Text to Image

The Text to Image project is our ongoing effort to push the limits of AI-powered visual synthesis—from prompt engineering to LoRA training, from aesthetic styling to architectural scene generation.

We don’t just consume generative models—we contribute back, actively releasing custom checkpoints, LoRA adapters, and dataset blends tuned for practical creativity and fidelity.

Focus Areas

Our work centers around three key contributions:

1. Custom Checkpoints

We’ve released and tested several custom checkpoints built on top of Stable Diffusion 1.5, SDXL, and Kandinsky. These are trained with:

Style-specific datasets (e.g., vaporwave, brutalist architecture, pixel noir)
Composition-enhanced pairs (captioned + scene-labeled samples)
Inpainting-focused augmentations for continuity in scene extensions

These checkpoints are optimized for fast inference, minimal artifacts, and consistent subject representation.

2. LoRA Contributions

We maintain a growing library of LoRA adapters, targeting niche styles and characters:

painterly-glow-v1: simulates oil-on-canvas with diffusion lighting
mech-anime-lora: trained on cel-shaded robot scenes and dynamic angles
aesthetic-portraits-v3: face-aware LoRA for consistent human stylization

These adapters are compatible with diffusers, ComfyUI, and AUTOMATIC1111, and available under open licenses via our model registry.

3. Prompt Engineering Toolkits

We actively test and publish prompt templates for:

Style transfer prompts (e.g., "in the style of...")
Scene composition constraints
Negative prompt balancing (e.g., avoiding extra limbs, text artifacts)
Batch generation scripts with automated curation scoring

We've built internal tools for ranking generations using CLIP, aesthetic scoring, and human feedback loops.

Internal Pipelines

Our training and generation stack includes:

TensorOne GPU Clusters for LoRA + DreamBooth training
Model merging workflows with weighted interpolation
Diffusion-based upscalers (ESRGAN, Real-ESRGAN, SwinIR variants)
Prompt-to-seed logging to ensure reproducibility

We can spin up a training job on demand using:

tensoronecli create clusters --gpuType "3090" --imageName "diffusers-train"

Community Impact

Contributed 20+ LoRA models to CivitAI
Shared fine-tuned SDXL models on Hugging Face
Collaborated on style datasets with open creative communities
Hosted internal contests to benchmark prompt → image fidelity

Next Experiments

Text → Sketch → Image cascaded diffusion
Multi-view generation for 3D lifting
Instruction-tuned LoRA for phrase-to-concept consistency
Batch alignment tools to ensure character consistency across outputs