import gradio as gr import transformers import os import re import json import random device = "cpu" model = None tokenizer = None def init_model(): global model, tokenizer model_id = os.environ.get("MODEL_ID") or "treadon/prompt-fungineer-355M" auth_token = os.environ.get("HUB_TOKEN") or True print(f"Using model {model_id}.") if auth_token != True: print("Using auth token.") model = transformers.AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True,use_auth_token=auth_token) tokenizer = transformers.AutoTokenizer.from_pretrained("gpt2") def format_prompt(prompt, enhancers=True, inspiration=False, negative_prompt=False): try: pattern = r"(BRF:|POS:|ENH:|INS:|NEG:) (.*?)(?= (BRF:|POS:|ENH:|INS:|NEG:)|$)" matches = re.findall(pattern, prompt) vals = {key: value.strip() for key, value,ex in matches} result = vals["POS:"] if enhancers: result += " " + vals["ENH:"] if inspiration: result += " " + vals["INS:"] if negative_prompt: result += "\n\n--no " + vals["NEG:"] return result except Exception as e: return "Failed to generate prompt." def generate_text(prompt, extra=False, top_k=100, top_p=0.95, temperature=0.85, enhancers = True, inpspiration = False , negative_prompt = False): global model, tokenizer try: if model is None: init_model() except Exception as e: print(e) return ["Try Again"] * 4 if model is None: return ["Try Again"] * 4 if not prompt.startswith("BRF:"): prompt = "BRF: " + prompt if not extra: prompt = prompt + " POS:" model.eval() # SOFT SAMPLE inputs = tokenizer(prompt, return_tensors="pt").to(device) samples = [] try: for i in range(1): print(f"Generating sample for prompt: {prompt}") outputs = model.generate(**inputs, max_length=256, do_sample=True, top_k=top_k, top_p=top_p, temperature=temperature, num_return_sequences=4, pad_token_id=tokenizer.eos_token_id) print(f"Generated {len(outputs)} samples.") for output in outputs: sample = tokenizer.decode(output, skip_special_tokens=True) sample = format_prompt(sample, enhancers, inpspiration, negative_prompt) samples.append(sample) except Exception as e: print(e) return samples if __name__ == "__main__": with gr.Blocks() as fungineer: with gr.Row(): gr.Markdown("""# Midjourney / Dalle 2 / Stable Diffusion Prompt Generator This is the 355M parameter model. There is also a 7B parameter model that is much better but far slower (access coming soon). Just enter a basic prompt and the fungineering model will use its wildest imagination to expand the prompt in detail. You can then use this prompt to generate images with Midjourney, Dalle 2, Stable Diffusion, Bing Image Creator, or any other image generation model. Treat this model more like a text-to-text model (simple prompt > complex prompt) rather than a generative model (prefix + word generation). It is a generative model under the hood.""") with gr.Row(): with gr.Column(): base_prompt = gr.Textbox(lines=5, label="Base Prompt", placeholder="An astronaut in space", info="Enter a very simple prompt that will be fungineered into something exciting!") extra = gr.Checkbox(value=True, label="Extra Fungineer Imagination", info="If checked, the model will be allowed to go wild with its imagination.") with gr.Accordion("Advanced Generation Settings", open=False): top_k = gr.Slider( minimum=10, maximum=1000, value=100, label="Top K", info="Top K sampling") top_p = gr.Slider( minimum=0.1, maximum=1, value=0.95, step=0.01, label="Top P", info="Top P sampling") temperature = gr.Slider( minimum=0.1, maximum=1.2, value=0.85, step=0.01, label="Temperature", info="Temperature sampling. Higher values will make the model more creative") with gr.Accordion("Advanced Output Settings", open=False): enh = gr.Checkbox(value=True, label="Enhancers", info="Add image meta information such as lens type, shuffter speed, camera model, etc.") insp = gr.Checkbox(value=False, label="Inpsiration", info="Include inspirational photographers that are known for this type of photography. Sometimes random people will appear here, needs more training.") neg = gr.Checkbox(value=False, label="Negative Prompt", info="Include a negative prompt, more often used in Stable Diffusion. If you're a Stable Diffusion user, chances are you already have a better negative prompt you like to use.") with gr.Column(): outputs = [ gr.Textbox(lines=2, label="Fungineered Text 1"), gr.Textbox(lines=2, label="Fungineered Text 2"), gr.Textbox(lines=2, label="Fungineered Text 3"), gr.Textbox(lines=2, label="Fungineered Text 4"), ] inputs = [base_prompt, extra, top_k, top_p, temperature, enh, insp, neg] submit = gr.Button(label="Fungineer",variant="primary") submit.click(generate_text, inputs=inputs, outputs=outputs) examples = [] with open("examples.json") as f: examples = json.load(f) for i, example in enumerate(examples): with gr.Tab(f"Example {i+1}", id=i): with gr.Row(): with gr.Column(): gr.Markdown(f"### Base Prompt") gr.Image(value=f"{example['base']['src']}") gr.Markdown(f"{example['base']['prompt']}") with gr.Column(): gr.Markdown(f"### 355M Prompt Fungineered") gr.Image(value=f"{example['355M']['src']}") gr.Markdown(f"{example['355M']['prompt']}") with gr.Column(): gr.Markdown(f"### 7B Prompt Fungineered") gr.Markdown(f"Coming Soon!") init_model() fungineer.launch(enable_queue=True, show_api=False, debug=True)