Download
3 variants available
fp8 SafeTensor
Ronin-V1-AIO-Fp8.safetensors
8-bit, smaller file • 16.61 GB
Verified: 12 hours ago
SafeTensor
Other Formats
V1


12.3K0 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9.0 1 2 3 4 5 6 7 8 9K
4570 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
7090 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
License:
Ronin

The last horse in the stables — and the most challenging to build.
Like Soliloquy and unlike Solstice and Serendipity, Ronin is not built on the official Krea 2 Turbo base. It is derived from a community merge/finetune — specifically Krea2 Turbo NSFW AIO V1 by bebaszna — whose full model lineage could not be traced or reconstructed. Where Solstice and Serendipity had the clarity of a known foundation, Ronin began with an unknown history and required a fundamentally different, more cautious approach.
The question that started it all: what happens if I take a strong but seemingly abandoned community base with strong character LoRA support and add a dash of Solstice's colour engine, sprinkle in some of Serendipity's micro-detail and prompt adherence, and fold in a measure of Soliloquy's compositional instinct?
Thousands of tests later, Ronin is a reality.
Ronin is an ongoing project and, alongside Solstice, one of my two daily-driver checkpoints — particularly for character LoRA work.
Soon available on Huggingface
Usage
Ronin is an AIO (All-In-One) checkpoint — the VAE and text encoder are baked into the model. Load it using the standard Load Checkpoint node in ComfyUI. No external VAE or text encoder required.
Overview
Ronin is a Krea 2 Turbo checkpoint created through iterative LoRA training and merging into a selected community base model.
It is not a simple checkpoint-to-checkpoint merge, and it is not a clone of its base. Multiple LoRAs were independently trained, each contributing different visual characteristics. These were evaluated, refined through extensive comparative testing, and merged — some iteratively — into the base to produce a model with its own distinct visual identity.
The base model is excellent in its own right, particularly for character LoRA compatibility. Ronin was built to preserve that strength while pushing the model's rendering intelligence significantly further.
Working with an untraceable lineage meant every merge decision carried more risk. Without knowing exactly what was already in the weights, each LoRA integration required careful A/B testing to ensure improvements were genuine and nothing was being inadvertently degraded. This slower, more iterative process ultimately shaped Ronin's development philosophy: test everything, assume nothing, trust only the output.
What Ronin Does Differently From Its Base
Ronin's improvements over the base model are consistent and directional. They are not random variation — they are systematic upgrades that hold across genres, subject matter, and prompt styles.
Material Rendering & Texture Fidelity Ronin produces richer, more physically convincing materials. Wet leather looks like wet leather — with weight, creasing, and sheen that respond to light direction. Silk has visible weave structure and embroidered patterning. Metal has accurate specular response. Chainmail resolves as individual interlocking rings rather than a texture overlay. Where the base model renders good surfaces, Ronin renders surfaces that behave like real materials under real light.
Contextual & Cultural Intelligence This is Ronin's defining trait. It does not simply generate "a temple" or "traditional clothing" — it generates the right temple and the right clothing for the context described. Wuxia scenes receive appropriate period hairstyling. Moroccan interiors get correct architectural layering — zellige tilework below, carved stucco frieze above. Viking equipment reflects actual Norse material culture. Japanese forge scenes include the correct workshop garments and tool types. Ronin understands what belongs in a scene, not just what looks good in one.
Environmental World-Building Ronin constructs more complete environments with greater depth, detail, and physical logic. Backgrounds are not afterthoughts — they are coherent spaces with architectural structure, atmospheric perspective, and contextually appropriate elements. A Gothic cathedral floor shows remnants of its original checkerboard tilework beneath the moss. A Bangkok fight ring has posters on the walls and a cameraman in the crowd. A rooftop has weathered planks and railing wire rather than generic concrete.
Atmospheric & Physical Accuracy Light, smoke, water, and weather behave more realistically. Smoke transitions from warm orange near a heat source to cool blue-grey as it rises and cools. Mud follows gravity down skin in rain rather than sitting as a uniform grime layer. Water sheets across reptile scales rather than beading artificially. Atmospheric haze separates foreground from background with realistic depth falloff. These are subtle improvements individually, but they accumulate into a meaningfully more believable image.
Color Separation & Chromatic Richness Where the base can lean toward monochromatic color washes — particularly in neon-lit or night scenes — Ronin introduces greater chromatic variety. Neon spills in multiple distinct hues. Puddle reflections carry color from surrounding light sources rather than defaulting to uniform tones. The overall palette is richer and more separated without becoming oversaturated.
Compositional Storytelling Ronin makes compositional choices that serve narrative over spectacle. A blacksmith is shown working at the forge with tongs gripping hot metal — not simply standing near fire. Hair is styled functionally for a fighter rather than defaulting to loose and dramatic. A cracked porcelain mask is held with visible emotional weight. The model tends toward scenes that feel inhabited and intentional rather than posed.
Character LoRA Compatibility Ronin inherits and preserves the base model's excellent LoRA compatibility — one of the primary reasons this base was selected. Trained character likenesses integrate cleanly with minimal interference to the trained features. This was a non-negotiable requirement throughout development. Any merge iteration that degraded LoRA performance was discarded.
Visual Character
Ronin tends toward:
Grounded photorealism with strong environmental context
Material authenticity — surfaces that behave like real materials under real light
Culturally and historically informed scene construction
Rich chromatic palette with natural color separation
Atmospheric depth with realistic haze, fog, smoke, and weather behavior
Compositional choices that serve narrative over spectacle
Functional character design — clothing, hair, and equipment that make contextual sense
Strong prompt adherence with intelligent interpretation of cultural and environmental details
Excellent character LoRA integration
Genre versatility — cyberpunk, wuxia, dark fantasy, urban photography, portraiture, wildlife, historical, science fiction
Rendering Character & Trade-Offs
Ronin's iterative merge process introduced a known trade-off worth documenting.
Compared to its siblings Serendipity and Solstice, Ronin exhibits a slight reduction in micro-texture density — most visible on skin pore detail, fine fabric weave, and granular surface noise when viewed at extreme zoom or in direct A/B comparison. This is a characteristic of the iterative merging process: each LoRA integration smooths the highest-frequency detail layer fractionally, and those fractions accumulate.
At normal viewing scale, this is not as apparent, however, it becomes visible primarily during pixel-level comparison testing against its siblings. It does not affect Ronin's material logic — the model still renders the right materials with the right physical behavior — but users who prioritize maximum micro-texture crunch at macro zoom should be aware of it.
This is not a deficiency — it is a rendering personality. Each model in this stable has its own. Serendipity has the most raw photographic texture. Solstice has the most polished cinematic finish. Soliloquy excels at close-range compositions and grit. Ronin has the deepest contextual intelligence. They are different lenses, not competing products.
Where Ronin Fits
Ronin is the fourth and final model in a stable of four, each with a distinct visual identity:
Soliloquy (The Photographer): Intimate photorealism, un-airbrushed skin, tactile surfaces, macro detail, beauty photography, close-range portraiture. The 85mm prime lens of the family.
Serendipity (The Director): Cinematic storytelling, camera placement, spatial depth, dynamic motion, multi-subject scenes, complex stage direction. Built around what is happening inside the frame.
Solstice (The Icon): Cinematic realism, elegance, visual harmony, refined subject presentation. Excels when you want fantasy, cyberpunk, glamour, action, or everyday scenes to feel photographed rather than illustrated.
Ronin (The World-Builder): Contextual intelligence, material authenticity, environmental coherence. Excels when the truth of the world inside the frame matters as much as its beauty. The model that knows what a Moroccan riad wall is actually made of.
These models are not ranked — they are different lenses for different intentions. Ronin does not replace its siblings. It offers something none of them do.
Content Notice
Ronin is fully NSFW capable.
The base model carries a strong NSFW bias, and Ronin inherits this. Efforts were made during development to reduce unsolicited nudity, and significant progress was achieved — but this remains a work in progress. The inherited tendency is persistent, and white or light-colored clothing on female subjects in particular can still trigger unintended results.
For strictly safe-for-work generations, describe clothing clearly and specifically, use appropriate negative prompting, and employ workflow-level safeguards where needed.
Recommended Use
Ronin responds well to detailed, descriptive prompts — particularly those that include environmental, material, and cultural specificity. Effective prompts tend to include:
Specific cultural or historical context — name the tradition, the region, the era
Material descriptions — embroidered silk, battle-scarred chainmail, weathered stone, wet teak, corroded brass
Environmental detail — architectural elements, vegetation type, weather conditions, time of day
Camera and lens language — wide shot, shallow depth of field, cinematic composition, documentary photography
Lighting direction — golden hour breaking through clouds, neon reflecting off wet pavement, forge light, overcast diffusion
Ronin rewards specificity. The more precisely you describe the world, the more precisely Ronin builds it.
Available Variants
FP8 — standard checkpoint.
BF16 — full-precision reconstruction. Loads using the standard "Load Diffusion Model" node (weight_dtype: default).
INT8 ConvRot — quantized variant using INT8 weights with Hadamard rotation applied to reduce quality loss versus plain INT8.
Closing Note
A ronin is a samurai without a master — a warrior who walks their own path.
This model began from someone else's work. The base was strong, the foundation was sound, but the lineage was unknown and the creator seemed to have moved on. Ronin picks up that blade, reforges it, and carries it forward — not knowing the full history of the steel, but trusting the thousands of strikes that shaped it into something new.
It is not the most intimate model in this stable. It is not the most cinematic. It is not the most polished. But when you need the world inside the frame to feel true — the right architecture, the right clothing, the right light on the right material — Ronin is the model that understands what it is building.
Build true.
Licensing & Disclaimer
Original Model Creators: All credit goes to KREA.ai for the original research, architecture, and weights.
Base Model: Krea2 Turbo NSFW AIO V1 by bebaszna. Full credit to the original creator for the foundation this model is built upon.
License: This model is subject to the KREA 2 License Agreement. Please read and comply with the official license terms before using these weights: KREA 2 Licensing Terms.
Official Distribution: This model is only officially available on Hugging Face and CivitAI. Unauthorised redistribution, rehosting, or modification of these model weights is strictly prohibited. If you encounter this model on any other platform, it has been uploaded without permission.

