A word becomes a field.

The theatre first appears from above. Drag to take the view, or skip the opening. The camera settles at an open overview; enter the theatre or select a word when ready. You then stand at the centre. Choose a word to enter its generation. The same points develop through recorded model states. Each reached state unfolds within a gauze threshold; the camera crosses its plane. Side View pauses the process and lets you orbit its reached stages and same-ID trajectories. View Journey returns to that moment; Continue returns and resumes. Completed forms remain at the outer edge with small labels. Overview and Side View allow free orbiting, panning and zooming. Top, Side and Reset move smoothly to a view; dragging takes over immediately. Press G to open a control panel for passage speed, history lines, the endpoint estimate, mist and exposure. Sound starts with your first click; M or SOUND turns it off.

This work uses route A: a small, class-conditioned joint point generator. It learns the positions and tones of a checkpoint-defined number of points. — RK4 steps preserve — actual states. Inference receives model weights, a class and a fresh Gaussian seed; no target image is supplied.

XY motion and tone come from the learned joint state. The generation view uses a uniform 0.35 display scale; the final exhibit returns to scale 1. Depth spreads actual steps — into separate viewing planes. Linear playback between saved states is not an additional model step. These trajectories are not FLUX pixel motion.

The scrims open and light responds when the camera crosses a reached stage. Camera, lighting, scale and fades are presentation. Side View changes only the viewing angle and visibility of reached records; future states stay hidden. Each line connects one actual point ID through saved states without invented bends. A low-strength local coverage layer connects the recorded samples visually; it adds no new detail. The point model still cannot reproduce all the fine texture of its high-resolution training images.

The faint form on the last plane is the model's own estimate of the final state at the current step. The network predicts this endpoint at every step and builds its velocity from it, so I read it back from each saved state and velocity. It is an estimate, not the result, and it fades before the real final state arrives. Thin lines join a sample of point IDs to their own estimated endpoints and shorten as the points arrive.

The same estimate is also laid on the word's scrim at 3.5 times the size, and fades before the form arrives. A narrow top light over the arrival point brightens as the starting noise leaves the state (1 − β(t) of the model's path); its floor pool shows the estimate's shadow. The side glances turn across the long axis of the estimate read at t = 0.125, and each crossing pulse follows how far the same points actually moved between checkpoints. After a journey the hall keeps one cord per word: 32 lines rise as an ordered ribbon from the central rostrum and end on the exhibit. The rising curve is drawn; the last stretch, from step 116 to the final state, is the record's own same-ID trajectories. The floor shows a weak reflection of each exhibit. Light, projection, shadow, threads and reflection are presentation; no saved state is changed.

The available vocabulary is determined by the trained checkpoint. It is a finite vocabulary, not an arbitrary text model. Detail and generalisation remain experimental. New choices sample fresh noise; saved fallback records are explicitly labelled.

The Creation Well holds the four recorded text-conditioned forms below the eight-word theatre. Each has 129 saved solver states, presented through seven spatial checkpoints. Select a form to enter its passage again, or rise back to the theatre. In the studio version, a submitted short phrase generates a new point record and adds it to this collection.

The text route uses a local FLUX model and eight actual decoded sampling states placed along time-depth. Fine points sample each source image; their lines follow the brightness history at the same image coordinates, not moving pixel identities or the XY trajectories of the eight-word point model. The prompt, seed and original states stay with each result. Only an explicitly submitted new description starts its passage automatically; restored works wait on their own layer.

PROCESS & DOCUMENTATION ↗

Runtime record

Loading

RECORDED WEB EDITION · ABOUT THE WORK ↗

Opening the Creation Well…