Content Generation Pipeline with LLM and SDXL

You will build a content generation pipeline where you provide a topic and get back a complete blog post with a matching feature image — the LLM writes the article while SDXL generates the header image, both running in parallel on the same GPU. The end result: a marketing team submits “GPU hosting for AI startups” and receives a 600-word article plus a 1024×1024 feature image within 30 seconds. Here is the full pipeline on dedicated GPU infrastructure.

Pipeline Architecture

With sequential execution, a 24GB GPU handles this using the 8B model for both text and image prompts. For parallel execution, use a 48GB+ GPU or run text and image on separate servers.

Article Generation

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")

def generate_article(topic: str, tone: str = "professional", length: int = 600) -> dict:
    response = client.chat.completions.create(
        model="meta-llama/Llama-3.1-8B-Instruct",
        messages=[{
            "role": "system",
            "content": f"You are a content writer. Write a {tone} blog post "
                       f"of approximately {length} words. Include an engaging title, "
                       f"introduction, 3-4 sections with H2 headings, and a conclusion. "
                       f"Return as JSON: {{\"title\": \"\", \"meta_description\": \"\", "
                       f"\"content\": \"(HTML with h2 tags)\", \"image_concept\": \"\"}}"
        }, {"role": "user", "content": f"Write about: {topic}"}],
        max_tokens=2000, temperature=0.7
    )
    return parse_json(response.choices[0].message.content)

The model returns an image_concept field describing what the feature image should depict — this feeds the image generation stage. Deploy via vLLM for efficient serving.

Image Prompt Engineering

def create_sdxl_prompt(image_concept: str) -> str:
    response = client.chat.completions.create(
        model="meta-llama/Llama-3.1-8B-Instruct",
        messages=[{
            "role": "system",
            "content": "Convert this image concept into an SDXL prompt. "
                       "Include style keywords: professional, modern, high quality, "
                       "4k, clean composition. Keep under 75 tokens."
        }, {"role": "user", "content": image_concept}],
        max_tokens=100, temperature=0.5
    )
    return response.choices[0].message.content

LLMs produce better SDXL prompts than raw topic descriptions because they add the stylistic keywords that SDXL responds to.

Image Generation

import torch
from diffusers import StableDiffusionXLPipeline

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16
).to("cuda")

def generate_feature_image(prompt: str, output_path: str):
    image = pipe(
        prompt=prompt,
        negative_prompt="text, watermark, blurry, low quality",
        width=1024, height=576,  # 16:9 for blog headers
        num_inference_steps=30
    ).images[0]
    image.save(output_path)
    return output_path

Orchestrated Pipeline

from fastapi import FastAPI
app = FastAPI()

@app.post("/generate-content")
async def generate_content(request: dict):
    topic = request["topic"]

    # Stage 1: Generate article (includes image concept)
    article = generate_article(topic, request.get("tone", "professional"))

    # Stage 2: Generate matching image
    sdxl_prompt = create_sdxl_prompt(article["image_concept"])
    image_path = generate_feature_image(sdxl_prompt, f"/output/{slugify(topic)}.png")

    return {
        "title": article["title"],
        "meta_description": article["meta_description"],
        "content": article["content"],
        "image_url": image_path,
        "image_prompt": sdxl_prompt
    }

Quality and Brand Control

For production content pipelines: add brand voice guidelines to the system prompt with examples of approved writing style; maintain a negative prompt library for SDXL to enforce brand-consistent imagery; implement a review queue where editors approve content before publishing; integrate with ComfyUI for advanced image workflows with LoRA style adapters; and use RAG retrieval from existing published content to maintain consistency. Store content in a CMS-ready format. Teams should review AI-generated content before publication. See model options for larger models that produce higher-quality text, private hosting for brand-sensitive content, and more tutorials for related workflows. Check content use cases for industry examples.