360° image toolset

Your local workspace for 360° images

Import, generate, edit, composite, compare, and finish equirectangular panoramas directly on your NVIDIA GPU. Use AI where it helps, and keep image processing on your own machine.

See the 360 workflow View pricing

Walkthrough: first launch to finished render, in the order you work. Nothing here is mocked.

workflow

One workspace, from source to final panorama

Four stages, one application. The work stays in one workspace, and your images don't round-trip to a server.

  1. 01

    Bring it in

    Open a panorama you already have, point at a folder or a saved project, or start a new image from a prompt or Pattern.

  2. 02

    Work on the sphere

    Navigate the live 360 viewer, then pick the tool that fits the region you want to change.

  3. 03

    Compare and commit

    Hold previews as variations, choose the take you want, and commit it as a step in the project graph.

  4. 04

    Finish and export

    Apply finishing passes, resize or upscale where the delivery needs it, and write out a supported format.

foundation

Work directly on the sphere

An equirectangular panorama is a 2:1 rectangle that behaves like a sphere. Treat it as an ordinary flat image and the poles lie to you: the same patch of sky spreads across a huge run of pixels, and a straight stroke lands as a smear.

360 Hextile works in sphere space instead. You navigate the live 360 viewer, and every tool acts on the region you are actually looking at, at its true angular size. The five points below are the same argument the front page makes emotionally. Here they are stated as mechanism.

The rectangle lies. Here is exactly where.

  • At the zenith and nadir an equirectangular projection over-samples: one small patch of the sphere is smeared across a wide band of pixels. That is pole pinch: a property of the projection, not a join. The app names it in the toolbar as De-Pinch, so the copy matches the button you press.

    Hexagonal tiling

  • The left and right edges of the rectangle are one place on the sphere. Asymmetric tiling closes that left-right join by generating across it, and does nothing at the poles, because the pole is not a seam. Already generating in ComfyUI? Keep doing that; asymmetric tiling closes the wrap there too. What it cannot reach is the pole, which is where sphere-space tiling earns its keep.

  • Every tool acts on the region you are looking at, at its real angular size on the sphere, so a brush, a mask, or a Spot region keeps its shape at any pitch instead of being stretched by the projection. The rectangle distorts; the sphere does not.

  • Rendering runs over the faces of a polyhedral mesh, up to a 92-face Goldberg sphere, so the poles are resolved at native resolution rather than as an after-thought stretched over a flat grid. The geometry closes tile boundaries; nothing is patched back in afterwards.

  • Fourteen geometries, from the cube up to the 92-face Goldberg sphere, each one a way to divide the sphere into faces that share context across their borders. You can rely on a truncated icosahedron to get the job done.

    How sphere-space tiling works

  • Ordinary upscalers make pole pinch worse. The projection-aware cubemap mode reprocesses the pole caps specifically as it enlarges, so the operation everyone assumes ruins the poles is the one that treats them properly. Real-ESRGAN x4plus at 2× / 4× / 8× / 16×, up to a 16,384px edge.

    Upsampling

Tiles that share context across their borders

Rendering runs over the faces of a polyhedron rather than a flat grid of strips, and neighbouring tiles are generated with shared context across the boundary between them. That reduces the boundary artifacts flat tiling leaves behind, and it is why the same geometry drives both the viewer and the renderer.

Rendering templates ship in 20-, 32-, and 44-tile layouts, so you can pick how finely the sphere is divided for a given job. When none of them fit, the custom template builder lets you define your own tile layout and render with it.

How sphere-space tiling works

14 Geometries. One Continuous Sphere.

Every shape Hextile can tile, in your hands. Drag to rotate, scroll to zoom, pick a shape, change render mode. This is the same viewer that lives inside the app.

Drag to orbit · Scroll to zoom · Switch shapes

CREATE

01 Already have an equirect? Jump to Edit

CREATE

Start with an equirectangular.

Creation starts with a 360 panoramic image known as equirectangular format: a sphere projected onto a rectangle.

  • Any 360 photo you shot, bring it in. Fix things, swap things out, clean things up, re-imagine it. A ~72 MP camera frame opens and works at full size, on far less VRAM. Less VRAM, never less GPU.

    Insta360Ricoh ThetaGoPro MAX
    See how local processing works

  • Import an equirectangular you rendered elsewhere. Bring an old skybox, or an equirect you are generating in ComfyUI right now. Keep generating there. Drop it here to fix the poles and edit it region by region through a perspective view on a versioned panorama, then restyle, re-detail, or upscale it and take it back out, flat or 360°. A finishing and editing stage, never a replacement pipeline.

    Unreal EngineBlenderCinema 4DUnity
    How importing works

  • Use 360 LoRAs to generate a low-res starter image you can upsample and enhance with tiling.

    SDXLFLUXQwen
    360 LoRA

  • Over 100 GPU procedural generators. Noise, cellular, and geometric tilings you can project onto the sphere.

    PerlinVoronoiReaction-DiffusionPenrose
    Patterns

  • 22 GPU fractal generators, mapped onto the sphere. Zoom, iterations, and palettes are editable.

    MandelbrotJuliaNewtonBurning Ship
    Fractals

  • Hand one Workflow file to a build script or your AI coding agent and get many worlds headlessly, on your own machine. Pick a template, override the creative bits (prompt, seed, count, look), run it. It drops straight into a game engine or a VR world. Export equirectangular PNG, JPG, HDR or EXR, the formats Unreal and Blender already read.

    UnrealBlenderUnityC4D3ds MaxMaya

    A world stops being a thing you make and becomes a build artifact.

    MCP

worlds

Or start from a world someone already wrote.

A World is a whole scene already written down: a global description plus a prompt for every cardinal direction, bound to an interpolation mode that carries each direction into its neighbours. Drill down through the picker, load one, and take it wherever you want.

The Worlds picker: categories, the worlds inside one, and a single world ready to load

Tap a category to drill in, a World to preview, then Load it into your render.

231 worlds, 11 categories

231 built-in Worlds across 11 categories (Alien Planet, Cyberpunk City, Northern Fjord, Crystal Cavern, and more), each a starting point you can take in any direction. Quality presets (sharp, HD, photo) and negative presets come with them.

Directional prompt interpolation

Give each cardinal direction its own prompt and watch them blend across the panorama: volcanic mountains to the north, frozen tundra to the south, in a single image. Five interpolation modes cover 2-Point Gradient, 3-Horizon, 4-Direction, 6-Direction, and Tiles; in Tiles mode you pick a polyhedron, let Qwen2.5-VL describe each face from your image, and get geometry-driven prompts that track the content as you pan.

2-Point Gradient 3-Horizon 4-Direction 6-Direction Tiles
Worlds How sphere-space tiling works

prompt editor

Then write it yourself.

A World hands you a scene already written. The Prompt Editor is where you write your own: the global scene, the negative, and every directional tile prompt in one surface, so you can aim a change at the field that needs it instead of flattening the whole sphere into one sentence.

Describe reads your image, on your GPU

Describe runs Qwen2.5-VL locally and writes the global, negative, or per-tile prompts from an input image. It is keyless: no provider, no account, no key to paste. The Qwen model has to be downloaded first.

Bring your own brain

The text actions are BYOK. Bring a Claude key (Opus, Sonnet, Haiku) or an OpenAI key (GPT-5, the GPT-4.1 family, o3, o4-mini): two providers, nine models. The Intelligent Prompt Engine can expand a simple idea into model-aware phrasing, or read your scene and fill every direction for you, with optional Claude or OpenAI expansion using your own API key.

Your API key is stored encrypted on your machine, sent only to the provider you choose, and never to 360 Hextile. The prompt text goes to that provider too, which is what makes the action work; the one fully-local, keyless writer is Describe.

Prompt editor

pipelines

Pick a brain. They're all local.

Everything runs locally on your GPU. No cloud uploads, no per-render fees. 8 model families cover everything from fast iteration to maximum quality. Generate panoramas from a text prompt or transform your own images, each with its own artistic character. Additional experimental pipelines ship in alpha for advanced users to explore.

Diffusion is one set of tools in the workspace, not the workspace itself. When you reach for it, the render covers the entire sphere rather than a flat frame you then have to wrap, it lands in the same graph as every other edit, and you can condition it with the ControlNet paths each pipeline actually ships models for (depth + canny), or load your own checkpoints and LoRAs alongside the included catalog.

8 model families ship ready for real work: SD 3.5, SDXL, SD 2.1, SD 1.5, FLUX.1 Schnell, FLUX.2 Klein, Qwen-Image, and Z-Image Turbo, plus Real-ESRGAN upscaling.

Eight brains, one workspace.

  • A modern MMDiT model with strong prompt adherence. The default when you want the description followed closely across a whole 360° scene.

    Explore SD 3.5

  • The workhorse, with the deepest community LoRA and checkpoint ecosystem. The safe choice when you want maximum control over style.

    Explore SDXL

  • Lightweight and quick, with a low VRAM footprint. Useful for fast blocking passes before you commit to a heavier model.

    Explore SD 2.1

  • The most-tuned base in existence, with an endless supply of community checkpoints and LoRAs. The model to bring your own weights to.

    Explore SD 1.5

  • A few-step flow transformer for fast drafts. Reach for it when you want a coherent world in a small number of steps.

    Explore FLUX.1

  • A 4B flow transformer from Black Forest Labs under Apache-2.0, around 13 GB VRAM, with a distilled few-step variant and a Base variant tuned for LoRA training. The newest addition to the catalog.

    See the pipelines Explore FLUX.2

  • Describe what you want in plain English. Qwen-Image understands natural-language instructions for both generating new panoramas and editing existing ones. A Generate mode and an Edit mode over the same model.

    See instruction-based editing Explore Qwen-Image

  • Fast iteration for when you are hunting for the shape of a scene and want to see many directions quickly before switching to a heavier model to finish.

    Explore Z-Image

  • TeleStyle is a style-transfer layer, and Chroma1-HD is a diffusion family held back in ALPHA. Both are experimental, in active development, not production-ready, and kept clearly separate from the eight production families above.

Denoise

Dial in the effect

From scratch gives full creative range; from your own image, dial denoise from a light restyle to a complete reimagining.

Light effectStrong effect
See the generation pipelines Read the full model and ControlNet matrix
02

edit

Navigate the sphere. Change what you are looking at.

TOOLS

The tools you reach for while editing.

Every one of them acts on the same sphere: one room, many tools. Every one of them knows it is on a sphere.

  • Resample the panorama to a target dimension when a delivery needs a specific size rather than more detail: a plain, exact resize that keeps the 2:1 ratio intact.

  • Three sphere-space warps behind one toolbar button. De-Pinch works the pinched, over-sampled areas at the zenith and nadir. Offset rotates and recentres the equirect so the seam falls where you want it. Möbius Warp applies a controlled Riemann-sphere transform that keeps the panorama continuous. All three act on the sphere, not on a flat rectangle.

  • Continue a render with another pass, or sweep back over the sphere to revisit it. Chain multiple diffusion passes in a single render, each with its own processor, template, strength, and prompt. The output of one feeds directly into the next through GPU memory. Resume from any completed pass on interruption, and save a multipass preset to repeat the workflow.

  • Put a region anywhere on the sphere and regenerate what is inside it through a perspective view, then composite it back into a versioned 360° panorama. The region keeps its true angular shape at any pitch instead of being stretched by the projection, and native Stable Diffusion inpaint models receive the real mask. Regional editing on a sphere, in a graph that remembers every step.

    Spot editor Read the Spot editing guide

  • Build layered 360° composites directly on the sphere. Place transparent artwork with seam-aware controls and drag it on the live sphere, then align or distribute several layers around the full 360°, or snap them to the horizon: moves a flat editor cannot make. Choose from 19 blend modes, add five editable effects (Stroke, Drop Shadow, Outer Glow, Color Overlay, and a multi-stop Gradient Overlay), and finish with GPU AI Cutout and 2× / 4× upscale. Commit the arrangement as a step in the graph, and reopen a saved recipe as an editable fork rather than a flattened dead end.

    Layer editor The graph
03

compare & commit

Hold the takes side by side. Commit the one you want.

Hold the view. Try another take.

Turn on Lock position and a Spot stops moving: yaw, pitch, roll, and field of view stay exactly where you put them. Every further preview aims at the same place and lands in the Variations filmstrip as a draft, so you can look at them side by side before deciding.

take 01
take 01: the marked region sits at the same yaw and pitch as the other takes, with a different result inside it
take 02 selected
take 02: the marked region sits at the same yaw and pitch as the other takes, with a different result inside it
take 03
take 03: the marked region sits at the same yaw and pitch as the other takes, with a different result inside it
Schematic: position and framing stay fixed; only the result changes.
Spot editor

Every committed edit has a place in the story

Tool previews can sit as drafts while you compare them. Commit the result you want, branch from an earlier image when you want to try a different direction, and see which operation produced each version.

Schematic of a derivation graph A draft node connects to a committed node, which continues along the chain to the node marked in the brand colour (the current head) and also runs down to a branch node.
Schematic: drafts are held for comparison; committing advances the project head.
draft
Tool work you are still considering. Drafts are held so you can compare them, and they do not move the project on their own.
commit
The take you selected becomes a project step and the graph head moves to it.
branch
Carry on from an earlier result without erasing the direction you already took.
The graph Derivation graph guide

Undo goes all the way back.

04

finish

Make it deliverable.

Aim a camera inside the finished world and pull the frame the delivery wants. The sphere is the source; the crops are what ships.

Upscale

Upsampling

Real-ESRGAN x4plus enlarges the panorama at 2× / 4× / 8× / 16×, up to a 16,384px edge, and its projection-aware cubemap mode reprocesses the pole caps as it goes. 8192×4096 native, up to 16384×8192+ with Real-ESRGAN upscaling. Never a bare maximum.

Upsampling

Effects and grading

Post Processing

29 GPU post-processing effects in seven groups (Detail, Color, Artistic, Film, Lens, Noise, and Finishing), applied across the whole sphere so the grade holds together at every heading. The colour system carries 59 bundled LUTs, .cube and zip import, LUT generation from one of your own photos, and a live preview against your current tile, plus ACES / AgX / Reinhard / Hable tonemapping with exposure, split toning, HSL, channel mixer, and gradient map.

Post processing

Harvest — pull flat stills

Reframe & Export

Aim a camera anywhere inside the finished world and pull flat, standard-ratio stills at production resolution (16:9, 2.39:1, square, or a custom aspect), written as PNG, JPG, HDR or EXR. Angles are aimed and exported one at a time, and your poses are saved so you can pull a new shot when the project changes. The sphere is your source, not necessarily your deliverable.

AUTOMATION & AGENTS

Two faces over one Workflow file.

Build one world in the studio, or hand the same Workflow to a build script or your AI coding agent and get many worlds headlessly. The GUI, the CLI, and the agent all open the same Workflow. No second config system, no export step.

  • The plugin starts a local MCP server. Claude Code, Codex, or Grok can run Workflows headlessly — agentic coding aimed at 360°. For a large share of users this is the interface they touch the product through, so it is a first-class surface, not a footnote.

    MCP

  • The hextile CLI runs a Workflow headlessly with override and merge. The template owns the pipeline, and the caller fills only the creative bits: prompt, seed, count, look. Named for vibe coders and game pipelines that want worlds produced by a build step rather than by hand.

    MCP

  • The procedural pattern generators are a first-class Workflow input alongside 360-LoRA and your own images, not just an in-app tab. A math-driven starting image the render passes then do the look on top of.

    Pattern

  • Batch, matrix, and scheduled runs are the game-pipeline face: hand the factory one template and a set of overrides and it produces the variants, each as its own Run with a ledger and resume, all on your own machine.

the machine

What it runs on

Check your machine against this before you buy. It requires Windows 10 or Windows 11 (64-bit) with a CUDA-capable NVIDIA GPU. You can get by on less VRAM, but never on less than an NVIDIA GPU.

Local by design

Image processing runs on your own NVIDIA GPU. Your panoramas do not have to be uploaded to a generation service to be worked on, and there is no per-image cloud-generation charge from 360 Hextile.

What needs the network

Renders run offline once the models are cached. The first download of model weights, software updates, licence activation, and this website are separate network paths, described plainly rather than folded into a blanket privacy claim. The software is a one-time licence, not a subscription.

See how local processing works

Will your card run it?

Pick your GPU and see, pipeline by pipeline, what it can run and at which quantization. This picker is the one place a VRAM number lives: a readout for your own card, not a single floor printed at everyone.

System Requirements

Windows 10 (64-bit, 21H2+) or Windows 11 · CUDA-capable NVIDIA GPU · CUDA 12.1+

No AMD, Intel Arc, Apple Silicon, or integrated graphics.

VRAM depends on the pipeline you run. SDXL (the default) needs 12 GB. Lighter pipelines can run with less. Use the GPU picker above to check your card. · 16 GB recommended · 24 GB optimal

RAM: 16 GB minimum → 32 GB recommended · Storage: 50 GB free disk recommended

Check before you buy. We meant it.

own it

One licence. Buy it once.

One-time purchase. The full 360 image workspace ships with every licence.

Standard

$360 one-time

  • The whole 360 image workspace: every tool on this page
  • 2 machine activations
  • Perpetual commercial licence
  • 30-day money-back guarantee
View pricing

No subscriptions. Never.

Bring your next 360° image into the workspace

One licence, your own GPU, and the whole equirectangular workflow in one place.

View licence options See finished panoramas

all feature pages

Every feature page, by stage

The same order as the page above: the sphere, then the four stages, then the agents and the machine that run them.

the sphere

01 create

02 edit

03 compare & commit

04 finish

agents

machine

technical reference

The full technical inventory

Every model, pipeline, compatibility note, ControlNet path, built-in and bring-your-own LoRA detail, and post-processing effect is listed in full below.

Supported Models & Pipelines

Diffusion pipelines, utility tools, and style transfer — each optimized for different creative workflows

Stable Diffusion XL

1024px native • 12GB VRAM

The most versatile pipeline. Rich ecosystem of LoRAs and IP Adapter style transfer. No ControlNet in this product.

LoRAs No ControlNet IP Adapter

Stable Diffusion 3.5

1024px native • 16GB VRAM

Latest MMDiT architecture with superior prompt adherence and IP Adapter support. Medium, Large, and Large Turbo variants.

LoRAs No ControlNet No IP Adapter

Stable Diffusion 1.5

512px native • 6GB VRAM

The largest community ecosystem. Massive LoRA library and low VRAM requirements. No ControlNet in this product.

LoRAs No ControlNet IP Adapter

Stable Diffusion 2.x

768px native • 8GB VRAM

Solid mid-range option with good depth-of-field rendering characteristics. No ControlNet in this product.

LoRAs No ControlNet No IP Adapter

Qwen-Image

Instruction-based • 13-48GB VRAM

20B MMDiT with natural language editing and generation. GGUF Q4 quantization runs on 16GB GPUs. Two modes: Generate and Edit. ControlNet = depth + canny via Qwen-Image ControlNet Union only.

Generate Mode Edit Mode ControlNet Union (depth + canny) GGUF Q4

TeleStyle ALPHA

Style transfer • 12GB VRAM

Diffusion-based artistic style transfer using a reference style.

Style Transfer Structure-Guided No ControlNet

Post Process ALPHA

Effects only • 0GB VRAM

Apply post-processing effects without AI inference. Sharpen, blur, color adjust, and more — runs on the CPU.

No GPU Required No inference

Format Converter ALPHA

Projection conversion • ~1GB VRAM

GPU-accelerated conversion between 5 projection formats. Experimental — projection conversion is in active development; not production-tested for all format combinations.

Equirectangular Fulldome (Domemaster) Fulldome Angular Cubemap Cross Rectilinear

Real-ESRGAN

Any resolution • 4GB VRAM

AI upscaler for boosting output resolution up to 16K+. Supports x2, x4, and anime-optimized models.

All licenses

FLUX.2 Klein

4B flow transformer • ~13GB VRAM (8GB GGUF Q4)

Compact production pipeline from Black Forest Labs. Qwen3 text encoder (512 tokens); Klein (8 steps) and Klein Base (20 steps, best for LoRA). Apache 2.0 licensed.

Production LoRAs Apache 2.0 No ControlNet

FLUX.1 Schnell

4-step generation • 12GB VRAM

Few-step flow matching from Black Forest Labs. Output in just 4 inference steps. Apache 2.0 licensed.

Fast LoRAs Apache 2.0 No ControlNet

Chroma1-HD ALPHA

High-res diffusion • 19GB VRAM

FLUX-derived high-resolution diffusion tuned for vivid color and crisp detail — a middle ground between SDXL coherence and FLUX prompt adherence.

High-res Vivid color LoRAs

Z-Image Turbo

Few-step turbo • 8-16GB VRAM

Turbo-distilled pipeline for fast iteration. Trade a touch of fidelity for dramatically lower step counts — ideal for rapid concept passes and prompt previews.

Fast Few-step Low VRAM

ALPHA ALPHA engines are experimental, in active development, and not production-ready, and kept clearly separate.

Feature SDXL SD 3.5 FLUX.1 SD 1.5 SD 2.x Qwen TeleStyle Real-ESRGAN
LoRAs 9 built-in 3 built-in 3 built-in
Custom LoRAs
IP Adapter / Style Transfer ViT-H ViT-H Diffusion
ControlNet depth + canny
Native Resolution 1024px 1024px 1024px 512px 768px Varies 1024px Any
Min VRAM 12 GB 16 GB 12 GB 6 GB 8 GB 13 GB 12 GB 4 GB

ControlNet Types by Pipeline

ControlNet is available only on Qwen-Image via ControlNet Union — depth and canny only. SD 1.5, SD 2.x, and SDXL have no ControlNet in this product.

ControlNet Type SDXL SD 1.5 SD 2.x Qwen-Image
Canny Edge detection Union
Depth Depth map conditioning Union
SoftEdge Soft edge detection
OpenPose Human pose detection
Scribble Sketch input
Lineart Line extraction
MLSD Straight line detection
Segmentation Semantic regions
Normal Map Surface normals

Built-in LoRA Styles

SDXL LoRAs (9 included)

  • Watercolor Style
  • Pixel Art XL
  • Papercut Style
  • Crayon Style
  • Embroidery Style
  • 3D Render Style
  • Manga Line Art
  • Voxel 3D Style
  • LCM Accelerator

SD 3.5 LoRAs (3 included)

  • Turbo Realism 2.0
  • Photorealistic
  • Anime Style

+ Install your own custom LoRAs

SD 1.5 LoRAs (3 included)

  • Pixel Art
  • Studio Ghibli
  • LCM Accelerator

+ Thousands of community LoRAs compatible

Bring Your Own Models & LoRAs

The built-in pipelines and styles are just the start. Extend any supported family with your own LoRAs and base models — all local, all yours, no cloud lock-in.

Bring Your Own LoRAs

Drop any compatible .safetensors LoRA into your pipeline's folder and load it from the LoRA tab — including the thousands of community LoRAs from sites like Civitai. Stack multiple and tune each one's strength.

  • Per-family folders: SD 1.5, SD 2.1, SD 3.5, SDXL
  • Stack and blend multiple LoRAs
  • Adjustable per-LoRA strength

Bring Your Own Models

Add your own base models by HuggingFace URL or import a local diffusers-format model from disk. 360 Hextile auto-detects the architecture and matches it to a supported family — you're never locked to the bundled checkpoints.

  • Add via HuggingFace repository URL
  • Import local diffusers-format models
  • Architecture auto-detected per family

29 Post-Processing Effects

GPU-accelerated effects available on every pipeline — insertable at 4 stages: pre-process, post-tile, post-stitch, final

Color
  • ACES Tonemap
  • Color Grade (LUT)
  • Channel Mixer
  • HSL Adjustment
  • Split Toning
  • Gradient Map
Detail
  • Detail Enhance
  • Unsharp Mask
  • CLAHE
  • Edge-Preserving
  • Guided Filter
  • Bilateral Denoise
Film
  • Film Grain
  • Chromatic Grain
  • Halation
  • Sensor Noise
Lens
  • Bloom
  • Chromatic Aberration
Noise
  • Reaction-Diffusion
  • Blue Noise Dither
  • Procedural Noise
  • Noise Toolkit
Artistic
  • Stylization
  • Turing Texture
  • Turing Dither
  • Kuwahara
  • Real-ESRGAN Detail
Finishing
  • Denoise + Regrain
  • Frequency Separation

We use cookies for analytics to understand how visitors use our site. Privacy Policy