Reproducing ChatGPT-4o's Ghibli Style Locally

Not trusting OpenAI? Concerned about privacy or subscription costs? Here's how you can create Studio Ghibli-style images with just one NVIDIA GPU and open-source tools.

TL;DR – jump straight to the hosted demos:
• Text-to-Image workflow 👉 https://agireact.com/workflow/ghibli_style
• Image-to-Image workflow 👉 https://agireact.com/gstyle

---

Why the Hype?

OpenAI's ChatGPT-4o introduced a Ghibli moment – social feeds are now saturated with whimsical, pastel-tinted illustrations reminiscent of Spirited Away and My Neighbor Totoro.

But using OpenAI's generator is either paid (ChatGPT Plus at $20/mo) or rate-limited (2 images/day for free users). Running your own model keeps data on-prem and costs nothing after the initial setup.

Hardware

• RTX 3090 (any 12 GB+ NVIDIA GPU works)

Software Stack

1. ComfyUI – node-based workflow editor
2. Flux models – Flux 1 / Flux 1-schnell or Stable Diffusion 3.5 M/L
3. LoRA fine-tunes – Flux-Ghibsky Illustration
4. ControlNet – for precise edge guidance (Image-to-Image)
5. Vision LLM – e.g. minicpm-v via ComfyUI-Ollama for automatic prompt extraction

Workflows

Text-to-Image

Simply load the Ghibli LoRA in your sampler node, set a base model (Flux), and start sampling.

Example prompt:

(ghibli style), a girl on a bicycle, sunset, soft pastel colours

Image-to-Image

1. Canny ControlNet on your photo
2. Vision LLM generates a descriptive prompt
3. Apply the Ghibli LoRA
4. Denoise with a low strength (0.3-0.5)

The full ComfyUI graph is in the repo and demoed on the site above.

Fun Fact

Indian users adopted the workflow fastest – uploading selfies to "Ghibli-fy" themselves!

Conclusion

Running locally gives you unlimited generations, lower latency, and full data ownership. Give it a go and share your Ghibli masterpieces!