Ray’s Models Zoo (Readme is WIP)

An AI repository of curated visual styles — a personal generative model collection.

TL;DR

A curated collection of custom-trained generative models built on personal assets across Flux-Dev, Krea, SDXL, Chroma, Qwen-Image, Z-Image, and others — all designed to deliver my own distinctive visual style, mostly photographic, but not exclusively.

Model Philosophy

All models in this zoo were trained on a unique, personally-curated dataset built around years of original street photography. Many of them were later expanded and refined with synthetic images generated from earlier generations of my own models — creating a self-reinforcing aesthetic loop.

The overarching goal is to capture a natural, analog-leaning photographic sensibility without sacrificing creative range. Many of the models inherit a subtle film-grain texture, balanced contrast, and an overall organic feel. Others explore alternate aesthetics — oil painting, digital illustration, stylized comic rendering — but always through the lens of my own visual taste.

Across the different families (SDXL, Flux, Qwen-Image, Chroma, Z-Image), the intention remains the same:
a deeply personal twist on photorealism and style-driven rendering, built from scratch and continuously refined.

These models are designed for everyone:

Artists and photographers wanting high-fidelity realism
Illustrators looking for stylistic control
Prompt engineers who want predictable behavior
Anyone exploring generative visuals with a distinctive mood

Some families are more specialized — like the SDXL branch, with portrait-centric Raymnants, painterly Rayctifier, and comic-styled Rayburn — while newer architectures such as Flux, Qwen-Image, and Z-Image tend to be more powerful generalists with a strong stylistic backbone.

It's also a proof-of-concept for blockwise LoRA merging as a method to finetune models rather than full training. Besides my SDXL models, most of the checkpoints here were done with the following process:

Train a LoRA on a specific concept (style or subject)
Turn the LoRA influence to 0 and turn it back on on every blocks of the models one by one to generate a control image
Identify the layers/blocks most affected my the concepts we want to reinforce
tune the values only in thos blocks by checking the control images until satisfactory
bake weights in the model
wash, rinse, repeat with a new LoRA

A typical vanilla-to-v1 transition takes usually 4-8 LoRAs. I oftentimes merge LoRAs thematically on a single model, but keep multiple model branches in parallel before merging them together (e.g. one Photo style branch, one character detail branch, etc.)

Repository Structure

CHROMA/
    Raydiant v1.0

FLUX/
    Rayflux Horndog
    Rayflux Photoplus
    Rayflux v1.0
    Rayflux v1.0 (UNet-only)
    Rayflux v3.0
    Rayflux v3.x (3.1–3.4)
    Rayflux v4.0

FLUX_KREA/
    Rayflux Krea v1.0

QWEN_IMAGE/
    RayQwest v0.65
    RayQwest v0.7
    RayQwest v1.0

SDXL/
    Rayburn 1.0
    RayburnBlitz 1.0
    Rayctifier 1.0
    Rayctifier 1.5
    Raymnants 3.0
    Raymnants 3.0 LightningW 4L1.0
    Raymnants 3.0 LightningW 4L2.0
    Raymnants 3.4
    Raymnants 3.4b
    Raymnants 4.0
    Raymnants 4.1
    Raymnants 5.0
    Raymnants 6.0
    Raymnants 6.1
    Raymnants 6.2

Z-IMAGE/
    Rayzist v1.0

Model Families Overview

CHROMA — Raydiant
Lightweight analog-photography oriented experiments, built on Chroma v48.
FLUX — Rayflux series
Flagship generalist line with analog grain, strong realism, and expressive control.
FLUX_KREA — Rayflux Krea
Experimental Flux-Krea refinement with excellent skin rendering and strong style.
QWEN_IMAGE — RayQwest
Qwen-Image based models that remove plasticness and restore analog grit and Ray-style texture.
SDXL — Raymnants / Rayctifier / Rayburn
Fast SDXL models spanning portrait realism, oil painting, and stylized comic rendering.
Z-IMAGE — Rayzist
Fast Turbo-tier model that performs above typical Turbo expectations, with good text rendering.

Flagship Lines

Rayflux (Flux)
RayQwest (Qwen-Image)
Raymnants (SDXL)
Rayzist (Z-Image Turbo)

FLUX Family — Rayflux Models

Default Rayflux settings (baseline suggestion):
Sampler: Euler · Guidance: Beta 2.5 · CFG: 1 · Steps: 16 · Shift / Max Shift: 1 / 1.5
(Override as needed per workflow.)

Rayflux Horndog

A model slightly overtuned on anatomy; it captures the photographic flair of the Rayflux line but tends to produce accidental NSFW content due to the way it latched onto certain patterns during training. High detail, vivid rendering, and strong contrast.

Rayflux Photoplus

A best-in-class photorealistic Rayflux variant with excellent analog-style grain, stable anatomy, and a broad creative range. Ideal for portraits, street scenes, and detailed natural photography with a cinematic twist.

Rayflux v1.0

Less photorealistic than Photoplus but more vibrant and creatively expressive. A versatile all-rounder that works well across photography, digital illustration, and hybrid “photo-illustration” styles.

Rayflux v1.0 (UNet-only)

Same aesthetic profile as Rayflux v1.0, packaged as a UNet-only checkpoint variant for compatibility and experimental pipelines.

Rayflux v3.0

Evolution of the v1.0 line, trained on the same dataset as Photoplus. Strikes a balanced hybrid identity: clean photographic detail with meaningful creative flexibility and a slightly more mature, cohesive visual language.

Rayflux v3.x (v3.1 → v3.4)

A consolidated series of refinement experiments designed to progressively “de-hornify” Horndog — cleaner anatomy, fewer unintended biases, and improved stability — while preserving the core Rayflux style.

Rayflux v4.0

A better-tuned evolution of Horndog, keeping its high detail and expressiveness while delivering much more stable, predictable, and controllable outputs.

CHROMA Family — Raydiant

Raydiant v1.0

Raydiant v1.0 is a Chroma-based model spun off from Chroma v48, tuned to produce uncensored analog-style photography with strong contrast, light fuzziness, and a natural film-grain feel. Despite its photographic backbone, it also performs surprisingly well across bold illustrated genres, maintaining expressive color and stylistic weight.

Intended use:
Generalist model, fully open source and uncensored, leaning toward analog photography and gritty realism.

Recommended settings:

Sampler: Euler Beta
Steps: 16
CFG: 4.5
Notes: Pad the T5 tokenizer (min_length = 3); works well with dedicated negative prompting.

Caveats:
Fully uncensored and tends toward colorful, vivid output by default. It's hard to control and quite prone to tantrums: bleeding concepts, garbled latents and unpolished hands/feet.

FLUX_KREA Family — Rayflux Krea

Rayflux Krea v1.0

Rayflux Krea v1.0 is a clean photorealistic checkpoint tuned for believable, natural skin textures and a polished photographic look. It carries a distinct artistic flair and handles a wide variety of styles well, from creative photography to expressive illustration. Compared to standard Rayflux models, it has its own flavor — stylistically stronger than vanilla Flux, but also slightly more resistant to absorbing new concepts.

Intended use:
Creative, photoreal, and stylistically rich imagery; strong at portraiture and mixed artistic styles.

Recommended settings:

Sampler: Euler / Euler Simple
Steps: 12–30+ (best results above ~30)
CFG: 1
Shift: low (0–1)

Performance notes:
Performs best at 30+ steps. At lower steps, an upscaler is recommended to smooth out residual noise.

QWEN_IMAGE Family — RayQwest Models

RayQwest v1.0

RayQwest v1.0 is designed to remove the characteristic “AI plasticness” often seen in Qwen-Image outputs, restoring a more natural analog grit and the recognizable Ray-style aesthetic. It maintains excellent prompt comprehension, handles a wide range of styles with ease, and runs fast — all while giving Qwen’s already powerful architecture a stronger artistic backbone.

Intended use:
A generalist, expressive model that produces natural-looking images with improved texture, reduced smoothness, and strong prompt adhesion. Works well for portraits, lifestyle photography, stylized art, and creative compositions.

Recommended settings:

Sampler: Euler / Euler Simple
Steps: 16
CFG: 1
Notes: Works very well with LoRAs — but reduce LoRA strength compared to Flux/SDXL models.

Caveats:
Carries some of Qwen’s usual softness — less sharp than Rayflux — but compensates with excellent prompt accuracy and expressive range.

RayQwest v0.7

RayQwest v0.7 is a strong pre-v1 iteration characterized by a heavier analog feel, noticeable grain, and a slightly flatter dynamic range. It’s a bit noisier than v1.0, but compensates with excellent aesthetic vibes, maintaining solid prompt comprehension and creative flexibility. Despite being an intermediate version, it remains highly usable and flavorful.

Intended use:
Analog-style, gritty photo generation with expressive textures; suitable for artistic photography, street scenes, and moody portraiture.

Recommended settings:

Sampler: Euler / Euler Simple
Steps: 16
CFG: 1
Notes: LoRAs work well (lower their strength).

Caveats:
More noise and less range than v1.0, but strong vibes and impact.

RayQwest v0.65

RayQwest v0.65 was trained on a different dataset entirely, giving it a more dreamlike, saturated, and softer look. It carries a unique atmosphere — less sharp than v0.7 and v1.0, but with an appealing ethereal quality that makes it stand apart. A charming early version that still holds its own artistically.

Intended use:
Dreamy, soft, colorful renders; stylized photography; expressive portraits; atmospheric scenes.

Recommended settings:

Sampler: Euler / Euler Simple
Steps: 16
CFG: 1
Notes: LoRAs work well (lower their strength).

Caveats:
Softest of the RayQwest line; not as sharp or grounded as the later versions.

SDXL Family

Raymnants 3.0

Realistic model with a good sense for Art. Great at portraits, wide ethnic and body type range. The model that started it all. What would happen if I trained a bunch of street photos into an image model? The answer was: Raymnants. The rudimentary methodology and complete lack of insight make it nothing short of a miracle that it came out that good. Still took half a dozen iterations before 3.0 though!

Intended use:
Portrait photography, character work, realistic scenes with artistic flair.

Recommended settings:

Sampler: DPM++ 2M Karras
Steps: ~30
CFG: 5–8

Caveats:
Early version; less polished than some later variants, occasionally raw or inconsistent, but still remarkably expressive.

Raymnants 3.0 LightningW 4L1.0

A Lightning variant of Raymnants 3.0 distilled with a specific method to accelerate sampling while preserving as much of the original style as possible.

Recommended settings:

Sampler: Euler / sgm_uniform
Steps: 8
CFG: 1

Notes:
Closer to the 3.0 look but thinner in detail and nuance.

Raymnants 3.0 LightningW 4L2.0

Another Lightning variant distilled with a different method, aiming for even faster sampling and a punchier look.

Recommended settings:

Sampler: DPM++ SDE / sgm_uniform
Steps: 6
CFG: 2

Notes:
More aggressive distillation; punchier but also more chaotic and further from base 3.0.

Raymnants 3.4 → 6.2 — Iterative Evolution Line

Raymnants 3.4 through 6.2 form an exploratory lineage built on the unexpected success of 3.0. At the time, the training process involved a lot of experimentation, resulting in stylistic “yoyo effects” between versions — sometimes more realistic, sometimes more expressive, sometimes more saturated.

Despite this fluctuating path, the 6.x models stand out as flexible, bright, saturated, and artistically dynamic, if slightly less realistic than 3.0.

Intended use:
Portraits, expressive character work, stylized realism, bright and saturated scenes.

Recommended settings (all versions):

Sampler: DPM++ 2M Karras
Steps: ~30
CFG: 5–8

General caveats:
Stylization oscillates across versions. Later models shift away from strict realism toward flexibility, brightness, and vivid color rendering.

Rayburn 1.0

Rayburn 1.0 is a bold, stylized comic-illustration SDXL model with a strong personality. It embraces saturation, thick stylized shapes, graphic lighting, and unapologetically expressive palettes. Ideal for scenes that need attitude, energy, and visual punch rather than realism.

Intended use:
Stylized illustrations, comic-panel aesthetics, graphic novel art, bold character portraits, expressive genre scenes.

Recommended settings:

Sampler: DPM++ 2M Karras
Steps: ~20
CFG: 2–4

Caveats:
Not suitable for realistic rendering — this model leans intentionally and heavily into stylization.

RayburnBlitz 1.0

RayburnBlitz 1.0 is the distilled, lightning-fast variant of Rayburn. It remains stylized and comic-driven but trades some of the original model’s extreme boldness for speed and responsiveness. Excellent for rapid iteration, thumbnails, and concept sketching where turnaround time matters.

Intended use:
Fast stylized renders, quick comic concepts, energetic sketches, storyboard-like shots.

Recommended settings:

Sampler: Gradient_Estimation (beta57)
Steps: 6
CFG: 1.5

Caveats:
The distillation process softens some of Rayburn’s signature exaggeration and intensity.

Rayctifier 1.0

Rayctifier 1.0 blends illustration with classical painting sensibilities, sitting stylistically between Raymnants (portrait realism), Rayburn (comic boldness), and traditional oil-painting aesthetics. It is soft, lush, and unapologetically expressive — with a warm, classical energy and painterly richness.

Intended use:
Classical illustration, oil-painting style portraits, dramatic scenes, lush artistic compositions.

Recommended settings:

Sampler: uni_pc (KL optimal)
Steps: 16
CFG: 4

Caveats:
Shines in painterly or classical styles — modern/clean aesthetics tend to fall flat.

Rayctifier 1.5

Rayctifier 1.5 refines everything introduced in the original model: a more defined painterly style, better flexibility, and stronger artistic structure. This is the preferred Rayctifier flavor — richer, more confident, and more adaptable.

Intended use:
Everything Rayctifier 1.0 excels at, but with improved consistency and detail.

Recommended settings:

Sampler: uni_pc (KL optimal)
Steps: 16
CFG: 4

Caveats:
Still prioritized for classical illustration — modern, flat, or minimalist styles remain outside its strengths.

Z-IMAGE TURBO Family — Rayzist

Rayzist v1.0

Rayzist v1.0 is a fast, generalist Z-Image Turbo model with unusually strong prompt adhesion for its class. Despite its distilled nature, it offers a broad conceptual range, performs well across photography, anime, and comic-style rendering, and remains mostly uncensored. It is also notably good at generating readable text inside the image, which is rare for Turbo-tier models.

Intended use:
Rapid ideation, multi-style image generation, low-VRAM workflows, photography and anime rendering, fast comic concepts, and text-in-image generation.

Recommended settings:

Sampler: Euler Simple
Steps: 8
CFG: 1

Caveats:
Still a distilled Turbo model — some visual concepts simply aren’t well encoded (e.g., cyclops / one-eyed characters). These gaps are unpredictable and usually only discovered through experimentation.

License

All models in this zoo are trained on personal and/or mixed-source data and inherit the license and usage constraints of their respective base models (see metadata above). Check the upstream repositories for detailed license terms.

Credits

All models, datasets, training, and curation by Ray / MutantSparrow.

Downloads last month: -

Model tree for MutantSparrow/Ray

Base model

Qwen/Qwen-Image

Finetuned

(56)

this model