kitaniai/morgana-1.2

Kitani Morgana v1.2

The photography model that doesn't look like it came from a model.

Morgana v1.2 is Kitani's image model, built in partnership with ShuttleAI and trained in part on the Shuttle 3 Jaguar dataset. It's tuned for one thing above all: images you'd believe were shot on a camera. Golden-hour landscapes, candid phone selfies, macro textures, moody night streets and painted art, all with natural light, honest color and real-world detail.

What it does best

• Realistic photography - soft window light, deep shadows, film-like grain and subtle lens character instead of plastic HDR renders
• Human faces that feel human - front-camera selfies, candid moments, visible pores, flyaway hair and imperfect framing
• Macro and material detail - water droplets, feather barbs, fabric weave, wood grain, rusted metal, worn leather
• Cinematic landscapes - aurora mirrored in alpine lakes, waves breaking on black sand, mist in the pines
• Atmospheric nights - neon glowing on wet asphalt, believable low-light exposure
• Art styles on demand - thick impasto oil paintings, luminous watercolor, film-still anime and hyperreal 3D renders

Quick start

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "kitani/morgana-v1.2",
    torch_dtype=torch.bfloat16,
).to("cuda")

image = pipe(
    prompt="A realistic macro photograph of a water droplet on a waxy green leaf, fine veins, soft daylight, shallow depth of field.",
    width=1024,
    height=1024,
    num_inference_steps=28,
    guidance_scale=4.0,
    max_sequence_length=512,
    generator=torch.Generator("cuda").manual_seed(12345),
).images[0]

How to write prompts Morgana loves

Describe the shot like a photographer would:

• Lighting - "soft window light", "golden hour", "overcast sky", "neon reflections"
• Lens and framing - "85mm", "shallow depth of field", "wide-angle", "off-center"
• Texture and reality - "visible pores", "natural skin texture", "film grain", "subtle imperfections"
• Keep it clean - end prompts with "no text" when you don't want signage

Notes

• Negative prompts are not supported by this pipeline.
• Readable text and signage are not Morgana's strength — ask for clean scenes when it matters.
• The right prompt turns good into great; the examples above are a strong starting point.

Credits

Built by Kitani in partnership with ShuttleAI, using the Shuttle 3 Jaguar dataset.

License

Licensed by Kitani, Inc. (https://kitani.ai). See the base model card for underlying model terms.