Skip to main content
KUNST & KI
INTERMEDIATE
11 MIN

Turn Any Photo into Ghibli Style – Free & Local with Flux Kontext (2026)

Real Ghibli look from any photo – your face and pose stay. The trick is a small, free Kontext LoRA. Tested locally on an RTX 3060 (12 GB), no subscription.

Turn a photo into real Ghibli style with Flux Kontext – a realistic portrait becomes a hand-painted Ghibli character while the face and pose stay the same
Turn a photo into real Ghibli style with Flux Kontext – a realistic portrait becomes a hand-painted Ghibli character while the face and pose stay the same

Drop your photo into one of the viral "Ghibli" generators and you get back a pretty character of someone else. Different face, different pose, often behind a subscription or a watermark.

Flux Kontext does the opposite: it repaints your photo instead of inventing a new one – face, pose and composition stay. But here's the catch almost every guide skips: the prompt "Studio Ghibli" alone gives only generic anime – glossy round eyes, no real Ghibli handwriting. The real look comes only with a small, free Kontext LoRA.

This guide shows both – the quick prompt route and the step to real Ghibli. Tested locally on an RTX 3060 with 12 GB of VRAM, free, no cloud required.

Updated: July 2026 · ComfyUI (current) · Flux.1 Kontext Dev as GGUF · LoRA Kontext-Style/Ghibli_lora

What you'll get:

  • Why Kontext keeps your subject – and why the prompt alone still isn't enough
  • The free Ghibli LoRA for Kontext – and exactly where the file goes
  • The full ComfyUI workflow – with the GGUF trick for 12 GB cards
  • The right prompt + the tested settings (LoRA strength, guidance)
  • Fine-tuning and fixes when the look drifts

Why Kontext keeps your subject

The key term is image conditioning. Kontext gets your original image not as a vague description but as a fixed reference point (the ReferenceLatent node). It applies the style change to your image – instead of rolling a brand-new picture from a text prompt.

In practice:

  • Same face, same pose, same framing – only the painting style changes.
  • Not a random anime character, but recognizably you (or your subject).
  • You can apply the same look consistently across a whole photo series.

That's half the battle. The other half is the style itself – and that's where it gets interesting.


Why the prompt alone isn't enough

The Flux Kontext base model barely knows "Ghibli" – the studio name is trademark-thin in the training data. Ask it for "Studio Ghibli style" via prompt and you get generic anime: big glossy eyes, garish colors, but not the soft, hand-painted Ghibli handwriting.

Comparison: the same prompt without the LoRA gives generic anime, with the Kontext Ghibli LoRA it produces a real Ghibli look
Comparison: the same prompt without the LoRA gives generic anime, with the Kontext Ghibli LoRA it produces a real Ghibli look

The fix is a LoRA – a small add-on file (~350 MB) that teaches the model a specific style. That's exactly what the Kontext-Style project is: a Ghibli LoRA trained specifically for Flux Kontext on paired image data. Free, local, one download – and the look lands.

🔑 Remember: Kontext provides the fidelity (your subject stays), the LoRA provides the real style. Only both together give real Ghibli instead of run-of-the-mill anime.


Requirements

WhatRecommendation
Graphics cardNVIDIA with ≥ 8 GB VRAM (tested on RTX 3060/12 GB)
SoftwareComfyUI (current version)
Custom nodeComfyUI-GGUF (for the GGUF loader)
Storage~ 12 GB (base models) + ~ 350 MB (LoRA)

💡 Tip: We use the GGUF variant of Flux Kontext (Q4_K_M, ~6.8 GB). GGUF is a compressed format – so a 12-billion-parameter model runs smoothly on a 12 GB card, where the full fp16 version (~24 GB) would be far too large.

⚠️ Warning: ComfyUI Desktop and the portable version do not share models. Put the files in the models folder of the version you actually launch.

RunPodEmpfohlenes Tool · Cloud-GPU

Keine starke Grafikkarte? Führe ComfyUI in der Cloud aus.

Flux & SDXL sind speicherhungrig. Statt ~1.800 € für eine eigene Karte mietest du bei RunPod eine 24-GB-GPU ab ca. 0,50 €/Stunde – und zahlst nur, was du nutzt.

GPU mieten & Flux starten →Affiliate-Link · für dich ohne Mehrkosten

Step 1: Install the custom node "ComfyUI-GGUF"

The GGUF loader isn't a default node – it comes via the Manager:

  1. Start ComfyUI → click Manager on the right.
  2. Open Custom Nodes Manager → search for ComfyUI-GGUFInstall.
  3. Fully restart ComfyUI.

Without this step, the "Unet Loader (GGUF)" node is missing later and the workflow stays red.


Step 2: Place the models and the Ghibli LoRA

Flux Kontext has four parts (main model, two text encoders, VAE). On top comes the Ghibli LoRA – that's the part that makes the real style.

FileSourceFolder
flux1-kontext-dev-Q4_K_M.gguf🤗 QuantStack/FLUX.1-Kontext-dev-GGUFmodels/unet/
t5xxl_fp8_e4m3fn.safetensors🤗 comfyanonymous/flux_text_encodersmodels/text_encoders/
clip_l.safetensors🤗 comfyanonymous/flux_text_encodersmodels/clip/
ae.safetensors🤗 black-forest-labs/FLUX.1-schnell (VAE)models/vae/
Ghibli_lora_weights.safetensors🤗 Kontext-Style/Ghibli_loramodels/loras/

The folder structure then looks like this:

📁 ComfyUI/
└── 📁 models/
    ├── 📁 unet/
    │   └── flux1-kontext-dev-Q4_K_M.gguf
    ├── 📁 text_encoders/
    │   └── t5xxl_fp8_e4m3fn.safetensors
    ├── 📁 clip/
    │   └── clip_l.safetensors
    ├── 📁 vae/
    │   └── ae.safetensors
    └── 📁 loras/
        └── Ghibli_lora_weights.safetensors
Which file goes where: unet, text_encoders, clip, vae and the Ghibli LoRA in the loras folder
Which file goes where: unet, text_encoders, clip, vae and the Ghibli LoRA in the loras folder

📌 Note: After copying, restart ComfyUI once, otherwise the models and LoRA won't show up in the dropdowns.


Step 3: Load the workflow – GGUF trick + hook in the LoRA

You build nothing by hand. ComfyUI ships the base workflow as a template:

  1. Click Workflow → Browse Templates → category Flux"Flux.1 Kontext Dev (Basic)". The full graph loads automatically.
  2. GGUF trick for 12 GB cards: delete the "Load Diffusion Model" node. Right-click → Add Node → bootleg → Unet Loader (GGUF) → select flux1-kontext-dev-Q4_K_M.gguf.
  3. Hook in the LoRA: right-click → Add Node → loaders → LoraLoaderModelOnly. Plug the GGUF loader's MODEL output into the LoRA node, choose Ghibli_lora_weights.safetensors there and set strength_model to 1.0. Route the LoRA node's MODEL output on to the KSampler.

Quickly check the remaining loaders:

  • DualCLIPLoadert5xxl_fp8_e4m3fn + clip_l, type: flux
  • VAELoaderae.safetensors
From photo to real Ghibli image in four steps: load models + LoRA, photo and prompt, Kontext sampling, result
From photo to real Ghibli image in four steps: load models + LoRA, photo and prompt, Kontext sampling, result

🔑 Remember: The chain is GGUF loader → LoRA loader → KSampler. The LoRA sits between model and sampler – that's the only way the Ghibli style activates.


Step 4: Load photo, write the prompt, generate

1. Load photo. Upload your source photo in the Load Image node. A clear portrait with good lighting works best.

2. Write the prompt. With the LoRA you need no prompt acrobatics – the recommended sentence is enough:

Turn this image into the Ghibli style.

3. Set the parameters. The tested values:

ParameterValueWhy
Samplereulerstable for Flux
Schedulersimplestandard for Kontext
Steps20clean look; 8 for a quick preview run
Guidance2.5Kontext standard – the LoRA does the style work
CFG1.0fixed for Flux
LoRA strength1.0full Ghibli look

4. Generate. Click Queue Prompt.

🧪 Try it: The first run takes ~2–3 minutes because both the model and the LoRA load into VRAM. Every further conversion runs at ~30–60 seconds on the RTX 3060.

Result: your photo becomes a hand-painted Ghibli character – face, pose and clothes stay, the painting style is real Ghibli.


Fine-tuning: LoRA strength controls the look

The one dial that makes the difference is strength_model on the LoRA node:

StrengthEffect
1.0Full Ghibli look – hand-painted, soft eyes, film feel
0.7–0.8Keeps more of the photo's realism (photo ↔ Ghibli blend)
0.5Just a light coat – clearly a photo, with a Ghibli tone

💡 Tip: The Kontext-Style project has not only Ghibli but a whole set of style LoRAs for Kontext (watercolor, cartoon, line-art and more). Same workflow, you just swap the LoRA file – so you build a whole style palette from one photo.


Doesn't look like Ghibli? Quick fixes

ProblemCauseFix
Models/LoRA not in dropdownComfyUI not restartedrestart ComfyUI
"Unet Loader (GGUF)" node missingComfyUI-GGUF not installedManager → install ComfyUI-GGUF
Result = generic animeLoRA not active / strength 0check the chain GGUF → LoRA → KSampler, set strength_model to 1.0
Look too weakLoRA strength too lowraise strength_model toward 1.0
Face becomes unrecognizableLoRA strength too high for your subjectlower to 0.8
Red nodes / errorsfile in the wrong foldercheck folders per the table above

No 12 GB card? Use the cloud

If your GPU won't keep up or lacks VRAM: run the exact same workflow (LoRA included) on a cloud GPU – an RTX 4090 (24 GB) from ~$0.59/h renders Kontext in seconds and with no quant compromise. The full setup is in our hub:

👉 Set up ComfyUI on RunPod (Network Volume)

RunPodEmpfohlenes Tool · Cloud-GPU

Keine starke Grafikkarte? Führe ComfyUI in der Cloud aus.

Flux & SDXL sind speicherhungrig. Statt ~1.800 € für eine eigene Karte mietest du bei RunPod eine 24-GB-GPU ab ca. 0,50 €/Stunde – und zahlst nur, was du nutzt.

GPU mieten & Flux starten →Affiliate-Link · für dich ohne Mehrkosten

Conclusion & next steps

Real Ghibli needs no subscription and no ChatGPT – just Flux Kontext plus a free LoRA. Kontext locks your subject in place, the LoRA brings the real style. The prompt alone only makes generic anime – the LoRA is the difference.

One photo, one LoRA, one sentence – and your snapshot becomes a scene that looks straight out of your favorite film.