Built entirely on Sogni
How this was made.
Sogni World is three explorable worlds — every image, every crossing, every voice, every piece of music and every 3D figure generated on the Sogni Supernet, and every line of code written by an LLM agent driving those models through the Sogni Creative Agent skill. Nothing here is stock, licensed or hand-modelled. Here is the whole recipe, so you can build your own.
The stack
Eight models, in this order.
Each one takes the exact output of the last. Model ids are what you pass to the Sogni API — they are the same names the agent used to build this.
- 01
Draw the places
krea2_turbo_fp8_scaledText to image. Every scene starts as one still, generated wide and left at native resolution.
In this build: A scene still is the source of truth for everything downstream, so it is generated once and never screenshotted, cropped or re-saved. Each one keeps its own SHA-256, and the app refuses to load a still whose bytes have changed.
- 02
Keep the same character, somewhere new
krea2_identity_edit_sogni_v0_3_alphaImage editing with context references. Give it two pictures — the world, and who has to stay recognisable — and it composes a new place with the same character in it.
In this build: This is how the traveller wears the same moss-green cloak in seven scenes, and how the hooded figure you can collect in The Dreaming Wilds turns out to be the sloth from another world entirely. A name is not an identity lock. The reference image is.
- 03
Make the objects clickable
sam3_image_segment_bf16Segment Anything 3. Ask for an object by name, by point, or by box, and it returns a pixel mask.
In this build: Every glowing outline in this app is a real returned mask, traced to an SVG contour and shipped with the hash of the image it came from. Nothing is hand-drawn, and the same model runs live behind the "Select anything" box while you are standing in the scene.
- 04
Travel between them, and let them speak
minimax-h3-fastvideo-int8_flf2v_turbominimax-h3-fl2va-fp8_flf2v_turboFirst-frame / last-frame video with generated audio in the same pass. Give it the still you are leaving and the still you are arriving at, and it renders the journey between them.
In this build: Anchoring both ends on exact retained stills is what makes a crossing feel continuous rather than like a cut: the clip opens on the frame you were already looking at and lands on the one you are about to explore. The same model, anchored on one still at both ends, is how characters talk — the scene holds, breathes and is spoken over.
- 05
Score it
minimax_music3Instrumental music from a prompt, long enough to loop.
In this build: One cue per region, crossfaded as you move, ducked under any clip that has its own audio.
- 06
Knock out the background
birefnet_image_background_removal_fp16Foreground from background, with no prompt at all. Where segmentation selects an object you name, this one separates whatever is in front from whatever is behind, and it does the edges far better.
In this build: The two are not interchangeable and this build uses both. SAM 3 decides which object and where — "the telescope, not the observatory around it" — and its bounding box crops the scene down to that object. BiRefNet then mattes the crop, where the thing being separated is the only thing left in frame. Worth knowing before you write the code: SAM 3 returns a strictly binary mask and BiRefNet returns a soft matte, so thresholding a matte at pure white will measure a fraction of the real coverage. Threshold at half-opaque instead.
- 07
Stand things up in 3D
pixal3d_int8_i23dImage to 3D. Hand it an object on transparency and it returns a GLB.
In this build: The input costs nothing to generate: the retained still, carrying that matte in its alpha channel, cropped to the object. What you turn over is literally the thing standing in the scene. Ask for far fewer triangles than the default — a browser wants 60,000, not 700,000 — and do not let anything downscale the cut-out on the way in. A third of the pixels lost off a portrait is a third lost off the face, which is the only part anyone looks at.
- 08
Give a recurring character one voice
Qwen3-TTS Voice CloneClone a voice once and speak any line in it, so a character sounds like themselves in every scene.
In this build: The dialogue in these three worlds came out of the video model itself, which generates picture and sound together and casts a fresh voice each time. A voice clone is what you reach for when the same character has to carry a whole world.
The part that is not a model
One loop, repeated a few hundred times.
The agent never calls a model directly from application code. Every asset goes through the same four steps, which is what makes a world of this size survivable.
- 01Prepare
Write the exact job — prompt, seed, references, sizes — to a plan file. Nothing is submitted. A plan that already exists must reproduce byte for byte.
- 02Render
Journal the submission before sending it, so a job whose outcome is unknown is never silently duplicated. A completed item is never re-rendered.
- 03Verify natively
Probe the returned file. Frame counts, dimensions, audio channels, triangle counts, mask coverage, SHA-256. Then look at it.
- 04Retain
Publish to immutable, hash-addressed storage and write the receipt into a manifest. The app only ever reads the manifest, so it cannot ship an asset that was never verified.
What that added up to
The size of it.
Read straight off the shipped manifests.
- 3 worlds, 24 scenesEvery one an original still, never a video frame
- 65 clickable objectsBacked by 80 retained SAM contours
- 72 crossingsPlus a pre-rendered reverse of each, for the way back
- 75 spoken momentsVideo and audio generated together, in one pass
- 10 collectible figuresCut out of the scenes they stand in, rebuilt as meshes
- 8 music cuesOne per region, looping and crossfaded
Sogni Unlimited · One plan, more possibilities
One image starts a world. Keep the story growing.
Try another opening. Give a character a new path. Turn the scene in your head into a place someone can explore. Eligible image and video generations use your Unlimited plan without spending Spark or SOGNI.
Fair use, queues, model and plan limits apply. External vendor models can cost extra. Current prices and any trial eligibility appear at checkout.
Already subscribed? Sign in with the same Sogni account and select your Unlimited plan before generating. Plan-only requests do not fall back to token spending.
Your turn
Every model on this page is available to you.
Image generation, image editing, segmentation, video with sound, music, voice cloning and image-to-3D run on the Sogni Supernet — a community GPU network. Choose Pay as you go or an eligible Sogni Unlimited plan and check coverage before generating. Point your agent of choice at the Sogni Creative Agent skill and it can drive all of them.