From watching AI-Generated worlds to exploring them with H3

The next major interface for artificial intelligence may not be a chatbot, an image, or even a video. It may be a world.

I recently generated an early implementation of that idea using MiniMax H3, a specialized equirectangular generation ComfyUI workflow with AI 360°prompting to achieve an interactive viewing concept that allows the viewer to control the viewport through the generated environment on mobile and desktop with continuous looping on Youtube and Facebook.

Watch the demonstration on YouTube.

The result is far from perfect, I generated the clip on my laptop with a consumer Nvidia RTX 3080 Ti 16GB VRAM, so the resolution is very limited and the current generation still shows visible seams on some moments of the video and other inconsistencies but those imperfections may be less important than what the experiment demonstrates. A Minimax H3 video is not necessarily something the viewer has to watch from the camera angle position chosen by the creator, the viewer can choose where to look using an immersive UI, that changes the relationship between a person and generative AI media; panoramic video exposes the full spherical observation domain in a single coordinate frame and is naturally suited to immersive telepresence, robotics, and scene-centric world modeling

For more than a century, cinema has been based on a simple relationship: The filmmaker/cinematographer controls the camera, the audience sees the result. Video games changed that relationship by giving the player control over the camera and increasingly the world itself. Virtual reality pushed it further by allowing people to look around an environment naturally. Generative AI introduces another possibility: The world itself can become generated content. I’m exploring this field on my github about Creative knowledge ecology for visual creation

Generative AI has progressed through several major stages, first came systems that generated images from text, then models learned to generate moving images. Video-generation systems dramatically expanded what could be created from a prompt: characters, environments, camera movements, actions and entire cinematic sequences but conventional generated video still has a fundamental limitation: the camera is predetermined so the model generates the sequence that the viewer sees, a viewer can pause it, replay it, or perhaps interact with an interface around it, but the underlying visual world remains a predetermined sequence; an immersive equirectangular representation changes that relationship, instead of generating only the narrow field of view that a conventional camera would record, the scene can be represented across a panoramic spherical projection (an everyday example is an Insta360 camera or a GoPro360). In an equirectangular image, the horizontal dimension represents a continuous 360-degree environment: the extreme left and right sides correspond to neighboring regions of the same spherical space. This new pipeline creates a different possibility: generate the environment first using multiple options like a single prompt, a first image, reference images or videos and then the user decides where to look afterward.

The 2:1 rule: the shape of the immersive world

A practical requirement of the equirectangular representation is its 2:1 aspect ratio. For a full spherical panorama: WH=2\frac{W}{H}=2 where WW is the panorama width and HH is its height. For example: W=3840,H=1920W=3840,\qquad H=1920 or: W=7680,H=3840.W=7680,\qquad H=3840. This is the format expected by common 360° video workflows and is particularly important when delivering immersive video to platforms such as YouTube and Facebook where the panoramic video must be interpreted as a spherical 360° environment rather than an ordinary flat video. For example, the H3 generation branch in my workflow uses 2112 × 1056 so the immersive representation and final delivery pipeline preserve the equirectangular 360° geometry.

3D Rotation Matrices for 360° Equirectangular Images

To manipulate or view a 360° equirectangular image or video correctly, computers use 3D rotation matrices to transform pixel coordinates between 2D image space and 3D spherical space.

Coordinate Transformation Pipeline

[ 2D Equirectangular Pixel (x, y) ]
                 │
                 ▼
       Convert to Spherical Coordinates
                 │
                 ▼
       [ Latitude & Longitude (θ, φ) ]
                 │
                 ▼
       Convert to 3D Cartesian Vectors
                 │
                 ▼
          [ 3D Point (X, Y, Z) ]
                 │
                 ▼
       ┌─────────────────────────┐
       │ 3D Rotation Matrix      │
       │        (3 × 3)          │
       └─────────────────────────┘
                 │
                 ▼
       [ Rotated 3D Point (X', Y', Z') ]
                 │
                 ▼
           Project back to 2D
                 │
                 ▼
[ New 2D Equirectangular Pixel (x', y') ]

The technical challenge is considerable because a conventional neural video model does not automatically understand that the left edge of an equirectangular image must connect seamlessly to its right edge. Treating those boundaries as ordinary image edges can produce discontinuities, duplicated structures, or visible seams.

The seamless equirectangular implementation used in this workflow addresses this problem at the generation and decoding stages rather than simply attempting to hide a seam afterward. The workflow includes specialized MiniMax H3 processing for equirectangular output, including an EquirectMiniMaxH3VAEDecode stage; this matters because the geometry of the output is different from ordinary flat video. The experiment combines several layers. At the generative level, the workflow can use MiniMax H3 with 360 prompt to video, reference-to-video and First last frame-to-video pathways, in the workflow an initial high-res equirectangular image can be generated also with Krea2 or Flux using Panorama Stickers (a small node set and frontend extension for handling 360 equirectangular panoramas developed by Nomadoor)

In general Artificial intelligence imaging use diffusion-based methods that refine scenes holistically so they need to solve a stochastic differential equation; for this workflow I used 20 steps of Euler Ancestral in the first stage and then a more precise second order solution of 4 steps res_2m/beta57 ETA=0 and 0.33-0.45 denoise using the Clownshark sampler of Res4lyf with the 3D latent upscaler that I implemented in my previous H3 Detailer FHD/2.5K Comfyui workflow. A three-step configuration can be used when speed or real-time generation is more important. er_sde/beta57 is a faster alternative, but can reduce detail and affect character acting, audio, and 360 motion consistency. res_2s_rkmk2e two-stage sampler it requires additional model evaluations compared with simpler one-evaluation-per-step approaches. The benefit is the potential for a more accurate numerical approximation of the denoising trajectory.

This is not simply a conventional video placed inside a 360-degree player. The workflow is specifically attempting to make the generated visual representation compatible with a spherical, wraparound viewing model and upscaled and detailed in the latent 3D space of Minimax

To correct the additional contrast and saturation produced by minimax_h3_turbo_v4_step600_ema_pruned it was used Orion4D FXMax DCTL Tone Mapper

The generated sequence can be presented as an immersive environment in which the viewer controls the viewing direction. On a phone, the viewer can interact with the scene; on a desktop, the camera can be moved manually. The sequence can also be looped forward and backward so that the environment continues rather than behaving like a single linear cinematic shot. The result is not yet a fully reconstructed 3D universe like a gaussian splatting. It is a time-varying immersive/equirectangular visual environment that can be explored interactively.

For H3 Minimax 360 prompting I wrote a public custom gpt and added 37000 tokens of H3 360 filmmaking reasoning

It’s possible to get a 360 Equirectacular video prompt using a mixed approach of probabilistic and deterministic prompt enrichment obviously with some key parameters:

1. It explicitly defines the output as a 360° equirectangular spherical panorama. It doesn't leave the model to interpret “immersive” vaguely. It's very important to prompt on H3 that the entire visual field must be represented as a spherical panorama. That establishes the fundamental spatial format.
2. It freezes the physical camera. The prompt must say the camera remains at one fixed physical position and height for the entire sequence, with no pan, tilt, zoom, tracking, dolly, truck, arc, roll, reframing, viewpoint change, or lens change. That is extremely important. In my tests it’s actually possible with Minimax H3 to move the camera but instead spend the temporal generation budget on that, I think in this early stage it is better to spend it generating a stable world around the camera.
3. It explicitly understands the mathematical problem of the panorama seam. I tell H3 that the extreme left and right edges are actually adjacent portions of the same physical environment and create a seam-safe exclusion zone where characters and distinctive objects cannot appear. That's not merely an artistic instruction. It's a representation-aware instruction. H3 is being told: left edge = right edge rather than: left edge | right edge, this distinction is fundamental to making an equirectangular image behave like a wrapped spherical environment.
4. reduce complexity exactly where the representation is most vulnerable, the prompt doesn't simply say “make it seamless.” It specifies that the lateral boundary regions should contain relatively simple environmental elements—terrain, smoke, dust, haze, sky avoiding distinctive objects such as characters,creatures, trees, banners, fires and explosions, a seam through an empty stretch of compatible terrain is much less destructive than a seam passing through a soldier's face or a dragon.
5. explicitly require environmental continuity across the wrap. The prompt calls for matching ground color, terrain geometry, atmospheric density, smoke distribution, illumination, exposure, contrast and sky appearance across the boundary. The prompt reserves approximately the outer 15% of the panorama width on each side as seam-safe environmental regions. The main action remains inside approximately the central 70%. If the panorama width is WW:0≤u<0.15W0 \leq u < 0.15W and 0.85W<u≤W0.85W < u \leq W are treated as seam-safe regions. The main action is approximately: 0.15W≤u≤0.85W.0.15W \leq u \leq 0.85W. I included specific video prompt enrichment version with parameters like for example a Mandatory spatial layer description

When the scene contains meaningful environmental depth, multiple subjects, movement through space, architecture, crowds, vehicles, landscapes, or any camera perspective where spatial relationships materially affect the image, Enriched cideo prompt Version B MUST explicitly describe the scene as spatial layers.

Do not treat "foreground/background relationships" as satisfied merely because the prompt mentions a foreground object and a background environment.

The reasoning engine must construct and explicitly describe:

FOREGROUND elements closest to the camera immediate surfaces, objects, people, vehicles, environmental details, atmospheric effects, or occluding elements their approximate spatial relationship to the camera and primary subject their motion, parallax, scale, and interaction with the camera when relevant

MIDDLEGROUND

• primary subjects and important nearby environmental elements
• subject placement relative to the camera
• secondary subjects, vehicles, props, architecture, or action occurring around the primary subject
• spatial separation between primary and secondary subjects
• relevant occlusion, parallax, and relative movement

BACKGROUND

• distant environmental structure and contextual scale
• architecture, skyline, landscape, distant subjects, distant vehicles, horizon, - - atmospheric depth, or other contextual elements
• their distance relative to the primary subject
• perspective convergence, scale reduction, atmospheric separation, and background motion when relevant

SPATIAL DEPTH / LAYER RELATIONSHIPS

• explicitly describe how foreground, middleground, and background relate to one another
• establish clear near-to-far ordering
• preserve coherent subject-to-environment scale
• describe parallax when the camera or subjects move
• describe occlusion when one layer passes in front of another
• preserve consistent spatial geography throughout continuous movement or across cuts
• ensure the camera perspective visibly communicates the intended depth

NATURAL EDGE TRANSITION:

• do not make the seam-safe zones appear artificially blank
• gradually reduce subject density toward both lateral boundaries
• populate the outer zones with scene-appropriate environmental elements
• use terrain, vegetation, water, sky, architecture, atmosphere, smoke, clouds, reflections, shadows, or equivalent environmental detail according to the scene
• environmental complexity should naturally soften toward the boundaries
• the result must feel like naturally open environment, not intentional empty margins

WRAPAROUND CONTINUITY:

• left and right boundaries represent adjacent portions of the same physical environment
• match lighting, exposure, color, texture, atmosphere, horizon, reflections, terrain, water, vegetation, architecture, and environmental depth across the wraparound
• avoid distinctive high-contrast objects directly on either boundary
• prevent abrupt environmental changes at the seam
• prevent subjects from appearing, disappearing, duplicating, crossing, straddling, or being clipped by the seam

SUBJECT DISTRIBUTION:

• concentrate important subjects and action within the central 70%
• allow natural variation in density inside the central region
• create gradual density falloff toward the seam-safe zones
• use distance, atmosphere, environmental occlusion, and natural spatial structure to make the transition subtle
• do not allow subject movement or crowd flow to migrate toward either boundary

MOTION:

• preserve the requested movement quality and intensity
• fast-moving subjects remain entirely within the central region
• plan movement trajectories with sufficient spatial margin from the boundaries
• do not reduce action intensity merely to protect the seam

This creates an entirely different pipeline: Prompt > AI generation > immersive representation > interactive camera > human exploration The prompt no longer has to describe only what should appear in front of a fixed camera. It can describe a world. That is the conceptual leap, if this generation process becomes sufficiently fast, coherent and inexpensive, the applications could extend far beyond experimental video:

Video games Instead of developers manually constructing every environment, AI could generate explorable spaces from natural-language descriptions. “Generate an alien ecosystem surrounding the player.” The difficult question would no longer be only how to render the world. It would be: How quickly can AI generate and maintain the world as the player explores it?

VR education Imagine asking an AI to create an immersive historical environment and then entering it. Instead of watching a documentary about ancient Rome, a student could potentially enter an AI-generated reconstruction and look around. The teacher could change the scenario through language: “Show the city before the fire.” That would transform AI from an information interface into an environment for learning.

AR world transformation The implications become even more interesting when the same concept is combined with augmented reality. A physical environment could become the canvas. A user might look at an ordinary street with some glasses and ask: “Transform this into a cyberpunk city.” “Show this neighborhood as it looked 500 years ago.” or “show me that car in blue with a representation of me as driver” The underlying physical world would remain present, but the AI-generated visual layer could continuously reinterpret it.

Interactive Cinema Movies could eventually become less linear. Instead of the director deciding exactly what every audience member sees at every moment, a film could provide a controlled environment in which viewers explore the scene themselves. The director would still control the story, performances, lighting, world design and narrative boundaries—but the audience could control the camera. That would not simply be another format for film. It would be a new relationship between cinema and audience.

AI worlds driven by AI agents AI agents could eventually generate the environments that humans and other AI agents interact with in real time.

Obviously the result is not production-ready, there are visible imperfections, the resolution is very limited with consumer hardware, seam quality is not perfect. Generated environments can exhibit inconsistencies, particularly when visual content moves through the panoramic coordinate system and around the poles of an equirectangular representation. but this experiment was produced on my laptop with an RTX 3080 Ti so an implementation using a cluster of high-performance computing server infrastructure could get amazing results immediately. The workflow includes low-VRAM attention, chunked processing and memory-management techniques specifically to make H3 more practical on constrained GPUs of the Open Community.

What happens when the same concept is given to H3 with more efficient inference, higher resolution, stronger temporal consistency and eventually real-time generation? We are accustomed to thinking about AI as something that answers questions, then AI learned to create pictures, then it learned to create moving pictures, the next step may be AI that creates immersive spaces. And once those spaces can respond to human input, the interface changes again. The user is no longer simply asking: “What should the AI show me?” The user can begin asking: “What world should exist around me?” AI-generated visual content does not have to remain a fixed sequence of images. It can become an environment that the viewer interacts with. That distinction may eventually matter as much as the transition from still images to generated video. Because the ultimate evolution of generative AI may not be from text to image to video, it can be: text > image > video > environment > world. And when AI can generate those worlds fast enough for us to explore them in real time, the computer screen may stop being a window onto generated content. It may become the doorway to a new world.

My H3 360 Test Prompt:

integrated_multimodal_description: [Shot 1] Live-action, photorealistic medieval science-fantasy battle presented exclusively as a true 360-degree equirectangular spherical panorama. The camera is absolutely stationary at one fixed physical position and fixed height for the complete 6-second duration. The spherical camera orientation never changes. There is no pan, tilt, zoom, push, pull, tracking, dolly, truck, arc, roll, handheld shake, reframing, camera repositioning, viewpoint change, lens change, or cut. The panorama is one uninterrupted spherical environment. All apparent motion is generated by the battle and environmental elements, never by the camera.

The horizontal equirectangular wraparound seam is a critical visual constraint. Treat the extreme left and extreme right edges as two adjacent portions of the same physical battlefield. Create a very wide, completely unpopulated seam-safe exclusion zone at BOTH lateral extremes. These zones must remain empty of every character for the entire shot, from the first frame through the final frame. No human soldier, blue alien, face, head, torso, arm, hand, leg, foot, sword, shield, spear, helmet, armor fragment attached to a character, character-carried banner, horse, or other living or character-associated element may enter, touch, cross, straddle, or be partially clipped by either boundary. Do not allow characters to run through the edges. Do not allow distant silhouettes to appear at the edges. Do not allow crowd density to gradually creep toward the edges. Keep all people and combat decisively confined to the broad central panoramic region.

The seam-safe lateral zones contain only open muddy battlefield, continuous low terrain, distant non-character environmental forms, smoke, dust, atmospheric haze, subtle ground debris, and sky. Keep the visual complexity lower in these zones. Avoid isolated rocks, poles, buildings, trees, weapons, banners, fires, explosions, dragons, or other distinctive objects touching the lateral boundaries. The environmental texture at the far-left edge must be compatible with and naturally continue into the far-right edge. Match ground color, terrain geometry, atmospheric density, smoke distribution, illumination, exposure, contrast, and sky appearance across the wraparound boundary. The panorama must be able to loop horizontally without a visible discontinuity caused by people or distinctive objects.

The central populated region contains an enormous medieval battle. A young adult human female warrior in battered red-painted plate armor over chainmail stands within the active central region, gripping a broad medieval longsword. Her armor has scratches, dents, chipped paint, rivets, leather straps, mud, and worn steel. Long brown hair streams from beneath her weathered helmet. Around her, the human army wears coordinated red armor and red medieval uniforms. Opposing them is a large army of blue-skinned humanoid aliens with clearly non-human facial proportions, varied physiques, worn medieval armor, blue cloth elements, and steel swords. The faction separation remains visually obvious: red humans against blue aliens.

FOREGROUND: Within the central panoramic region, muddy ground, deep boot impressions, broken shields, discarded weapon shafts, small debris, dust, and rapidly moving fighters establish immediate depth. Near-camera combatants move aggressively and quickly, but none are permitted to approach the lateral seam-safe zones.

MIDDLEGROUND: The primary battle occurs here, with dense red human infantry and blue alien infantry colliding at high speed. Fighters sprint, lunge, duck, parry, slash, block, recover, and counterattack in rapid succession. Sword swings are fast and decisive with believable acceleration and follow-through. Shields absorb impacts and rotate under force. Bodies react immediately to contact. Footwork is rapid and grounded, with boots digging into and kicking up muddy earth. Avoid slow-motion behavior, floating movement, delayed combat reactions, or synchronized choreography.

BACKGROUND: Behind the central battle, enormous formations continue fighting through smoke and dust. In the center-background only, several large fires burn among damaged battlefield structures and debris. Short, violent explosions erupt intermittently behind the fighting, producing brief expanding fireballs, dark smoke bursts, flying sparks, and illuminated dust. Keep all explosions well inside the central panoramic region and far from both lateral seam-safe boundaries.

High above the central battlefield, several large medieval-fantasy dragons fly through the smoky sky. They remain entirely inside the central and upper-central panoramic region, never approaching either lateral boundary. Their wings beat strongly and naturally, with visible wing flex and body momentum. Some dragons circle above the battle while others cross behind smoke columns, creating large-scale aerial motion without intersecting the empty seam-safe zones. Their silhouettes remain clearly separated from the battlefield below.

SPATIAL DEPTH / LAYER RELATIONSHIPS: Maintain a strong near-to-far hierarchy around the fixed spherical camera. Foreground fighters and debris create the strongest local motion; the central melee provides the primary action; distant formations, fires, and explosions move more slowly; dragons occupy the high background sky. Character density decreases deliberately toward both lateral extremes, transitioning into empty environmental space before either image boundary. This gradual reduction in character density is intentional and must remain stable for the full six seconds.

From 00:00 to 00:01.500, both armies accelerate toward one another rather than walking. Red human soldiers sprint forward with swords and shields while blue alien soldiers rush toward them with weapons already raised. The central warrior immediately shifts into an active combat stance and prepares for rapid engagement. Smoke and dust move strongly through the battlefield while the lateral seam-safe regions remain completely empty of people.

From 00:01.500 to 00:03.500, the armies collide at high speed in the central region. Multiple sword fights begin almost simultaneously but remain individually varied. Fighters execute rapid slash-parry-counterattack sequences with short recovery times. Sword impacts produce brief sparks, shields buckle under force, armored bodies recoil and immediately regain balance, and boots kick mud into the air. The action should feel physically fast and violent rather than theatrical or slow.

From 00:03.500 to 00:05.000, several explosions erupt in the central background behind the combatants. Each explosion has a visible physical progression: ignition, rapid expansion of flame and debris, upward-rising smoke, and fading embers. Firelight briefly illuminates nearby smoke, armor, and airborne dust. At the same time, dragons sweep through the upper-central sky with powerful wingbeats, remaining far from both lateral boundaries.

From 00:05.000 to 00:06.000, the central melee reaches maximum speed and density. The red human formation and blue alien formation continue rapid close-quarters combat while another controlled explosion flashes in the central background and dragons remain visible overhead. The central red-armored warrior performs a fast defensive block followed immediately by a forceful counter-swing. Her hair, tabard, chainmail, and loose straps react naturally to her acceleration. End with active combat, burning fires, drifting smoke, and dragons in the central sky.

Throughout the entire sequence, preserve the fixed 360-degree spherical geometry. The camera never moves. The lateral seam-safe zones never contain people. No character may suddenly appear at an edge, disappear through an edge, cross the panoramic seam, duplicate across the seam, or become partially clipped at the boundaries. All battle action, explosions, fires, and dragons remain safely inside the central panoramic region so the left and right edges remain visually compatible and the horizontal panorama wraps cleanly.

overall_soundscape: The fixed viewpoint is surrounded by an intense battlefield sound field with rapid running footsteps, accelerating armor movement, chainmail rattling, fast sword swings, frequent metallic impacts, shield collisions, heavy breathing, and short exertion cries. Each central explosion is synchronized to a sharp blast followed by a deeper reverberating boom, crackling fire, falling debris, and expanding smoke, while powerful dragon wingbeats produce deep rhythmic air movement overhead. Distant combat remains audible around the central battle without suggesting characters occupy the empty lateral seam-safe zones.

non_diegetic_music: A very fast orchestral battle score drives the sequence with rapid low-string ostinatos, deep war drums, layered percussion, aggressive cellos and basses, horns, and low brass. Percussion accelerates into dense rhythmic patterns during the melee, with powerful brass and drum accents synchronized to the central explosions and the final rapid sword clash.

H3 Immersive video workflow

---

References / Credits

MINIMAX H3 https://huggingface.co/MiniMaxAI/MiniMax-H3

Comfy UI - https://comfy.org/

360-degree AI video nodes by Burgstall Labs ComfyUI-Seamless-Equirectangular provides the seam-inference and geometry-enhancement layer that complements projection, outpainting, and LoRA workflows for immersive video. https://github.com/Burgstall-labs/ComfyUI-Seamless-Equirectangular

Nomadoor Panorama Stickers https://github.com/nomadoor/ComfyUI-Panorama-Stickers

Orion4D FXMax by Philippe Joye https://github.com/orion4d/Orion4D_FXMax

MiniMax H3 Turbo Lora by drbaph https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

ClownSharksampler nodes: https://github.com/ClownsharkBatwing/RES4LYF/

Jukka Seppänen Kijai nodes https://github.com/kijai/ComfyUI-KJNodes

Minimax 3D latent Upscaler https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler

SpheRoPE: Zero-Shot Optimization-Free 360◦ Panorama Generation with Spherical RoPE Or Hirschorn , Aaron Olender , Eli Alshan , Ianir Ideses , Lior Fritz , Sagie Benaim∗ Amazon Prime Video, Tel-Aviv University, Hebrew University of Jerusalem https://arxiv.org/pdf/2606.32033