File size: 7,604 Bytes
0dd5017 3eb76a3 d29c244 aa80887 0dd5017 d5f5cef b0e7a93 140d8bd 3eb76a3 0dd5017 3eb76a3 0dd5017 3eb76a3 0dd5017 3eb76a3 0dd5017 25f8a79 0dd5017 3eb76a3 0dd5017 3eb76a3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 | ---
license: mit
base_model:
- Qwen/Qwen-Image-2.1
tags:
- lora
- qwen
- image
- '2.1'
- top
- best
- diffusion
- transformer
pipeline_tag: image-text-to-image
library_name: diffusers
---
# Introducing Pixelship 🚀
We present to you a diffusion transformer called Pixelship.
## Model Details
This is a fine tuned LoRA of Qwen Image 2.1 meant to create images from very detailed prompts.
### Reasoning
This model was once going to be an autoregressive fine tune of Qwen 3.5 4B to produce image tokens. That did not go out well as the model didn't have a concept of image generation really. That made me come up with a new plan.
### Research
Ideogram 4 has proved that training an AI on detailed JSON prompts could boost image quality. The DiT backbone is in charge of the layout of the image. With this info, I was able to come up with this: Fine tune a DiT on style and listening to more of the prompt. But also gives a question: Which model to train?
### Why Qwen Image 2.1
Qwen Image 2.1 is a great model, however, it has many issues that this model can fix and improve. It is pretty good at following a LoRA and other stuff. Training this leaves a better model with little error.
### Prompt Layout
{
"prompt": "Create a highly detailed [ASPECT RATIO] image at [RESOLUTION] depicting [MAIN SUBJECT OR SCENE]. The overall composition is [COMPOSITION DESCRIPTION], with [PRIMARY SUBJECT] positioned at [LOCATION] and [SECONDARY ELEMENTS] arranged around it. The setting is [ENVIRONMENT], featuring [SPECIFIC BACKGROUND DETAILS]. The main subject has [EXACT APPEARANCE, MATERIALS, COLORS, TEXTURES, EXPRESSION, AND POSE]. The lighting is [LIGHTING DESCRIPTION], with [SHADOWS, REFLECTIONS, AND HIGHLIGHTS]. The color palette consists of [EXACT COLORS]. The foreground contains [FOREGROUND DETAILS], the middle ground contains [MIDDLE-GROUND DETAILS], and the background contains [DISTANT ELEMENTS]. Include the exact visible text '[TEXT 1]' at [POSITION], using [FONT STYLE, SIZE, WEIGHT, AND COLOR]. Include '[TEXT 2]' at [POSITION] with [TYPOGRAPHIC DETAILS]. Every visible object, symbol, interface element, and text label must have intentional placement, scale, and appearance. The visual hierarchy should emphasize [FOCAL POINT], followed by [SECONDARY FOCAL POINT]. Use [PHOTOGRAPHIC, GRAPHIC DESIGN, ILLUSTRATION, OR UI STYLE] with [SPECIFIC RENDERING CHARACTERISTICS]. Maintain realistic spatial relationships, consistent perspective, correct proportions, and clear separation between elements. Avoid [UNWANTED DETAILS]. The final image should look like [DESIRED FINAL RESULT], not a generic AI-generated composition.",
"math": {
"coordinate_system": "normalized",
"width": 1.0,
"height": 1.0
},
"objects": [
{
"description": "[PRIMARY SUBJECT: exact identity or type, appearance, colors, materials, pose, expression, orientation, lighting, and visual details]",
"x": 0.25,
"y": 0.2,
"width": 0.5,
"height": 0.55
},
{
"description": "[PRIMARY SUBJECT DETAIL 1: specific facial feature, clothing component, logo, button, or other independently visible element]",
"x": 0.35,
"y": 0.3,
"width": 0.15,
"height": 0.1
},
{
"description": "[PRIMARY SUBJECT DETAIL 2: describe a second individually positioned component]",
"x": 0.5,
"y": 0.3,
"width": 0.15,
"height": 0.1
},
{
"description": "[SECONDARY OBJECT 1: precise shape, appearance, color, material, orientation, and relationship to the primary subject]",
"x": 0.05,
"y": 0.4,
"width": 0.18,
"height": 0.25
},
{
"description": "[SECONDARY OBJECT 2: independently described object with specific visual characteristics]",
"x": 0.77,
"y": 0.4,
"width": 0.18,
"height": 0.25
},
{
"description": "[FOREGROUND OBJECT: exact visual appearance, texture, color, and perspective]",
"x": 0.1,
"y": 0.78,
"width": 0.3,
"height": 0.17
},
{
"description": "[VISIBLE TEXT 1: exact wording, font family or style, font weight, capitalization, text color, alignment, and line wrapping]",
"x": 0.1,
"y": 0.06,
"width": 0.8,
"height": 0.08
},
{
"description": "[VISIBLE TEXT 2: exact wording and detailed typography, positioned separately from all other text]",
"x": 0.2,
"y": 0.15,
"width": 0.6,
"height": 0.04
},
{
"description": "[ICON OR SYMBOL 1: exact shape, stroke width, fill, color, and meaning]",
"x": 0.08,
"y": 0.7,
"width": 0.06,
"height": 0.06
},
{
"description": "[ICON OR SYMBOL 2: separately described visual symbol with exact appearance]",
"x": 0.86,
"y": 0.7,
"width": 0.06,
"height": 0.06
},
{
"description": "[ADDITIONAL DETAIL: a separately positioned decorative or functional component]",
"x": 0.45,
"y": 0.85,
"width": 0.1,
"height": 0.06
}
],
"background": {
"description": "[COMPLETE BACKGROUND DESCRIPTION: environment, base color, architecture, landscape, patterns, textures, lighting, atmosphere, depth, distant objects, shadows, and relationship to foreground elements. Specify whether the background is flat, photographic, illustrated, or three-dimensional.]",
"objects": [
{
"description": "[BASE BACKGROUND: exact color, gradient, texture, or environment]",
"x": 0.0,
"y": 0.0,
"width": 1.0,
"height": 1.0
},
{
"description": "[BACKGROUND ELEMENT 1: exact object, surface, structure, or environmental feature]",
"x": 0.0,
"y": 0.1,
"width": 0.35,
"height": 0.45
},
{
"description": "[BACKGROUND ELEMENT 2: independently described distant visual component]",
"x": 0.65,
"y": 0.15,
"width": 0.3,
"height": 0.4
},
{
"description": "[GROUND, FLOOR, HORIZON, OR LOWER BACKGROUND SURFACE]",
"x": 0.0,
"y": 0.7,
"width": 1.0,
"height": 0.3
}
]
},
"style": {
"style_prompt": "[DESIRED ARTISTIC STYLE]. Render the scene with [SPECIFIC VISUAL QUALITIES], [MATERIAL AND TEXTURE CHARACTERISTICS], [LIGHTING STYLE], [COLOR GRADING], and [LEVEL OF REALISM]. Use [CAMERA OR RENDERING CHARACTERISTICS] and maintain [COMPOSITION RULES]. Typography should be [TYPOGRAPHIC REQUIREMENTS], with exact readable text and consistent spacing. All visual elements should follow the normalized coordinates and maintain correct proportions, perspective, and alignment. Emphasize [MOST IMPORTANT STYLE CHARACTERISTICS]. Avoid [UNWANTED ARTIFACTS, STYLISTIC BIASES, DISTORTIONS, AND GENERIC AI APPEARANCE].",
"style_intensity": {
"realism": 0.8,
"human_made": 0.95,
"smartphone_photo": 0.0,
"casual": 0.1,
"imperfection": 0.05,
"motion_blur": 0.0,
"sensor_noise": 0.0,
"compression": 0.0,
"sharpness": 0.95,
"professional_photography": 0.7,
"cinematic": 0.4,
"polished": 0.95,
"typography_precision": 1.0,
"vector_like": 0.3,
"hand_drawn": 0.0,
"film_grain": 0.02,
"depth_of_field": 0.15
}
}
}
### Conclusion
In conclusion, this is all of how Pixelship 2.0 was reasoned and works.
- **Developed by:** [retronic]
- **Model type:** [Diffusion Transformer, Qwen Image 2.1 Lora]
- **License:** [MIT License]
- **Finetuned from model [Qwen Image 2.1] |