import sys import os # Get the full path to the ComfyUI folder and its comfy submodule comfyui_path = os.path.join(os.getcwd(), "ComfyUI") comfy_module_path = os.path.join(comfyui_path, "comfy") # Add both to sys.path so Python can import from them if comfyui_path not in sys.path: sys.path.append(comfyui_path) if comfy_module_path not in sys.path: sys.path.append(comfy_module_path) import sys sys.path.append("ComfyUI") import gradio as gr from gnx_flux_lora import import_custom_nodes, NODE_CLASS_MAPPINGS, get_value_at_index import torch import random # Initialize once (model loading outside inference function for Hugging Face ZeroGPU) import_custom_nodes() unetloader = NODE_CLASS_MAPPINGS["UNETLoader"]() unetloader_3 = unetloader.load_unet( unet_name="flux1-dev.sft", weight_dtype="default" ) dualcliploader = NODE_CLASS_MAPPINGS["DualCLIPLoader"]() dualcliploader_8 = dualcliploader.load_clip( clip_name1="clip_l.safetensors", clip_name2="t5xxl_fp16.safetensors", type="flux", device="default", ) loraloader = NODE_CLASS_MAPPINGS["LoraLoader"]() loraloader_12 = loraloader.load_lora( lora_name="gauravjuneja4/gauravjuneja4.safetensors", strength_model=1, strength_clip=1, model=get_value_at_index(unetloader_3, 0), clip=get_value_at_index(dualcliploader_8, 0), ) vaeloader = NODE_CLASS_MAPPINGS["VAELoader"]() vae_9 = vaeloader.load_vae(vae_name="ae.safetensors") cliptextencode = NODE_CLASS_MAPPINGS["CLIPTextEncode"]() fluxguidance = NODE_CLASS_MAPPINGS["FluxGuidance"]() ksampler = NODE_CLASS_MAPPINGS["KSampler"]() vaedecode = NODE_CLASS_MAPPINGS["VAEDecode"]() emptylatentimage = NODE_CLASS_MAPPINGS["EmptyLatentImage"]() def run_inference(prompt: str): with torch.inference_mode(): # Text encodings positive = cliptextencode.encode( text=prompt, clip=get_value_at_index(loraloader_12, 1), ) negative = cliptextencode.encode( text="", clip=get_value_at_index(loraloader_12, 1), ) latent = emptylatentimage.generate( width=1024, height=1024, batch_size=1 ) guided = fluxguidance.append( guidance=3.5, conditioning=get_value_at_index(positive, 0) ) sample = ksampler.sample( seed=random.randint(1, 2**64), steps=20, cfg=1, sampler_name="euler", scheduler="simple", denoise=1, model=get_value_at_index(loraloader_12, 0), positive=get_value_at_index(guided, 0), negative=get_value_at_index(negative, 0), latent_image=get_value_at_index(latent, 0), ) decoded = vaedecode.decode( samples=get_value_at_index(sample, 0), vae=get_value_at_index(vae_9, 0), ) result = get_value_at_index(decoded, 0) return result # Gradio UI demo = gr.Interface( fn=run_inference, inputs=gr.Textbox(label="Prompt", placeholder="e.g. gjnx is driving ferrari on a road in germany"), outputs=gr.Image(label="Generated Image"), title="GNX Flux LoRA Generator", description="Enter a prompt using your trained LoRA with Flux .1 Dev." ) if __name__ == "__main__": demo.launch()