args)
+                else:
+                    raise ValueError(f"Unknown request type: {type(request)}")
+            elif isinstance(request, any([PhotoMakerStableDiffusionXLPipeline, PhotoMakerStableDiffusionXLControlNetPipeline])):
+                if request.controlnet_config:
+                    pipe_args['pipeline'] = PhotoMakerStableDiffusionXLControlNetPipeline.from_pipe(**pipe_args)
+                else:
+                    pipe_args['pipeline'] = PhotoMakerStableDiffusionXLPipeline.from_pipe(**pipe_args)
+            else:
+                raise ValueError(f"Invalid request type: {type(request)}")
+        
+    return pipe_args
+
+
+def load_scheduler(pipeline, scheduler):
+    schedulers = {
+        "dpmpp_2m": (DPMSolverMultistepScheduler, {}),
+        "dpmpp_2m_k": (DPMSolverMultistepScheduler, {"use_karras_sigmas": True}),
+        "dpmpp_2m_sde": (DPMSolverMultistepScheduler, {"algorithm_type": "sde-dpmsolver++"}),
+        "dpmpp_2m_sde_k": (DPMSolverMultistepScheduler, {"algorithm_type": "sde-dpmsolver++", "use_karras_sigmas": True}),
+        "dpmpp_sde": (DPMSolverSinglestepScheduler, {}),
+        "dpmpp_sde_k": (DPMSolverSinglestepScheduler, {"use_karras_sigmas": True}),
+        "dpm2": (KDPM2DiscreteScheduler, {}),
+        "dpm2_k": (KDPM2DiscreteScheduler, {"use_karras_sigmas": True}),
+        "dpm2_a": (KDPM2AncestralDiscreteScheduler, {}),
+        "dpm2_a_k": (KDPM2AncestralDiscreteScheduler, {"use_karras_sigmas": True}),
+        "euler": (EulerDiscreteScheduler, {}),
+        "euler_a": (EulerAncestralDiscreteScheduler, {}),
+        "heun": (HeunDiscreteScheduler, {}),
+        "lms": (LMSDiscreteScheduler, {}),
+        "lms_k": (LMSDiscreteScheduler, {"use_karras_sigmas": True}),
+        "deis": (DEISMultistepScheduler, {}),
+        "unipc": (UniPCMultistepScheduler, {}),
+        "fm_euler": (FlowMatchEulerDiscreteScheduler, {}),
+    }
+    scheduler_class, kwargs = schedulers.get(scheduler, (None, {}))
+    
+    if scheduler_class is not None:
+        scheduler = scheduler_class.from_config(pipeline.scheduler.config, **kwargs)
+    else:
+        raise ValueError(f"Unknown scheduler: {scheduler}")
+    
+    return scheduler
+
+
+def load_loras(pipeline, loras, fast_generation):
+    for i, lora in enumerate(loras):
+        pipeline.load_lora_weights(lora['repo_id'], adapter_name=f"lora_{i}")
+    adapter_names = [f"lora_{i}" for i in range(len(loras))]
+    adapter_weights = [lora['weight'] for lora in loras]
+    
+    if fast_generation:
+        hyper_lora = hf_hub_download(
+            "ByteDance/Hyper-SD",
+            "Hyper-FLUX.1-dev-8steps-lora.safetensors" if isinstance(pipeline, FluxPipeline) else "Hyper-SDXL-2steps-lora.safetensors"
+        )
+        hyper_weight = 0.125 if isinstance(pipeline, FluxPipeline) else 1.0
+        pipeline.load_lora_weights(hyper_lora, adapter_name="hyper_lora")
+        adapter_names.append("hyper_lora")
+        adapter_weights.append(hyper_weight)
+    
+    pipeline.set_adapters(adapter_names, adapter_weights)
+
+
+def load_xl_embeddings(pipeline, embeddings):
+    for embedding in embeddings:
+        state_dict = load_file(hf_hub_download(embedding['repo_id']))
+        pipeline.load_textual_inversion(state_dict['clip_g'], token=embedding['token'], text_encoder=pipeline.text_encoder_2, tokenizer=pipeline.tokenizer_2)
+        pipeline.load_textual_inversion(state_dict["clip_l"], token=embedding['token'], text_encoder=pipeline.text_encoder, tokenizer=pipeline.tokenizer)
+
+
+def resize_images(images: List[Image.Image], height: int, width: int, resize_mode: str):
+    for image in images:
+        if resize_mode == "resize_only":
+            image = image.resize((width, height))
+        elif resize_mode == "crop_and_resize":
+            image = image.crop((0, 0, width, height))
+        elif resize_mode == "resize_and_fill":
+            image = image.resize((width, height), Image.Resampling.LANCZOS)
+
+    return images
+
+
+def get_controlnet_images(controlnets: List[str], control_images: List[Image.Image], height: int, width: int, resize_mode: str):
+    response_images = []
+    control_images = resize_images(control_images, height, width, resize_mode)
+    for controlnet, image in zip(controlnets, control_images):
+        if controlnet == "canny" or controlnet == "canny_xs" or controlnet == "canny_fl":
+            processor = Processor('canny')
+        elif controlnet == "depth" or controlnet == "depth_xs" or controlnet == "depth_fl":
+            processor = Processor('depth_midas')
+        elif controlnet == "pose" or controlnet == "pose_fl":
+            processor = Processor('openpose_full')
+        elif controlnet == "scribble":
+            processor = Processor('scribble')
+        else:
+            raise ValueError(f"Invalid Controlnet: {controlnet}")
+    
+        response_images.append(processor(image, to_pil=True))
+    
+    return response_images
+
+
+def check_image_safety(images: List[Image.Image]):
+    safety_checker_input = feature_extractor(images, return_tensors="pt").to("cuda")
+    has_nsfw_concepts = safety_checker(
+        images=[images],
+        clip_input=safety_checker_input.pixel_values.to("cuda"),
+    )
+    
+    return has_nsfw_concepts[1]
+
+
+def get_prompt_attention(pipeline, prompt, negative_prompt):
+    if isinstance(pipeline, (FluxPipeline, FluxImg2ImgPipeline, FluxInpaintPipeline, FluxControlNetPipeline)):
+        prompt_embeds, pooled_prompt_embeds = get_weighted_text_embeddings_flux1(pipeline, prompt)
+        return prompt_embeds, None, pooled_prompt_embeds, None
+    elif isinstance(pipeline, StableDiffusionXLPipeline):
+        prompt_embeds, prompt_neg_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds = get_weighted_text_embeddings_sdxl(pipeline, prompt, negative_prompt)
+        return prompt_embeds, prompt_neg_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+    else:
+        raise ValueError(f"Invalid pipeline type: {type(pipeline)}")
+
+
+def get_photomaker_images(photomaker_images: List[Image.Image], height: int, width: int, resize_mode: str):
+    image_input_ids = []
+    image_id_embeds = []
+    photomaker_images = resize_images(photomaker_images, height, width, resize_mode)
+    
+    for image in photomaker_images:
+        image_input_ids.append(img)
+        img = np.array(image)[:, :, ::-1]
+        faces = analyze_faces(face_detector, image)
+        if len(faces) > 0:
+            image_id_embeds.append(torch.from_numpy(faces[0]['embeddings']))
+        else:
+            raise ValueError("No face detected in the image")
+    
+    return image_input_ids, image_id_embeds
+
+
+def cleanup(pipeline, loras = None, embeddings = None):
+    if loras:
+        pipeline.disable_lora()
+        pipeline.unload_lora_weights()
+    if embeddings:
+        pipeline.unload_textual_inversion()
+    gc.collect()
+    torch.cuda.empty_cache()
+
+
+# Gen function
+def gen_img(
+    request: SDReq | SDImg2ImgReq | SDInpaintReq
+):
+    pipeline_args = get_pipe(request)
+    pipeline = pipeline_args['pipeline']
+    try:
+        pipeline.scheduler = load_scheduler(pipeline, request.scheduler)
+        
+        load_loras(pipeline, request.loras, request.fast_generation)
+        load_xl_embeddings(pipeline, request.embeddings)
+        
+        control_images = get_controlnet_images(request.controlnet_config.controlnets, request.controlnet_config.control_images, request.height, request.width, request.resize_mode) if request.controlnet_config else None
+        photomaker_images, photomaker_id_embeds = get_photomaker_images(request.photomaker_images, request.height, request.width) if request.photomaker_images else (None, None)
+        
+        positive_prompt_embeds, negative_prompt_embeds, positive_prompt_pooled, negative_prompt_pooled = get_prompt_attention(pipeline, request.prompt, request.negative_prompt)
+        
+        # Common args
+        args = {
+            'prompt_embeds': positive_prompt_embeds,
+            'pooled_prompt_embeds': positive_prompt_pooled,
+            'height': request.height,
+            'width': request.width,
+            'num_images_per_prompt': request.num_images_per_prompt,
+            'num_inference_steps': request.num_inference_steps,
+            'guidance_scale': request.guidance_scale,
+            'generator': [torch.Generator(device=device).manual_seed(request.seed + i) if not request.seed is any([None, 0, -1]) else torch.Generator(device=device).manual_seed(random.randint(0, 2**32 - 1)) for i in range(request.num_images_per_prompt)],
+        }
+        
+        if isinstance(pipeline, any([StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline, StableDiffusionXLInpaintPipeline,
+                                     StableDiffusionXLControlNetPipeline, StableDiffusionXLControlNetImg2ImgPipeline, StableDiffusionXLControlNetInpaintPipeline])):
+            args['clip_skip'] = request.clip_skip
+            args['negative_prompt_embeds'] = negative_prompt_embeds
+            args['negative_pooled_prompt_embeds'] = negative_prompt_pooled
+        
+        if isinstance(pipeline, FluxControlNetPipeline) and request.controlnet_config:
+            args['control_mode'] = pipeline_args['control_mode']
+            args['control_image'] = control_images
+            args['controlnet_conditioning_scale'] = request.controlnet_conditioning_scale
+        
+        if not isinstance(pipeline, FluxControlNetPipeline) and request.controlnet_config:
+            args['controlnet_conditioning_scale'] = request.controlnet_conditioning_scale
+        
+            if isinstance(request, SDReq):
+                args['image'] = control_images
+            elif isinstance(request, (SDImg2ImgReq, SDInpaintReq)):
+                args['control_image'] = control_images
+        
+        if request.photomaker_images and isinstance(pipeline, any([PhotoMakerStableDiffusionXLPipeline, PhotoMakerStableDiffusionXLControlNetPipeline])):
+            args['input_id_images'] = photomaker_images
+            args['input_id_embeds'] = photomaker_id_embeds
+            args['start_merge_step'] = 10
+        
+        if isinstance(request, SDImg2ImgReq):
+            args['image'] = resize_images([request.image], request.height, request.width, request.resize_mode)
+            args['strength'] = request.strength
+        elif isinstance(request, SDInpaintReq):
+            args['image'] = resize_images([request.image], request.height, request.width, request.resize_mode)
+            args['mask_image'] = resize_images([request.mask_image], request.height, request.width, request.resize_mode)
+            args['strength'] = request.strength
+        
+        images = pipeline(**args).images
+        
+        if request.refiner:
+            images = refiner(
+                prompt=request.prompt,
+                num_inference_steps=40,
+                denoising_start=0.7,
+                image=images.images
+            ).images
+        
+        cleanup(pipeline, request.loras, request.embeddings)
+        
+        return images
+    except Exception as e:
+        cleanup(pipeline, request.loras, request.embeddings)
+        raise ValueError(f"Error generating image: {e}") from e
+
+
+# CSS
+css = """
+@import url('https://fonts.googleapis.com/css2?family=Poppins:wght@300;400;600&display=swap');
+body {
+    font-family: 'Poppins', sans-serif !important;
+}
+.center-content {
+    text-align: center;
+    max-width: 600px;
+    margin: 0 auto;
+    padding: 20px;
+}
+.center-content h1 {
+    font-weight: 600;
+    margin-bottom: 1rem;
+}
+.center-content p {
+    margin-bottom: 1.5rem;
+}
+"""
+
+
+flux_models = ["black-forest-labs/FLUX.1-dev"]
+with open("data/images/loras/flux.json", "r") as f:
+    loras = json.load(f)
+
+
+# Event functions
+def update_fast_generation(model, fast_generation):
+    if fast_generation:
+        return (
+            gr.update(
+                value=3.5
+            ),
+            gr.update(
+                value=8
+            )
+        )
+
+
+def selected_lora_from_gallery(evt: gr.SelectData):
+    return (
+        gr.update(
+            value=evt.index
+        )
+    )
+
+
+def update_selected_lora(custom_lora):
+    link = custom_lora.split("/")
+    
+    if len(link) == 2:
+        model_card = ModelCard.load(custom_lora)
+        trigger_word = model_card.data.get("instance_prompt", "")
+        image_url = f"""https://huggingface.co/{custom_lora}/resolve/main/{model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None)}"""
+        
+        custom_lora_info_css = """
+        <style>
+            .custom-lora-info {
+                font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', 'Roboto', 'Oxygen', 'Ubuntu', 'Cantarell', 'Fira Sans', 'Droid Sans', 'Helvetica Neue', sans-serif;
+                background: linear-gradient(135deg, #4a90e2, #7b61ff);
+                color: white;
+                padding: 16px;
+                border-radius: 8px;
+                box-shadow: 0 4px 6px rgba(0, 0, 0, 0.1);
+                margin: 16px 0;
+            }
+            .custom-lora-header {
+                font-size: 18px;
+                font-weight: 600;
+                margin-bottom: 12px;
+            }
+            .custom-lora-content {
+                display: flex;
+                align-items: center;
+                background-color: rgba(255, 255, 255, 0.1);
+                border-radius: 6px;
+                padding: 12px;
+            }
+            .custom-lora-image {
+                width: 80px;
+                height: 80px;
+                object-fit: cover;
+                border-radius: 6px;
+                margin-right: 16px;
+            }
+            .custom-lora-text h3 {
+                margin: 0 0 8px 0;
+                font-size: 16px;
+                font-weight: 600;
+            }
+            .custom-lora-text small {
+                font-size: 14px;
+                opacity: 0.9;
+            }
+            .custom-trigger-word {
+                background-color: rgba(255, 255, 255, 0.2);
+                padding: 2px 6px;
+                border-radius: 4px;
+                font-weight: 600;
+            }
+        </style>
+        """
+
+        custom_lora_info_html = f"""
+        <div class="custom-lora-info">
+            <div class="custom-lora-header">Custom LoRA: {custom_lora}</div>
+            <div class="custom-lora-content">
+                <img class="custom-lora-image" src="{image_url}" alt="LoRA preview">
+                <div class="custom-lora-text">
+                    <h3>{link[1].replace("-", " ").replace("_", " ")}</h3>
+                    <small>{"Using: <span class='custom-trigger-word'>"+trigger_word+"</span> as the trigger word" if trigger_word else "No trigger word found. If there's a trigger word, include it in your prompt"}</small>
+                </div>
+            </div>
+        </div>
+        """
+
+        custom_lora_info_html = f"{custom_lora_info_css}{custom_lora_info_html}"
+
+        return (
+            gr.update( # selected_lora
+                value=custom_lora,
+            ),
+            gr.update( # custom_lora_info
+                value=custom_lora_info_html,
+                visible=True
+            )
+        )
+
+    else:
+        return (
+            gr.update( # selected_lora
+                value=custom_lora,
+            ),
+            gr.update( # custom_lora_info
+                value=custom_lora_info_html if len(link) == 0 else "",
+                visible=False
+            )
+        )
+
+
+def add_to_enabled_loras(model, selected_lora, enabled_loras):
+    lora_data = loras
+    try:
+        selected_lora = int(selected_lora)
+        
+        if 0 <= selected_lora: # is the index of the lora in the gallery
+            lora_info = lora_data[selected_lora]
+            enabled_loras.append({
+                "repo_id": lora_info["repo"],
+                "trigger_word": lora_info["trigger_word"]
+            })
+    except ValueError:
+        link = selected_lora.split("/")
+        if len(link) == 2:
+            model_card = ModelCard.load(selected_lora)
+            trigger_word = model_card.data.get("instance_prompt", "")
+            enabled_loras.append({
+                "repo_id": selected_lora,
+                "trigger_word": trigger_word
+            })
+    
+    return (
+        gr.update( # selected_lora
+            value=""
+        ),
+        gr.update( # custom_lora_info
+            value="",
+            visible=False
+        ),
+        gr.update( # enabled_loras
+            value=enabled_loras
+        )
+    )
+
+
+def update_lora_sliders(enabled_loras):
+    sliders = []
+    remove_buttons = []
+    
+    for lora in enabled_loras:
+        sliders.append(
+            gr.update(
+                label=lora.get("repo_id", ""),
+                info=f"Trigger Word: {lora.get('trigger_word', '')}",
+                visible=True,
+                interactive=True
+            )
+        )
+        remove_buttons.append(
+            gr.update(
+                visible=True,
+                interactive=True
+            )
+        )
+    
+    if len(sliders) < 6:
+        for i in range(len(sliders), 6):
+            sliders.append(
+                gr.update(
+                    visible=False
+                )
+            )
+            remove_buttons.append(
+                gr.update(
+                    visible=False
+                )
+            )
+    
+    return *sliders, *remove_buttons
+
+
+def remove_from_enabled_loras(enabled_loras, index):
+    enabled_loras.pop(index)
+    return (
+        gr.update(
+            value=enabled_loras
+        )
+    )
+
+
+@spaces.GPU
+def generate_image(
+        model, prompt, negative_prompt, fast_generation, enabled_loras,
+        lora_slider_0, lora_slider_1, lora_slider_2, lora_slider_3, lora_slider_4, lora_slider_5,
+        img2img_image, inpaint_image, canny_image, pose_image, depth_image,
+        img2img_strength, inpaint_strength, canny_strength, pose_strength, depth_strength,
+        resize_mode,
+        scheduler, image_height, image_width, image_num_images_per_prompt,
+        image_num_inference_steps, image_guidance_scale, image_seed,
+        refiner, vae
+    ):
+        base_args = {
+            "model": model,
+            "prompt": prompt,
+            "negative_prompt": negative_prompt,
+            "fast_generation": fast_generation,
+            "loras": None,
+            "resize_mode": resize_mode,
+            "scheduler": scheduler,
+            "height": int(image_height),
+            "width": int(image_width),
+            "num_images_per_prompt": float(image_num_images_per_prompt),
+            "num_inference_steps": float(image_num_inference_steps),
+            "guidance_scale": float(image_guidance_scale),
+            "seed": int(image_seed),
+            "refiner": refiner,
+            "vae": vae,
+            "controlnet_config": None,
+        }
+        base_args = SDReq(**base_args)
+
+        if len(enabled_loras) > 0:
+            base_args.loras = []
+            for enabled_lora, lora_slider in zip(enabled_loras, [lora_slider_0, lora_slider_1, lora_slider_2, lora_slider_3, lora_slider_4, lora_slider_5]):
+                if enabled_lora.get("repo_id", None):
+                    base_args.loras.append(
+                        {
+                            "repo_id": enabled_lora["repo_id"],
+                            "weight": lora_slider
+                        }
+                    )
+        
+        image = None
+        mask_image = None
+        strength = None
+        
+        if img2img_image:
+            image = img2img_image
+            strength = float(img2img_strength)
+            
+            base_args = SDImg2ImgReq(
+                **base_args.__dict__,
+                image=image,
+                strength=strength
+            )
+        elif inpaint_image:
+            image = inpaint_image['background'] if not all(pixel == (0, 0, 0) for pixel in list(inpaint_image['background'].getdata())) else None
+            mask_image = inpaint_image['layers'][0] if image else None
+            strength = float(inpaint_strength)
+            
+            base_args = SDInpaintReq(
+                **base_args.__dict__,
+                image=image,
+                mask_image=mask_image,
+                strength=strength
+            )
+        elif any([canny_image, pose_image, depth_image]):
+            base_args.controlnet_config = ControlNetReq(
+                controlnets=[],
+                control_images=[],
+                controlnet_conditioning_scale=[]
+            )
+            
+            if canny_image:
+                base_args.controlnet_config.controlnets.append("canny_fl")
+                base_args.controlnet_config.control_images.append(canny_image)
+                base_args.controlnet_config.controlnet_conditioning_scale.append(float(canny_strength))
+            if pose_image:
+                base_args.controlnet_config.controlnets.append("pose_fl")
+                base_args.controlnet_config.control_images.append(pose_image)
+                base_args.controlnet_config.controlnet_conditioning_scale.append(float(pose_strength))
+            if depth_image:
+                base_args.controlnet_config.controlnets.append("depth_fl")
+                base_args.controlnet_config.control_images.append(depth_image)
+                base_args.controlnet_config.controlnet_conditioning_scale.append(float(depth_strength))
+        else:
+            base_args = SDReq(**base_args.__dict__)
+
+        images = gen_img(base_args)
+        
+        return (
+            gr.update(
+                value=images,
+                interactive=True
+            )
+        )
+
+
+# Main Gradio app
+with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
+    # Header
+    with gr.Column(elem_classes="center-content"):
+        gr.Markdown("""
+        # 🚀 AAI: All AI
+        Unleash your creativity with our multi-modal AI platform.
+        [![Sync code to HF Space](https://github.com/mantrakp04/aai/actions/workflows/hf-space.yml/badge.svg)](https://github.com/mantrakp04/aai/actions/workflows/hf-space.yml)
+        """)
+
+    # Tabs
+    with gr.Tabs():
+        with gr.Tab(label="🖼️ Image"):
+            with gr.Tabs():
+                with gr.Tab("Flux"):
+                    """
+                    Create the image tab for Generative Image Generation Models
+                    
+                    Args:
+                    models: list
+                        A list containing the models repository paths
+                    gap_iol, gap_la, gap_le, gap_eio, gap_io: Optional[List[dict]]
+                        A list of dictionaries containing the title and component for the custom gradio component
+                        Example:
+                        def gr_comp():
+                            gr.Label("Hello World")
+                        
+                        [
+                            {
+                                'title': "Title",
+                                'component': gr_comp()
+                            }
+                        ]
+                    loras: list
+                        A list of dictionaries containing the image and title for the Loras Gallery
+                        Generally a loaded json file from the data folder
+                    
+                    """
+                    def process_gaps(gaps: List[dict]):
+                        for gap in gaps:
+                            with gr.Accordion(gap['title']):
+                                gap['component']
+                    
+                    
+                    with gr.Row():
+                        with gr.Column():
+                            with gr.Group() as image_options:
+                                model = gr.Dropdown(label="Models", choices=flux_models, value=flux_models[0], interactive=True)
+                                prompt = gr.Textbox(lines=5, label="Prompt")
+                                negative_prompt = gr.Textbox(label="Negative Prompt")
+                                fast_generation = gr.Checkbox(label="Fast Generation (Hyper-SD) 🧪")
+                            
+                            
+                            with gr.Accordion("Loras", open=True): # Lora Gallery
+                                lora_gallery = gr.Gallery(
+                                    label="Gallery",
+                                    value=[(lora['image'], lora['title']) for lora in loras],
+                                    allow_preview=False,
+                                    columns=[3],
+                                    type="pil"
+                                )
+                                
+                                with gr.Group():
+                                    with gr.Column():
+                                        with gr.Row():
+                                            custom_lora = gr.Textbox(label="Custom Lora", info="Enter a Huggingface repo path")
+                                            selected_lora = gr.Textbox(label="Selected Lora", info="Choose from the gallery or enter a custom LoRA")
+                                        
+                                        custom_lora_info = gr.HTML(visible=False)
+                                        add_lora = gr.Button(value="Add LoRA")
+                                        
+                                        enabled_loras = gr.State(value=[])
+                                        with gr.Group():
+                                            with gr.Row():
+                                                for i in range(6): # only support max 6 loras due to inference time
+                                                    with gr.Column():
+                                                        with gr.Column(scale=2):
+                                                            globals()[f"lora_slider_{i}"] = gr.Slider(label=f"LoRA {i+1}", minimum=0, maximum=1, step=0.01, value=0.8, visible=False, interactive=True)
+                                                        with gr.Column():
+                                                            globals()[f"lora_remove_{i}"] = gr.Button(value="Remove LoRA", visible=False)
+
+                            
+                            with gr.Accordion("Embeddings", open=False): # Embeddings
+                                gr.Label("To be implemented")
+                            
+                            
+                            with gr.Accordion("Image Options"): # Image Options
+                                with gr.Tabs():
+                                    image_options = {
+                                        "img2img": "Upload Image",
+                                        "inpaint": "Upload Image",
+                                        "canny": "Upload Image",
+                                        "pose": "Upload Image",
+                                        "depth": "Upload Image",
+                                    }
+                                    
+                                    for image_option, label in image_options.items():
+                                        with gr.Tab(image_option):
+                                            if not image_option in ['inpaint', 'scribble']:
+                                                globals()[f"{image_option}_image"] = gr.Image(label=label, type="pil")
+                                            elif image_option in ['inpaint', 'scribble']:
+                                                globals()[f"{image_option}_image"] = gr.ImageEditor(
+                                                    label=label,
+                                                    image_mode='RGB',
+                                                    layers=False,
+                                                    brush=gr.Brush(colors=["#FFFFFF"], color_mode="fixed") if image_option == 'inpaint' else gr.Brush(),
+                                                    interactive=True,
+                                                    type="pil",
+                                                )
+                                            
+                                            # Image Strength (Co-relates to controlnet strength, strength for img2img n inpaint)
+                                            globals()[f"{image_option}_strength"] = gr.Slider(label="Strength", minimum=0, maximum=1, step=0.01, value=1.0, interactive=True)
+                                    
+                                    resize_mode = gr.Radio(
+                                        label="Resize Mode",
+                                        choices=["crop and resize", "resize only", "resize and fill"],
+                                        value="resize and fill",
+                                        interactive=True
+                                    )
+                        
+                        
+                        with gr.Column():
+                            with gr.Group():
+                                output_images = gr.Gallery(
+                                        label="Output Images",
+                                        value=[],
+                                        allow_preview=True,
+                                        type="pil",
+                                        interactive=False,
+                                    )
+                                generate_images = gr.Button(value="Generate Images", variant="primary")            
+                            
+                            with gr.Accordion("Advance Settings", open=True):
+                                with gr.Row():
+                                    scheduler = gr.Dropdown(
+                                        label="Scheduler",
+                                        choices = [
+                                            "fm_euler"
+                                        ],
+                                        value="fm_euler",
+                                        interactive=True
+                                    )
+
+                                with gr.Row():
+                                    for column in range(2):
+                                        with gr.Column():
+                                            options = [
+                                                ("Height", "image_height", 64, 1024, 64, 1024, True),
+                                                ("Width", "image_width", 64, 1024, 64, 1024, True),
+                                                ("Num Images Per Prompt", "image_num_images_per_prompt", 1, 4, 1, 1, True),
+                                                ("Num Inference Steps", "image_num_inference_steps", 1, 100, 1, 20, True),
+                                                ("Clip Skip", "image_clip_skip", 0, 2, 1, 2, False),
+                                                ("Guidance Scale", "image_guidance_scale", 0, 20, 0.5, 3.5, True),
+                                                ("Seed", "image_seed", 0, 100000, 1, random.randint(0, 100000), True),
+                                            ]
+                                            for label, var_name, min_val, max_val, step, value, visible in options[column::2]:
+                                                globals()[var_name] = gr.Slider(label=label, minimum=min_val, maximum=max_val, step=step, value=value, visible=visible, interactive=True)
+                                
+                                with gr.Row():
+                                    refiner = gr.Checkbox(
+                                        label="Refiner 🧪",
+                                        value=False,
+                                    )
+                                    vae = gr.Checkbox(
+                                        label="VAE",
+                                        value=True,
+                                    )
+
+
+                    # Events
+                    # Base Options
+                    fast_generation.change(update_fast_generation, [model, fast_generation], [image_guidance_scale, image_num_inference_steps]) # Fast Generation # type: ignore
+                    
+
+                    # Lora Gallery
+                    lora_gallery.select(selected_lora_from_gallery, None, selected_lora)
+                    custom_lora.change(update_selected_lora, custom_lora, [custom_lora, selected_lora])
+                    add_lora.click(add_to_enabled_loras, [model, selected_lora, enabled_loras], [selected_lora, custom_lora_info, enabled_loras])
+                    enabled_loras.change(update_lora_sliders, enabled_loras, [lora_slider_0, lora_slider_1, lora_slider_2, lora_slider_3, lora_slider_4, lora_slider_5, lora_remove_0, lora_remove_1, lora_remove_2, lora_remove_3, lora_remove_4, lora_remove_5]) # type: ignore
+
+                    for i in range(6):
+                        globals()[f"lora_remove_{i}"].click(
+                            lambda enabled_loras, index=i: remove_from_enabled_loras(enabled_loras, index),
+                            [enabled_loras],
+                            [enabled_loras]
+                        )
+                    
+
+                    # Generate Image
+                    generate_images.click(
+                        generate_image, # type: ignore
+                        [
+                            model, prompt, negative_prompt, fast_generation, enabled_loras,
+                            lora_slider_0, lora_slider_1, lora_slider_2, lora_slider_3, lora_slider_4, lora_slider_5, # type: ignore
+                            img2img_image, inpaint_image, canny_image, pose_image, depth_image, # type: ignore
+                            img2img_strength, inpaint_strength, canny_strength, pose_strength, depth_strength, # type: ignore
+                            resize_mode,
+                            scheduler, image_height, image_width, image_num_images_per_prompt, # type: ignore
+                            image_num_inference_steps, image_guidance_scale, image_seed, # type: ignore
+                            refiner, vae
+                        ],
+                        [output_images]
+                    )
+                with gr.Tab("SDXL"):
+                    gr.Label("To be implemented")
+        with gr.Tab(label="🎵 Audio"):
+            gr.Label("Coming soon!")
+        with gr.Tab(label="🎬 Video"):
+            gr.Label("Coming soon!")
+        with gr.Tab(label="📄 Text"):
+            gr.Label("Coming soon!")
+
+
+demo.launch(
+    share=False,
+    debug=True,
+)