zjourney - Fantasy Realism Refiner Ideogram 4

A lightweight LoRA trained on 44 curated images generated with the MidjourneyV1 Z Image Turbo LoRA, built to refine Ideogram 4 fantasy outputs toward a more realism aesthetic.

Why I made this: Ideogram 4 produces stunning fantasy art, but that's often the problem. No matter how I prompted it (photographic, photorealism, cinematic realism, etc.), fantasy subject matter almost always defaults to an illustrated or painterly art style. zjourney was built to push past that and ground those outputs in something that feels genuinely photographic.

What it does: Refines Ideogram 4 outputs with richer photorealism, improving surface texturing, cinematic lighting, environmental depth, and material detail across fantasy and action subjects. It also works well as a general realism refiner beyond pure fantasy content.

zjourney - Fantasy Realism Refiner | v2 Update

V2 expands the training dataset, focusing on further grounding outputs in photographic realism. Trained on 44 film stills alongside 6 MidjourneyV1 Z Image Turbo LoRA generations.

Usage:

• Standalone recommended strength: 0.55 0.35-0.6

• For best results, stack with v1:

  • v1 @ 0.35 + v2 @ 0.45

• First 2 example images generated standalone with v2

• Final 5 example images generated with v1 + v2 combined

What's next - v3:
Currently curating a dedicated dataset focused on environmental detail and depth, with an emphasis on faces at a distance, one of the weaker areas in the current versions. This will be folded into the v2 dataset for v3 training.

Key improvements over base Ideogram 4:

• More detailed and grounded environments (ruins, foliage, wet stone)

• Richer skin, fur, and fabric textures

• Stronger cinematic color grading and atmospheric depth

• More convincing crowd and secondary figure detail

Known limitations:

• Close-up shots with heavy focus on skin texture can see a slight degradation in quality, works best at mid-range and wider compositions

Usage:

• Recommended strength: 0.4–0.7

• Sample images generated at 0.5

• Works best alongside detailed fantasy or action prompts; can also be applied to general realism scenes

• Works with and without trigger word, the trigger word was trained in the aesthetics placement in the json prompt structure