AI Generation
Open-weight diffusion models for 360° worlds and environments — 8 families from Stable Diffusion to FLUX.2 Klein, all on your NVIDIA GPU
Diffusion is one set of tools in the workspace, not the workspace itself. Eight model families — SD 3.5, SDXL, SD 2.1, SD 1.5, FLUX.1 Schnell, FLUX.2 Klein, Qwen-Image, and Z-Image Turbo — plug into the same sphere-space tiling and project graph as every other edit. Build a brand-new 360° world or environment from a text prompt, or transform a render you already have — pick the pipeline that fits the job.
The same Workflow runs in the GUI, from a build script, or through your AI coding agent (Claude Code and Codex over MCP) — build one world by hand, or many headlessly on your own machine.
Available Pipelines
SDXL Inpaint
Stable Diffusion XL with native inpainting support. Rich detail and coherent compositions at 1024x1024 native resolution. Excellent for atmospheric environments and layered visual depth.
SD 3.5 Large
Stability AI's newest architecture with improved prompt following and superior text rendering. Available in Medium, Large, and Large Turbo variants. MMDiT transformer architecture.
SD 3.5 Turbo
Distilled version of SD 3.5 optimized for speed. Generates images in just 4 steps. Suited to rapid iteration and fast preview workflows.
SD 1.5 Inpaint
The largest community ecosystem with thousands of compatible LoRAs. Low VRAM requirements make it accessible on most NVIDIA GPUs. No ControlNet in this product.
FLUX.1 Schnell
4-step generation from Black Forest Labs using flow matching. CFG-free distilled architecture means zero negative-prompt overhead. CLIP + T5-XXL dual text encoders with 512-token budget.
FLUX.2 Klein
Compact 4B flow transformer from Black Forest Labs with a Qwen3 text encoder (512-token budget). Two variants: Klein (distilled, 8 steps) and Klein Base (undistilled, 20 steps, best for LoRA). Runs on consumer VRAM; Apache 2.0 licensed.
SD 2.x Inpaint
Mid-range option with 768px native resolution. Good depth-of-field rendering and solid general-purpose quality.
Qwen-Image
Instruction-based editing powered by Qwen2.5-VL. Describe changes in natural language — generate new panoramas or edit existing ones with two distinct modes. ControlNet via Qwen-Image ControlNet Union — depth and canny only (the only ControlNet in this product).
Z-Image Turbo
Few-step turbo-distilled pipeline for fast iteration. Trade a touch of fidelity for dramatically lower step counts — ideal for rapid concept passes and previewing prompts.
Post Process
Effects-only pipeline with zero AI inference. Apply sharpening, color adjustments, blur, and other post-processing effects on the CPU — no GPU required.
Generation Capabilities
Text-to-Image (360° LoRA)
Generate a brand-new equirectangular panorama from a text prompt — no source image needed. Our built-in 360° LoRA on SDXL produces sphere-space results with polyhedral boundary handling, straight from words.
Image-to-Image
Transform existing panoramas guided by your original. Adjust denoise strength to control how much the AI modifies your source image.
Inpainting
Selectively regenerate masked regions while keeping the rest intact. Useful for fixing artifacts or replacing unwanted elements.
ControlNet
Guide generation with depth and canny conditioning via Qwen-Image ControlNet Union — the only ControlNet in this product. SD 1.5, SD 2.x, and SDXL have no ControlNet.
Bring your own: extend any supported family with your own LoRAs (drop in any compatible .safetensors) and your own base models (HuggingFace URL or local diffusers-format import).
How Hextile AI Generation Works
Unlike traditional tiling that leaves hard edges at tile borders, 360 Hextile uses a hexagonal projection system that processes overlapping tiles and blends them with GPU-accelerated alpha gradients. Each tile is processed by your chosen AI model, then merged back onto the sphere.
- Image is projected onto overlapping hexagonal tiles
- Each tile is processed by the selected diffusion model
- Tiles are blended back using feathered alpha gradients
- Boundaries and poles are blended in sphere-space, not left as hard edges
Which Pipeline Is Right for You?
Choose based on your creative goals and hardware
Low VRAM (4-8 GB)
Use SD 1.5 (6 GB) or SD 2.x (8 GB) — smaller models with large LoRA ecosystems. No ControlNet on these pipelines.
General Purpose (12 GB)
Use SDXL — proven quality, rich LoRA ecosystem, and IP Adapter support. No ControlNet on SDXL; for depth + canny use Qwen-Image ControlNet Union.
Best Quality (16+ GB)
Use SD 3.5 Large for vivid detail, or Qwen-Image for instruction-based control.
Fast Iteration (12 GB)
Use SD 3.5 Turbo — 4-step generation for rapid previews and fast iteration.
Compact High Quality
Use FLUX.2 Klein — 4B flow transformer with strong prompt adherence; ~13 GB, or ~8 GB with GGUF Q4.
Utilities (0-1 GB)
Use Post Process for effects-only passes, or Format Converter for projection conversions.
Model Comparison
| Model | Quality | Speed | VRAM | Best For |
|---|---|---|---|---|
|
SDXL Inpaint
|
Excellent | Medium | ~12 GB | Final production, rich environments, detailed scenes |
|
SD 3.5 Large
|
Excellent | Slower | ~16 GB | Complex prompts, text rendering, vivid detail |
|
SD 3.5 Turbo
|
Good | Very Fast | ~12 GB | Rapid iteration, previews, fast workflows |
|
FLUX.1 Schnell
|
Good | Very Fast | ~12 GB | 4-step generation, flow matching |
|
FLUX.2 Klein
|
Excellent | Fast | ~13 GB | Compact 4B flow transformer, strong prompt adherence, Apache 2.0 |
|
SD 1.5 Inpaint
|
Good | Fast | ~6 GB | Low VRAM setups, large LoRA ecosystem |
|
SD 2.x Inpaint
|
Good | Medium | ~8 GB | Mid-range option, depth-of-field rendering |
|
Qwen-Image
|
Excellent | Slower | 13-48 GB | Instruction-based editing, ControlNet Union (depth + canny) |
|
Post Process
|
N/A | No inference | 0 GB | Effects-only passes, no AI inference |
|
Format Converter
|
N/A | No inference | ~1 GB | Equirectangular output (additional projection recipes not claimed until validated) |
Technical Specifications
| Supported Models | SD 3.5, SDXL, SD 2.1, SD 1.5, FLUX.1 Schnell, FLUX.2 Klein, Qwen-Image, Z-Image Turbo, Real-ESRGAN, Post Process, Format Converter |
| Generation Modes | Image-to-Image, Inpainting, ControlNet, Instruction Editing, Style Transfer |
| Tile System | Hexagonal projection with overlapping tiles |
| Native Tile Resolution | 1024x1024 (SDXL, SD3.5, FLUX.1) |
| Output Resolution | Up to 16384x8192+ (8192x4096 default) |
| GPU Requirements | NVIDIA GPU with 4-48 GB VRAM (pipeline dependent), CUDA 12.1+ |
| Scheduler Options | Euler, DPM++, DDIM, LMS, and more |
| Memory Optimization | xformers attention, model offloading |
| Availability | Standard |
Work on your own GPU
One-time licence from $360. No subscription, no per-render cloud charge.
30-day money-back guarantee on every purchase.