Living whitepaper / Updated August 30, 2026
A local visual-generation lab that turns art direction into repeatable still and motion workflows, then keeps model claims behind measured evidence and human approval.
Product truth
ImageForge is the workflow around the model: versioned direction, fixed controls, local execution, evidence capture, review criteria, and an explicit decision about whether an artifact can ship.
Motion evidence
Both clips use the same local model, seed, output size, frame count, frame rate, step count, CFG, sampler, and scheduler. The changed variable is source conditioning.
A five-second generation testing object stability, liquid behavior, camera restraint, and commercial composition without a source image.
Finding: Stable cup, saucer, liquid stream, and composition. Requested steam was weak or absent.
A source-conditioned test of face identity, cafe geometry, espresso-machine structure, CourseForge chrome, navigation placement, and text preservation.
Finding: Identity and major interface geometry remain coherent. Small text softens and some letterforms drift, so deterministic UI animation remains the production path.
AudioForge voice layer
These samples use the local speech pipeline to turn frame text into reviewable narration. Each voice is a controlled avatar for tone testing, not a final learner-facing persona.
Direct, concise, and procedural. Good for safety notes, task setup, and military-adjacent courseware.
More assertive and field-instruction flavored. Useful for checkpoints, warnings, and hands-on procedural demos.
Steady and neutral. Good for policy, accessibility, compliance, and high-clarity enterprise training.
Warmer and more instructional. Better for walkthroughs, introductions, and longer lesson narration.
Available now: American English and British English voices, plus Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese, and Mandarin voice sets. Australian English has local candidate samples under review. African English accents are still research-only.
Comparable CourseForge narration sample using Kokoro voice af_heart.
Comparable CourseForge narration sample using Kokoro voice am_michael.
Comparable CourseForge narration sample using Kokoro voice bf_emma.
Comparable CourseForge narration sample using Kokoro voice bm_daniel.
Ruby is the strongest current Orpheus Australian-English candidate. Human review is still required before it becomes a supported CourseForge narration voice.
Controlled comparison
| Model | Seed | Steps | CFG | Sampler | Scheduler |
|---|---|---|---|---|---|
| Wan 2.2 TI2V 5B | 20260830 | 20 | 5 | uni_pc | simple |
These are local capability probes on one workstation, not general performance guarantees. Video has no generated audio. Human temporal review remains required.
Current status
The ledger records the latest verified state. Promotion requires evidence, not model availability.
Loading current status…
Progress record
Next controlled cycle
The next image-to-video cycle will keep model-generated motion away from UI labels and composite deterministic text back over the clip. Success means the motion can feel alive while the interface remains crisp, readable, and governed by source copy.