# The Architect's Stillness Video Model Sweep

- Published: 2026-05-09
- Run ID: `architects-stillness-20260508T232959Z`
- Remote movie directory: `/srv/neonflux/shared/chat-assets/movie/architects-stillness-20260508T232959Z`
- Models tested: 10
- Completed: 3
- Failed: 7
- Best completed model: CogVideoX 5B

## Model Outcomes

| Model | Status | Latency | Prompt Match | Verdict |
| --- | --- | ---: | ---: | --- |
| Wan 2.1 T2V 1.3B | completed | 343.94s | 0.5/10 | Rendered a valid MP4 but the image is effectively black, so it fails the visual and narrative requirements. |
| CogVideoX 2B | completed | 209.27s | 2.5/10 | Produces a coherent Gothic/industrial surface with cracked-glass energy, but it is almost static and omits the central story objects. |
| CogVideoX 5B | completed | 714.53s | 3.5/10 | Best completed clip: it suggests a cracked dome/cathedral mood, but remains mostly environmental and does not execute the prompt sequence. |
| Mochi 1 Preview | failed | 20.73s | n/a | CUDA out of memory during model load/generation. |
| SkyReels V2 DF 14B 540P | failed | 7200.01s | n/a | Exceeded the 7,200,000 ms HF video runtime timeout at 960x544, 49 frames. |
| LTX Video 13B Distilled | failed | 48.88s | n/a | CUDA out of memory despite supplied source keyframe. |
| HunyuanVideo I2V | failed | 300.56s | n/a | Runtime tensor/indexing failure after source-image handoff; likely adapter/source conditioning incompatibility. |
| SkyReels V1 Hunyuan I2V | failed | 5.75s | n/a | Repository layout mismatch: missing model_index.json for the current diffusers loader path. |
| SkyReels V2 I2V 14B 540P | failed | 607.73s | n/a | Tensor channel mismatch in image-conditioning path; model adapter does not match runtime input shape. |
| Stable Video Diffusion XT | failed | 7.71s | n/a | Runtime API mismatch: guidance_scale was passed to a pipeline that does not accept it. |

## Operational Notes

- Average GPU utilization during sampled run: 85.7%.
- Peak temperature: 83.0C.
- Multi-generator overlap samples: 24.

## Recommendations

- Add low-memory presets for Mochi, LTX, and 14B SkyReels before public default use.
- Exclude unsupported kwargs per pipeline, especially Stable Video Diffusion guidance_scale.
- Split complex cinematic prompts into keyframe-first workflows with explicit source frames for each major beat.
- Fix image-conditioning adapters for Hunyuan/SkyReels I2V before judging creative quality.
- After SSH timeouts, explicitly terminate the remote process group to prevent stale GPU jobs.