CineForge-Wan-Models / VALIDATION.md
TheBaldDudeCo's picture
Add native Wan runtime support files and validation record
493b7c8 verified
|
Raw
History Blame Contribute Delete
1.45 kB
# Validation record
## Current state
- Four core files uploaded: passed
- Remote byte sizes: passed
- Remote SHA-256 identifiers: passed
- Native load: passed on NVIDIA GeForce RTX 4070 (12 GB)
- End-to-end generation: passed without ComfyUI
- Progress telemetry: passed at both Wan expert steps
- Output validation: passed; five finite 256x256 frames exported to a decodable MP4
- Clean-machine install: not passed
The pack is a release candidate. It must still pass the clean-machine installer test before stable designation.
## 2026-08-14 native smoke test
- Pipeline: Wan 2.2 I2V A14B high-noise + low-noise experts
- Runtime: CineForge native PyTorch/Diffusers loader; ComfyUI absent
- GPU: NVIDIA GeForce RTX 4070, 12 GB VRAM
- Resolution / frames / steps: 256x256 / 5 / 2 diagnostic steps
- Scheduler: UniPC flow-prediction scheduler, flow shift 3.0
- Seed: 42
- Latents: finite after both expert stages
- Frames: finite, range 0.0-1.0, standard deviation 0.081485
- MP4: exported successfully; 19,970 bytes
- Wall time: 231.1 seconds
## Evidence required
- CineForge commit and model repository commit
- Windows, NVIDIA driver, GPU, VRAM, RAM, and disk usage
- input image hash and non-sensitive prompt fixture
- seed, resolution, frame count, frame rate, and inference settings
- wall time, peak VRAM, and peak system RAM
- output video hash and media probe results
- startup and shutdown logs with secrets and private paths removed