⚠️ DEVELOPMENT PREVIEW
Pixeled is currently in active development. The files available in this repository represent an early development version of the model and are primarily intended for testing the Pixeled training approach and Pixelship 2.0 pipeline.
This is not the final or largest Pixeled model. A larger, more extensively trained version of Pixeled is planned and will be uploaded later as development progresses.
Expect the model, dataset, recommended settings, prompt format, and behavior to change between releases.
HF Inference Providers aren't recommended to run the model, as the settings they enforce aren't ideal for Pixeled.
Pixeled
Introducing Pixeled
Pixeled is the image-generation model being developed for Pixelship 2.0. It is a fine-tune of FLUX.2-dev designed to understand Pixelship's highly specific structured JSON prompts.
The version currently available is a development build. It exists to test the training format, structured-prompt understanding, generation quality, and integration with Pixelship 2.0 before the larger model is released.
Instead of only describing the general idea of an image, Pixelship can specify composition, individual objects, approximate positions, colors, backgrounds, typography, and other visual details. Pixeled is trained on this format so these instructions have a stronger influence over the final generation.
The goal is not to turn FLUX.2-dev into a deterministic renderer. Pixeled is still a generative image model. Instead, the goal is to make highly detailed prompting more useful, allowing Pixelship to communicate much more information about what an image should look like.
Development Status
Pixeled is currently in the development and experimentation stage.
The model uploaded here should be considered a preview of the Pixeled approach rather than the final Pixelship 2.0 image model.
This version is being used to evaluate:
- Structured JSON prompt understanding
- Composition and object placement
- Typography and text rendering
- Complex multi-object generations
- Graphic design and website generation
- LoRA strength and inference settings
- Dataset quality
- Integration with the Pixelship 2.0 pipeline
A larger Pixeled model trained with more data and further refinement is planned for a later release.
Because of this, results from the current model should not be treated as representative of the final capabilities of Pixeled or Pixelship 2.0.
Recommended Settings
The recommended settings are:
{
"lora_scale": 0.5,
"guidance_scale": 5,
"seed": "random",
"num_inference_steps": 50
}
FLUX.2-dev + Pixeled LoRA @ 0.5
Pixeled is intended to enhance the prompt understanding of FLUX.2-dev rather than completely overpower the base model.
At these settings, the base model can retain more of its original visual capabilities while Pixeled contributes its learned behavior for interpreting Pixelship's structured prompts.
Higher or lower values can still be experimented with, but 0.5 is currently recommended.
Because Pixeled is still under development, these settings may change in future releases.
Technicals
| Property | Value |
|---|---|
| Name | Pixeled |
| Development status | Experimental / Development Preview |
| Base model | FLUX.2-dev |
| Type | Text-to-image fine-tune |
| Current format | LoRA |
| Recommended LoRA scale | 0.5 |
| Prompt format | Structured JSON |
| Composition system | 16×16 normalized planning grid |
| Color representation | Hexadecimal colors |
| Object representation | Nested structured objects |
| Larger model | Planned for a later release |
Pixeled's training prompts follow a structure similar to:
{
"detailedprompt": "...",
"math": "...",
"palette": [
"#FFFFFF",
"#000000"
],
"objects": [
{
"description": "...",
"x": 8,
"y": 8,
"size_px": {
"width": 4,
"height": 4
},
"color": "#FFFFFF",
"objects": []
}
],
"background": {
"description": "...",
"color": "#FFFFFF",
"objects": []
}
}
The different fields provide the model with several types of information about the intended generation.
detailedprompt contains the main description of the complete image.
math contains additional information about composition and spatial relationships.
palette identifies important colors.
objects describes individual visual elements, including their approximate position, scale, color, description, and child objects.
background separates environmental information from the primary foreground objects.
The x and y values use a conceptual 16×16 planning grid. They are not literal output pixels or deterministic coordinates. An object positioned around (8, 8), for example, is intended to appear approximately around the center of the image.
Pixeled learns these structures as text conditioning.
Pixelship JSON
↓
Text Conditioning
↓
FLUX.2-dev + Pixeled
↓
Generated Image
There is no conventional JSON-to-canvas renderer between the prompt and the image.
Training Style
Pixeled is trained from image + structured prompt pairs.
Each image in the training dataset is accompanied by a detailed description following the Pixeled JSON format.
IMAGE ←→ STRUCTURED JSON DESCRIPTION
IMAGE ←→ STRUCTURED JSON DESCRIPTION
IMAGE ←→ STRUCTURED JSON DESCRIPTION
IMAGE ←→ STRUCTURED JSON DESCRIPTION
│
▼
LoRA Training
│
▼
Pixeled
The descriptions attempt to capture more than the basic subject of an image. They can contain information about layout, visual hierarchy, colors, object placement, backgrounds, text, nested elements, and other details visible in the source image.
The training dataset includes multiple types of imagery, particularly material where detailed prompting is useful:
- Graphic design
- Posters
- Advertisements
- Websites
- UI and screenshots
- Realistic imagery
- Cinematic imagery
- Comics
- Illustrations
- Magazines
- Artwork
- Typography-heavy designs
- Complex multi-object scenes
This gives Pixeled examples of the structured format being applied across very different visual styles instead of training it around one narrow type of image.
The current development release does not represent the full planned training scale. Dataset size, diversity, training duration, and model behavior are expected to expand as development continues.
Current Limitations
Although the development model can already produce strong results, it is not a finished release and currently has several known limitations.
Very Big Paragraphs
Pixeled may produce gibberish or incorrect characters when asked to generate very large paragraphs of text.
Shorter pieces of typography generally provide a more realistic target for the current development model.
Development Model
The currently uploaded weights are an early version of Pixeled.
They are intended to demonstrate and test the underlying approach rather than represent the maximum planned quality of the project.
Prompt Sensitivity
Because Pixeled is learning a highly structured prompt format, changing inference settings or significantly changing the structure of the JSON may affect results.
The format itself may also continue to evolve during development.
Using Pixeled with Pixelship 2.0
Pixeled is intended to sit at the end of the Pixelship 2.0 generation pipeline.
A user does not necessarily have to manually write the JSON.
User Prompt
↓
Pixelship 2.0
↓
Detailed Image Planning
↓
Structured JSON Prompt
↓
FLUX.2-dev + Pixeled @ 0.5
↓
Final Image
Pixelship handles the planning and prompt construction, while Pixeled handles the image generation itself.
The separation allows Pixelship to reason about what should be generated before asking the image model to render it.
As both Pixelship 2.0 and Pixeled are still being developed, this pipeline may change before the final release.
Future Releases
The current repository represents the development stage of Pixeled.
Development is continuing toward a larger model with expanded training and further refinement.
Future versions are expected to focus on:
- A larger training run
- More training data
- Better structured-prompt adherence
- Improved typography
- Better handling of complicated layouts
- More reliable object placement
- Better performance on graphic design and UI
- Improved handling of extremely detailed prompts
- Further Pixelship 2.0 integration
The larger Pixeled model will be uploaded later.
Until then, the currently available model should be treated as an experimental preview for testing, development, and evaluation.
Conclusion
Pixeled is an experiment in giving an existing image model a more precise language for describing complicated images.
FLUX.2-dev provides the underlying image-generation capabilities, while the Pixeled fine-tune teaches the model how to better interpret the structured descriptions generated by Pixelship 2.0.
The currently uploaded version is not the final Pixeled model. It is an active development release intended to test the architecture, training data, prompting system, and Pixelship pipeline before a larger model is released.
Its prompts can be huge, its training images can be huge, and the format will continue to evolve as the model is tested.
The current recommended configuration is:
FLUX.2-dev
+
Pixeled Development LoRA @ 0.5
+
Pixelship Structured JSON
↓
Pixelship 2.0
Current
Pixeled Development Preview
↓
Testing + Dataset Expansion
↓
Further Training
↓
Larger Pixeled Model
↓
Pixelship 2.0
More detail. More control. Better images.
Pixeled is still being built. The bigger model is coming later.
- Downloads last month
- 8
Model tree for retronic/pixeled-beta-image
Base model
black-forest-labs/FLUX.2-dev