*EDIT: I found the bug that casued Bf16 & Int8ConvRot variants to error out - fixed and tested and uploaded - enjoy!
Available Variants
Alongside the original (fp8) checkpoint, I've uploaded two additional variants:
BF16 — a full-precision reconstruction of the model. Loads like any normal checkpoint using the standard "Load Diffusion Model" node (weight_dtype: default). No extra nodes required.
INT8 ConvRot — a quantized variant (roughly half the file size) using INT8 weights with a Hadamard rotation applied to reduce quality loss versus plain INT8.
Overview
Soliloquy is a custom Krea 2 checkpoint created from a LoRA trained entirely on my own photography from my former studio, then permanently integrated into a carefully selected base model (see content notice further below).
It is not a conventional full-parameter finetune, but it is also more than a simple checkpoint merge.
Soliloquy is intended for cinematic portraiture, atmospheric environments, visual storytelling and surreal concepts that still feel as though they were captured through a real camera.
The Dataset
The LoRA used to create Soliloquy was trained exclusively on organic photographic data.
No synthetic or AI-generated images were used in the training dataset. Every image originated from my own photography, and every caption and tag was written by hand rather than generated through automated captioning.
Visual Character
Soliloquy tends toward:
Cinematic and directional lighting
Rich atmospheric depth
Strong subject separation
Natural skin and material texture
Warm practical light against cooler environments
Dramatic environmental compositions
Photographic interpretations of surreal or impossible scenes
A subtle analogue and DSLR-inspired character
The model is not limited to a single genre. It can move between portraiture, fantasy, fashion, landscapes, science fiction and surreal imagery while retaining a fairly consistent photographic eye.
Version 1
This is the first public version of Soliloquy and should be considered an evolving release.
The current model can produce particularly strong contrast in some generations. While this contributes to its dramatic character, certain renders may push shadows or highlights further than intended. I am currently refining the balance toward a more dynamic and controlled contrast response without losing the model’s depth and atmosphere.
Feedback and example generations will help guide future versions.
Content Notice
Soliloquy retains the broader generation capabilities of its underlying base model, including NSFW capability.
As a result, occasional unprompted nudity may appear, particularly in prompts involving bodies, minimal clothing, bathing, fantasy attire or ambiguous wardrobe descriptions. Users who want strictly safe-for-work output should describe clothing clearly and use appropriate negative prompting or workflow-level safeguards.
Recommended Use
Soliloquy responds well to descriptive natural-language prompts, especially when the prompt includes:
The intended light source
Time of day
Weather or atmospheric conditions
Camera position and framing
Materials and surface texture
Emotional tone
Foreground and background relationships
You generally do not need to overload prompts with long quality-tag strings. A clear scene, a strong visual intention and a few carefully chosen photographic details tend to work best.
Closing Note
The name Soliloquy refers to the act of speaking one’s thoughts aloud.
This model is, in a sense, a continuation of that idea: old photographs, visual instincts and memories from a former studio translated into a new generative medium.
It is not an attempt to reproduce one fixed style.
It is an attempt to preserve a way of seeing.
Description
Soliloquy is a custom Krea 2 checkpoint created from a LoRA trained entirely on my own photography from my former studio, then permanently integrated into a carefully selected base model (see content notice further below).
It is not a conventional full-parameter finetune, but it is also more than a simple checkpoint merge.
Soliloquy is intended for cinematic portraiture, atmospheric environments, visual storytelling and surreal concepts that still feel as though they were captured through a real camera.
FAQ
Comments (14)
Will you upload a int8 convrot version?
Yes, it should be up later today or by tomorrow
Hello @Fatbuns , do you mind explaining what it is? Thank you 🙌
@GlowingGuardianGirl It's a model type, much like there's fp8, fp16, fp32, bf15, gguf etc.
"INT8 ConvRot" is the same type - just 1.5x to 2.5x faster than the usual speed and the quality is pretty much the same as an in-between of fp8 and fp16, with the caveat, that it only supports RTX 20th/30th gen to 50th gen.
In short, it's basically a faster bf16
@Fatbuns Thank you, I needed that
Int8-ConvRot version available alongside Bf16
@Raxephion int8 is available (thank you!), bf16 not yet.
@civit77899 yeah the upload took a little longer than anticipated 🤞
Is it possible to download bf16 of your wonderful model somewhere? Unfortunately I can't use quantized versions on my setup yet.
There should be one up soon
Will you consider publishing the lora you trained? Would make a much more versatile application when used with the Raw version of Krea-2.
I actually am in the process of creating a style LoRA based on this which will be more comprehensive than the one used in the merge :)
this is the error im getting from your int8 version, even though other checkpoint's int8 works fine for me
Hmmm... I will have another look for you as I did check and was performing flawlessly my side, but lemme double check for you and confirm