Model Personalities — Same Scene, Six Models

Choosing a model isn’t just a quality decision — it’s a creative one. Every model has a visual personality shaped by its text encoder, training data, and architecture. This guide runs the same concepts through six models so you can see the difference for yourself.

All images were generated on an RTX 4090 with modl v0.2.9. Every image includes full params so you can reproduce them exactly.

Quick reference

If you already know what you’re after:

The rest of this guide is the evidence. Read on to see why each recommendation holds.

Before we compare: three things people get wrong

1. Text encoders don’t read the same prompt the same way

This is the most important thing to understand before comparing models. Each model family uses a different language model to interpret your prompt:

This means the same English sentence means different things to different encoders. “A Pomeranian warrior in golden plate armor guarding an ancient Greek temple at dawn” is parsed as keywords by CLIP (pomeranian, warrior, golden, armor, greek, temple, dawn) but understood as a full compositional instruction by Qwen3-8B.

Write prompts for your model’s encoder:

2. Seeds are not comparable across models

Same seed on Klein 9B and SDXL produces completely different starting noise — different schedulers, different latent spaces, different architectures. Seed 42 on Klein and seed 42 on SDXL share nothing.

Same seed is useful when comparing the same model with one variable changed (prompt A vs B, step count, LoRA strength). It’s meaningless across models.

3. When to control variables vs. when to explore

• Same seed, one variable changed: testing prompt wording or LoRA settings on the same model
• Multiple seeds, same prompt: evaluating a model’s range and consistency
• Different models, same concept: what this guide does — comparing personalities, not noise

The models

Six models, four text encoder architectures, wildly different generation times:

Scene 1: The Oracle’s Chamber

All images: seed 2847, 1024×1024. SDXL prompt includes ”, masterpiece, best quality, dramatic lighting”.

This prompt has six elements that need to coexist: statue, candlelight, pool, smoke, incense burners, inscriptions. The Qwen-encoded models (Klein, Z-Image) parse this as a spatial layout — they place elements deliberately because the encoder understands “smoke rises FROM bronze incense burners” as a relationship, not two keywords. CLIP-based SDXL picks up the aesthetic keywords (ancient, golden, candlelight) but treats element placement as optional. Chroma’s T5 encoder sits in between — it understands the language but without CLIP’s aesthetic anchor, it leans into mood over precision.

Scene 2: The Clockwork Cathedral

All images: seed 9012, 1024×1024.

The key detail here: “stained glass windows depicting mechanical angels.” That’s a compound concept — the encoder needs to understand that the angels themselves should look mechanical, not that there are mechanical things near stained glass near angels. Qwen-encoded models parse this as a nested instruction. CLIP-based models treat “mechanical,” “angels,” and “stained glass” as independent keywords and let the diffusion model sort it out.

Look at the material rendering too. The prompt specifies brass, copper, stone, and glass — four distinct surfaces in one scene. How each model differentiates these materials tells you about its training data distribution across textures and lighting interactions.

Scene 3: Celestial Garden

All images: seed 3388, 1024×1024.

This is the most abstract prompt in the set, and it’s where model personalities diverge the most. “Fireflies made of tiny stars” and “liquid starlight” are metaphorical — no training image is literally labeled this way. Each model has to extrapolate. The differences you see here are pure personality: how each model’s latent space interpolates between “firefly” and “star,” between “liquid” and “starlight.” Models trained on more diverse artistic data (Chroma, SDXL) tend to interpret these metaphors more freely. Models optimized for photographic accuracy (Klein, Z-Image) try to make them physically plausible.

Character LoRAs: same dog, different worlds

All three scenes above used the same concept without a specific character. Now we add one: Maxi, a Pomeranian with a trained LoRA on each model. Same prompt, same seed, different LoRA per model.

All images: seed 4242, 1024×1024. Each model uses its own Maxi LoRA trained on the same 24-photo dataset. SDXL prompt includes quality tags.

Each LoRA was trained on the exact same 24 Pomeranian photos — same trigger word, same dataset prep (see the character LoRA guide for the full training breakdown). The identity is the same dog. But the fur texture, the temple aesthetic, the lighting treatment — all different. The LoRA captures who the character is. The base model decides what the world around him looks like.

Style LoRAs: the aesthetic override

Style LoRAs add another dimension — they force a trained aesthetic onto any model. But even then, the base model’s personality bleeds through. Here’s the same concept rendered through a kids drawing style LoRA on two different models:

Dragon: seed 1122. Underwater: seed 6677. Both 1024×1024.

The LoRA defines the style; the model defines how that style is rendered. Z-Image Turbo’s version has its characteristic sharpness — clean lines, vivid color separation. SDXL’s version is softer, with more blended crayon-like textures. Same LoRA training data, same concept — different personality underneath.

SDXL’s superpower: the LoRA ecosystem

SDXL came out in 2023 — three generations behind the newest models. But it has something no newer model can match: thousands of community-trained style LoRAs that transform it into entirely different art tools.

Pixel art

Fairy tale / Ghibli

Midjourney aesthetic (MJ52)

The MJ52 LoRA brings that distinctive Midjourney v5.2 look to local generation — rich detail, cinematic color grading, hyper-polished compositions:

Z-Image Turbo: dioramas and material rendering

Z-Image Turbo generates in 6 seconds and has a distinctive sharp, volumetric rendering style. The same characteristics that make it less ideal for soft organic portraits make it exceptional for hard surfaces — glass, metal, stone, and miniature-scale detail.

All Z-Image Turbo, 8 steps, 1024×1024. ~6s each on RTX 4090.

Traditional art: the Qwen encoder advantage

Traditional Asian art styles are where the Qwen-encoded models pull ahead most visibly. It’s not just about style — it’s about the encoder’s training data. Qwen was trained on multilingual data including Chinese and Japanese, so concepts like “shan shui” and traditional composition principles are represented directly in the encoder’s vocabulary, not approximated through English translations.

All three: seed 1199, 1024×1024.

Compare how each model handles the ink gradient — the gradual fade from solid black to transparent wash. That’s a specific technique (渲染, xuànrǎn) with its own conventions. Chroma’s T5 encoder understands the English description of what ink wash looks like, but the Qwen-encoded models have a more direct representation of the tradition itself.

Chroma: cinematic mood

Chroma’s T5-only encoder and Apache 2.0 training data give it a distinctive visual personality — moody, painterly, with a film-still quality. It’s also the only Flux-architecture model that supports negative prompts, which gives you direct control over what not to render.

All three: seed 4455, 1024×1024.

The Viking fjord comparison distills each model’s default mood. Chroma goes for atmosphere — mist, reflection, stillness. Z-Image Turbo renders the geometry with precision — crisp mountain edges, sharp water reflections. Klein 9B finds a middle ground, leaning photographic. Three valid interpretations of the same scene; your preference is the tiebreaker.

Gallery

Same prompt, three models, no commentary needed.

Botanical illustration

”Detailed botanical illustration of exotic orchids and hummingbirds, scientific accuracy with artistic beauty, hand-painted watercolor on aged cream paper, golden labels in elegant calligraphy, natural history museum quality” · seed 3322

Art Nouveau · Dutch still life

What to take from this

Every image in this guide was a single modl generate command — no compositing, no upscaling, no manual retouching.