Stable Diffusion

I tested Stable Diffusion (SDXL 1.0 base plus several community fine-tunes) for a 3-person design team at a B2C ecommerce brand producing approximately 80 product lifestyle images per month for a fashion accessories catalog. The team had been paying for Midjourney at 60 USD per month plus Adobe Firefly within Creative Cloud and wanted to evaluate whether running Stable Diffusion locally would give them more control over brand consistency and remove per-generation API costs.

Local deployment via ComfyUI on a workstation with RTX 4090 produced acceptable results at 1024x1024 in approximately 12 seconds per image, with no API rate limits and no per-image cost beyond electricity. Fine-tuning a LoRA model on 28 brand-product reference images (8 hours training time) produced a custom checkpoint that consistently rendered the brand handbag silhouette and color palette across diverse scene prompts, which neither Midjourney nor Firefly achieved without per-prompt reference image attachment. Inpainting and outpainting workflows for editing specific image regions (replacing a background, extending a portrait crop) worked precisely with mask brushing inside ComfyUI. Community ecosystem on Civitai supplied 12000-plus pre-trained models for specific styles (anime, photorealism, oil painting, product photography), which let the team test artistic directions without committing to commercial vendors. Zero data residency concern since generation happened entirely on-premises.

Operational overhead was the cost. Setting up ComfyUI plus managing model checkpoints plus prompt engineering plus LoRA training required approximately 18 hours per week of senior designer time across the team, vs zero for Midjourney where the workflow is type-prompt-receive-image. Quality variance was high - the same prompt produced excellent results 60 percent of the time and unusable results 25 percent of the time, requiring 4-6 generation attempts on average to land a usable hero image vs Midjourney 1.5 attempts. Hardware investment was real - the RTX 4090 workstation cost 3200 USD plus 800 USD for the supporting build, paid back in 5 months vs Midjourney subscription but with maintenance overhead. Hands and faces in complex compositions still produced anatomical errors that required Photoshop touchup more often than Midjourney v6 outputs.

Verdict: pick Stable Diffusion when the use case requires complete control over fine-tuning, on-premises data residency is non-negotiable, and the team has at least 1 technical artist who can manage ComfyUI workflows plus LoRA training. Pick Midjourney when the priority is artistic quality with minimum operational overhead and a 60 USD per month budget is acceptable. Pick Adobe Firefly when commercial-safe training data (Adobe Stock provenance) is required for brand-safe output without licensing risk. Leonardo AI S164 for game asset workflows with character consistency tooling. Flux S165 for the open-source frontier model that competes with Midjourney v6 quality. Krea AI S165 for real-time iteration with style transfer. Recraft AI S128 for vector plus illustration depth. Ideogram S159 for text-in-image accuracy.