| --- |
| 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] |
