retronic/pixeled-image

HF Inference Providers aren't recommended to run the model, as the settings they enforce aren't very good with Pixeled image generation.

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 extremely detailed image-planning prompts.

Instead of only describing the general idea of an image, Pixelship provides Pixeled with a long natural-language prompt containing several kinds of information.

The prompt first gives a detailed description of the complete image, including its appearance, style, subjects, text, colors, lighting, and other visible details.

It can then provide composition math to help describe proportions, alignment, spacing, centering, and approximate positions. Pixelship uses a conceptual 16×16 grid when useful for describing where elements belong.

The prompt then explains where the important objects should appear. This can include their approximate grid positions, sizes, relationships to other objects, and what each object should look like.

Finally, it describes background elements separately, including where they appear and what they look like. This helps distinguish the main foreground composition from scenery, decorations, environmental elements, textures, and other background details.

All of this information is written as ordinary descriptive text rather than JSON or another structured data format.

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 the intended image before generation.

---

Recommended Settings

The recommended settings are:

LoRA Scale: 0.5
Guidance Scale: 5
Seed: Random
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 detailed prompts.

Higher or lower values can still be experimented with, but 0.5 is currently recommended for the LoRA scale.

---

Technicals

Pixeled is is pretty good model, but this shows the actual backbone toward the model.

Diffusion Transformer

The main generative network in FLUX.2-dev is a Diffusion Transformer (DiT).

During generation, the model begins with a noisy latent representation. The transformer processes this representation together with the conditioning produced from the input prompt.

Across the inference process, the model progressively transforms the latent representation toward the final image.

Pixeled retains this architecture completely. The Pixeled LoRA modifies learned behavior within FLUX.2-dev rather than replacing its underlying generative architecture.

Text Prompt
     ↓
Text Conditioning
     ↓
FLUX.2-dev DiT + Pixeled LoRA
     ↓
Latent Generation
     ↓
Image Decoder
     ↓
Final Image

---

Training Style

Pixeled is trained from image + detailed prompt pairs.

Each image in the training dataset is accompanied by a natural-language description that attempts to reconstruct the image through detailed visual and spatial information.

IMAGE  ←→  DETAILED IMAGE DESCRIPTION
IMAGE  ←→  DETAILED IMAGE DESCRIPTION
IMAGE  ←→  DETAILED IMAGE DESCRIPTION
IMAGE  ←→  DETAILED IMAGE 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, typography, spatial relationships, 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 detailed prompting method being applied across very different visual styles instead of training it around one narrow type of image.

---

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 construct these extremely detailed prompts.

User Prompt
     ↓
Pixelship 2.0
     ↓
Detailed Image Planning
     ↓
Composition + Object Placement
     ↓
Detailed Natural-Language 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.

---

Conclusion

Pixeled is an experiment in giving an existing image model a much more precise description of complicated images.

FLUX.2-dev provides the underlying image-generation capabilities, while the Pixeled fine-tune teaches the model to better interpret the detailed descriptions generated by Pixelship 2.0.

Instead of requiring a rigid structured format, Pixelship communicates the image plan through detailed natural language: describing the complete image, providing composition information, explaining where important objects belong, and separately describing the placement and appearance of background elements.

The recommended configuration is:

FLUX.2-dev
     +
Pixeled LoRA @ 0.5
     +
Pixelship Detailed Prompt
     ↓
Pixelship 2.0

More detail. More control. Better images.

---