Motion-Plate Experiment — Staff Review

Outcome

Ten silent image-to-video clips completed across Pruna P-Video, Seedance 2.0 Mini, Seedance 2.0, Hailuo 02 Fast, Kling 2.1 Standard and Veo 3 Fast. Three Seedance attempts to animate the Sam source were rejected by the same automated sensitivity check, including a neutral five-second retry using the documented [Image1] reference workflow.

The Replicate account reports $6.45 / approximately £4.96 of usage for the experiment. Its $0 outstanding balance means prepaid account credit covered that usage; it does not mean the runs were free. A deliberately conservative preflight estimate that included failed requests was $8.48, but the account's actual usage figure is authoritative.

Initial engineering assessment

• Generated video should provide people, rooms, hands, light, reflections and restrained camera motion—not final typography, reports, charts or interfaces.
• No tested model reliably preserved the report's fine print. The best production method is a generated environmental plate with the exact A4 report and word extraction composited deterministically in HTML/video rendering.
• Flagship Seedance created the largest reinterpretation of the evidence frame: polished as a new film shot, but unsuitable when exact page geometry matters.
• Seedance Mini retained more of the source composition than flagship Seedance, but still rewrote small text. It remains worth testing on non-text visuals.
• Pruna produced 1080-class output cheaply and preserved the source composition comparatively well. It is a strong value candidate for restrained inserts.
• Hailuo is useful for fast, inexpensive motion studies and is now the default volume-plate tier. Its 512p output is accepted only when the stable selected section survives the 1080p composition; faces and hero close-ups move to Kling.
• Kling and Veo both produced usable character candidates. The viewer grading should decide whether their motion quality justifies their price over Pruna.
• Every downloaded clip is silent as requested; existing narration, music and effects remain the editorial source of truth.

Viewer verdict

The viewer identified Hailuo 02 Fast as seriously cheap and very good, and preferred the Kling 2.1 character result f9gayha5khrmt0czzg08y1b89c overall. This supports a two-tier production workflow:

• Hailuo 02 Fast — default motion plates and inexpensive iteration at $0.10 for a six-second 512p output.
• Kling 2.1 Standard — selected human close-ups and story-critical hero shots at $0.25 for five seconds at 720p.
• Code-native compositor — reports, readable documents, charts, interfaces, dialogue and exact animated reveals.
• Existing audio edit — narration, score and effects; generated clips remain silent.

At three Hailuo plates and one Kling hero plate per episode, seven episodes cost about $3.85 / £2.96 per generation pass before retakes. A two-pass allowance is approximately $7.70 / £5.92. This is the recommended series budget model.

Prompt conclusion

For this series, a prompt should behave like a shot contract:

1. state what must remain unchanged;
2. allow one small subject action;
3. allow one camera move;
4. name only subtle environmental motion;
5. prohibit speech, identity drift, cuts and generated text;
6. request a one-second steady hold for the edit.

The full reusable formula and Seedance-specific input guidance are in PROMPT-GUIDE.md.

Current production status

The viewer grading and production-specific tests are complete. The current rule is Hailuo for volume coverage, Kling for selected human hero shots, and the deterministic compositor for every readable interface or document. Each new paid take still requires a continuity-approved room/character/screen packet, a cost gate, a stable-section review and a recorded prediction ID before integration.

The grading controls remain available as the historical experiment record; they are no longer a release blocker for the already-selected workflow.