Serendipity — The Director’s Checkpoint - V3
Sample Generations
Note: Serendipity V3 is now live! Available in fp8, int8convrot, and bf16 variants to suit your VRAM and workflow needs.
What's New in Serendipity V3:
NSFW Capability: V3 can generate NSFW content natively without requiring external LoRAs for general use. This is not a hardcore model. For more explicit or specialized NSFW content, pairing with a dedicated LoRA is still recommended.
V3 retains the full cinematic staging engine — OTS framing, spatial depth, radiosity, multi-subject blocking, suspended object physics — all intact, all verified through systematic A/B testing against V2 across 13 diagnostic prompts.
Improved Object Intelligence: V3 demonstrates stronger understanding of how objects are actually constructed — armor layering, mechanical internals, structural accuracy — not just how they look on the surface.
Tighter Prompt Adherence: V3 follows specific compositional instructions more precisely, particularly single-hand vs two-hand grips, action staging, and spatial placement of prompted elements.
Stronger Stylized Mode: Cel-shaded anime and Western comic art now produce bolder line weights, better spatial anchoring, and more dynamic action staging than V2.
Whether you are:
- storyboarding a film,
- building comic panels,
- creating atmospheric concept art,
- composing cinematic stills, or
- generating high-energy action sequences,
Serendipity is designed to give you greater control over where the camera is, where the subject is, and what is happening inside the frame.
How Serendipity Was Made
Serendipity did not begin as a synthetic merge recipe.
Its foundation was a LoRA trained on a carefully curated collection of my own older photography from my former studio.
The training dataset was selected manually from my photographic archive and hand-tagged image by image, with an emphasis on the visual principles I wanted the model to learn from real photography:
- Composition
- Subject placement
- Environmental depth
- Camera distance
- Lighting
- Perspective
- Motion
- Visual storytelling
- Photographic atmosphere
No automated tagging pipeline was used for the core dataset.
The LoRA was then trained, evaluated through repeated generations, refined through testing, and ultimately merged directly into the official Krea 2 Turbo base checkpoint to create Serendipity as a permanent standalone model.
This distinction is important.
Serendipity is built directly on the official Krea 2 Turbo base.
The intention was not simply to make Krea 2 look different.
The goal was to inject photographic composition and cinematic staging learned from real-world photography while preserving the capabilities of the original base as much as possible.
Serendipity is therefore best thought of as a:
Photography-derived directional refinement of Krea 2 Turbo, focused on composition, spatial awareness, cinematic staging, and visual storytelling.
Design Philosophy
Serendipity was built around a simple idea:
A strong image model should understand the scene, not just the subject.
The model is intended to respond more deliberately to instructions involving:
- Wide and environmental compositions
- Full-body framing
- Foreground / midground / background separation
- Over-the-shoulder shots
- Camera height and camera distance
- Low-angle and high-angle perspectives
- Multiple subjects occupying distinct spatial positions
- Suspended or moving objects
- Mid-air action
- Dramatic environmental lighting
- Scene blocking
- Narrative interaction between subjects
- Cinematic depth
Rather than forcing every prompt toward a portrait, Serendipity aims to preserve the relationship between subject, camera, and environment.
Best Suited For
Serendipity is particularly well suited to:
- Cinematic photography
- Narrative scenes
- Film stills
- Storyboarding
- Comic and graphic-novel panels
- Environmental portraiture
- Action photography
- Fantasy and science-fiction scenes
- Multi-character compositions
- Atmospheric concept art
- Dynamic camera perspectives
- Dramatic lighting
- Large-scale environments
Prompting Tips for Serendipity
Use Camera Directions: Serendipity responds heavily to cinematic framing tokens like wide shot, extreme wide shot, over-the-shoulder, low-angle tracking shot, and anamorphic lens flare.
Describe Stage Action: Don't just prompt a character—prompt what they are doing with their environment (e.g., leaning casually against the wooden bar, flipping a silver coin, mid-air acrobatic leap).
Layer Your Depth: Mention elements in the foreground, mid-ground, and background to activate Serendipity’s deep bokeh engine.
Licensing & Disclaimer
- Original Model Creators: All credit goes to KREA.ai for the original research, architecture, and weights.
- 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.
Model Lineage
Base: Official Krea 2 Turbo checkpoint Refinement: Custom photography-trained LoRA Dataset: Personally curated real-world photography Captioning: Manual / hand-tagged Final Model: LoRA merged permanently into the Krea 2 Turbo base
No synthetic imagery was used as the foundation of the photographic training dataset.
Variants
Current
- Fp8
- BF16
- Int8ConvRot
Serendipity is an ongoing project, and future versions will continue refining composition, cinematic staging, lighting behavior, spatial awareness, and photographic character.
Serendipity will continue evolving as I test new training data, refinements, quantization approaches, and model-development techniques.
Final Note
Serendipity grew out of photography rather than a desire to simply create another aesthetic checkpoint.
The photographs used to build its original LoRA represent years of real-world experimentation with cameras, subjects, lighting, framing, and composition.
The goal of the project is to carry some of that photographic intuition into a generative model.
Not just:
“What should the subject look like?”
But:
“Where is the camera?” “Where is the subject?” “What is the environment doing?” “Where is the light coming from?” “What is happening in this frame?”
That is the idea behind Serendipity — The Director’s Checkpoint.
Samples
"Sample #1"
"Sample #2"
"Sample #3"
"Sample #4"
"Sample #5"
"Sample #6"
"Sample #7"
"Sample #8"
"Sample #9"