Manyform · Independent product preview

Data that changes shape without losing its history.

Curate real observations, transform them in controlled ephemeral workspaces, validate against held-out evidence and release versioned datasets and edge-ready models.

Per-asset lineageProtected holdoutsEdge-aware evaluation
Lineage recordAST-8821 / package deformation
Review passed
Demonstration packaging line with deformed cartons on a conveyor
Derived assetSynthetic · demo
package deformation
SRC-0418Source observation
RCP-009Controlled edit
AST-8821Derived asset
DSV-0.3.2Frozen version
GeneratorQwen edit adapter
Source lockedSHA-256 verified
Split policyTrain eligible
From evidence to releaseOne inspectable chain
01Observe
02Curate
03Create
04Evaluate
05Release
Curated collections

Industrial evidence you can inspect.

Purpose-built collections with explicit real, simulated and derived provenance.

View the catalog
Controlled creation

Augment the gap. Preserve the evidence.

Turn Qwen editing models into repeatable industrial recipes. Lock geometry and identity, control the intended change, then route every candidate through automated checks and expert review.

  • Exact model revision, prompt, seed and source recorded
  • Invariants and forbidden edits defined before a run
  • Synthetic descendants remain distinct and searchable
  • Protected holdout lineage cannot leak into training
Open Transform
Recipe / PKG-DEFORM-V2Ready
Packaging line demonstration source
Editable mask · package only
Controlled instruction

Introduce one visibly deformed carton while preserving conveyor geometry, camera position and surrounding equipment.

Camera Geometry Equipment Package condition
Verification Bench

Performance with operational context.

A model result only matters when you can see the evidence, slices and target it was measured against.

Open Bench
Evaluation / EV-228

Industrial visual event benchmark

Real held-out evidence across camera, lighting, condition and site slices.

Real holdout91.8macro F1
p95 latency142 mson QCS6490
False alarms2.4%real-only
ModelTargetReal scorep95Status
EdgeSight VLM Smallv1.4.1 · Visual event analysisQCS649091.8142 msReleased
Package State Detectorv0.8.3 · Object detectionJetson Orin NX94.227 msEvaluated
Cell Boundary VLMv0.5.0 · VLM evaluationAWS g5.xlarge89.4618 msRelease pending
Trust is a data structure

Every claim leads back to evidence.

Manyform preserves source rights, transformations, human decisions, hashes, dataset membership and evaluation results as first-class records—not prose added at publication time.

Inspect the trust model
Plans and usage

Predictable curation. Transparent generation.

Start at €79 per month with 200 ImageForge generation credits. Every run shows its credit cost before it starts, and overage is consolidated into one monthly statement.

€79Pilot / month200candidates included€0.35additional credit
View plans and credit policy
Lateral Tinkering

Experiments can be playful. Releases still need proof.

Notes, recipes and failure cases from the Manyform lab become reusable methods only after their assumptions and evidence are visible.