Technical route · Open, honest, verifiable
Generation gets you to 70%.
Engineering gets you to 100%.
This page explains exactly how Qountless works — no buzzwords, no "magic".
Every capability is labeled for what it is: a real AI model, a classical algorithm, or a deterministic pipeline.
01 · The architecture
One suite, three layers.
The industry is converging on AI generation. Our bet is different: generation is becoming a commodity — the durable value is in what happens after generation.
Generation layer via API · we ride the best models
Text-to-image (Seedream-class), text-to-video (Seedance-class), reference-based regeneration for character consistency. We don't compete with foundation models — we plug in the strongest available and stay model-agnostic. AI here produces candidates and drafts, never final deliverables.
Repair & finishing layer local · deterministic · zero marginal cost
Super-resolution, denoise, deflicker, deband, frame interpolation (RIFE, feature-gated), face-drift detection and geometric alignment, smart crop, precise typography. Executed by FFmpeg/Rust pipelines on your machine — the same input always yields the same output, every step is previewable and undoable. This is the layer that turns 70% drafts into 100% deliverables.
Synapa — the full-chain harness orchestration · gated autonomy
An LLM understands your intent and decomposes it into steps; deterministic executors carry them out; every step is audited (.qsn), reversible, and labeled (C2PA). Autonomy where it's safe (detect, verify, repair), human gates where it matters (plan approval, final delivery). You direct; Synapa executes.
02 · The honest math
Why not just generate?
A fair question — here is our engineering answer, in the open.
character-consistent AND text-exact AND composition-right AND spec-compliant
// Generative models sample from a learned distribution.
// Joint satisfaction ≈ product of per-constraint rates:
P(deliverable) ≈ 0.9 × 0.9 × 0.9 × 0.9 ≈ 66% // optimistic case
// Re-rolling is global resampling — the 90% that was right gets thrown away too.
// Repair is local surgery — only the failing region is touched.
cost(re-roll) grows with 1/p // p collapses as constraints stack
cost(repair) stays local // and runs offline, for free
Generation has no verifier
Models are trained to stay close to a distribution, not to satisfy your spec. They cannot check themselves against your intent — verification is a tool job, not a model job. It stays ours.
Precision is symbolic
Exact text, brand colors, logo geometry, frame rates — these are discrete, checkable properties. Neural sampling approximates; it never guarantees. Our repair layer guarantees, then proves it.
Better models grow this market
We don't bet on models stalling — we bet on progress. Every model generation produces more 70% drafts. The finishing market scales with it.
No film set ships raw footage
A century of filmmaking never skipped post-production. AI made "shooting" nearly free — post-production demand didn't shrink. It exploded.
03 · Locator Grid
Point at the problem. By number.
"Fix this" only works when you and the AI share coordinates. The Locator Grid overlays a numbered chess-style grid on any image or video frame — say the cell, Synapa goes there.
→ "修复 C4 的手" / "fix the hand at C4" — grid + time = a complete space-time address (C4@00:12)
- ①Deterministic, offline, free. Cell → coordinates is pure math. No model call, no latency, works in the browser and on a plane.
- ②Zero ambiguity. "C4" parses identically in every language. Chess notation is a global convention — no translation needed.
- ③Three ways to point. Grid cell, drawn box, or plain language ("the right hand on the left person") resolved by a vision model — all three produce the same region object.
- ④The grid is the address, not the outline. A selected cell seeds a segmentation snap (SAM-class, local ONNX) so repair follows the true object boundary — then tracks it across video frames.
- ⑤Auditable. Every region action is logged in .qsn: who pointed (you or a detector), which cells, which tool, before/after.
Approach informed by public research: Microsoft Research's Set-of-Mark prompting (numbered marks measurably improve visual grounding). Independent engineering — not an affiliation.
04 · Synapa harness
Autonomy where it's safe.
Gates where it matters.
Synapa is built as a full-chain harness: a Think–Act–Observe loop where the loop itself is the easy part — the discipline is in the gating.
Plan task decomposition → GATE ① you approve the plan
Act deterministic executors mutate timeline / canvas / store
Observe region verification against your delivery spec // pass / fail, per constraint
Correct failing regions → routed repair → re-verify (budgeted) // loops, doesn't guess
Deliver GATE ② you accept the result → export with C2PA label
// Every step: audited (.qsn) · reversible (per-step undo) · honestly labeled
| Region issue | Detected by | Routed repair |
|---|---|---|
| Face drift | Local detector, per-frame (deterministic) | Geometric alignment → reference regeneration |
| Composition / black edges | FFmpeg cropdetect | Smart crop |
| Low spec / noise / flicker | Spec check + signal analysis | Super-res · denoise · deflicker · RIFE interpolate |
| AI artifacts, broken text | Vision scan (AIGC defect labels) | Regional repair → regeneration candidates |
05 · Honesty & roadmap
Labeled AI. C2PA on every export.
We publish what is a real model and what is a classical algorithm — in the product and on this page. Every export carries C2PA content credentials, meeting China's AI labeling rules and the EU transparency code.
Locator Grid + detection upgrade
Grid overlay, address grammar, "fix C4" end-to-end, existing detectors promoted to region issues. No new models required.
Vision grounding channel
"Fix the hand on the left person" — vision model proposes a region; you confirm; local regional repair pipeline executes.
Local segmentation snap
Edge-class SAM ONNX via the same gated path as RIFE: download to enable, rectangular fallback honestly labeled.
Defect scan + auto-correct loop
Synapa finds problems you didn't point at; budgeted repair loops; we report first-pass rate and auto-fix rate — as metrics, not slogans.
Red line, always: no "one-click finished product" promises. AI proposes; you dispose. Read the full honesty note →
See the route in the product.
Free forever for creation features. AI billed by credits, transparently.