# Wan2.2 one-render pilot background pattern

Use this when the user explicitly allows exactly one final pilot render and wants a quality-assessment video, not a variant batch or platform upload.

## Pattern

Wan2.2 can be used as a **single text-free background/motion layer** under a deterministic renderer when:

- the approval is for exactly one final pilot render
- no styleframe/keyframe batch is approved
- platform APIs/posting remain blocked
- all captions/labels/arrows/factual text must remain renderer-owned

This is not the ideal multi-scene Wan pipeline, but it is a controlled way to evaluate overall video feel without creating a render series.

## Gate checklist before submitting `/prompt`

- Candidate ID, version, and script hash match the explicit approval.
- Claim gate valid.
- AI disclosure valid.
- Resource gate allows `WAN_FINAL_RENDER` only.
- Publishing tier still blocks.
- Platform API tier still blocks.
- Queue is checked; optionally call `/free` with `unload_models` and `free_memory`.

## API observations

On a remote A5000 ComfyUI host, `/object_info` is useful for confirming available Wan nodes and exact dropdown model names before a long job.

A minimal native Wan2.2 TI2V text-to-video background job can use:

- `WanVideoModelLoader`: `wan2.2_ti2v_5B_fp16.safetensors`
- `LoadWanVideoT5TextEncoder`: an available Wan T5 encoder such as `umt5-xxl-enc-fp8_e4m3fn.safetensors` or the scaled variant if the node supports it
- `WanVideoTextEncode`
- `WanVideoEmptyEmbeds`
- `WanVideoSampler`
- `WanVideoVAELoader`: `wan2.2_vae.safetensors`
- `WanVideoDecode`
- `CreateVideo`
- `SaveVideo`

Conservative 16 GB A5000 test/final-pilot background settings:

```text
416x736
61 frames
12 fps
8–12 steps
CFG ~4.5
scheduler unipc
VAE tiling on
CPU/offload text/model settings if memory is tight
```

## Prompting

Prompt for abstract motion/texture and safe negative space only. Example positive prompt shape:

```text
Vertical 9:16 cinematic abstract motion-graphics background for a short about [mechanism].
[Objects/cards/paths] flow through [abstract system].
Modern dark editorial tech style, subtle parallax, smooth camera push, clean negative space for captions and overlays, no readable text.
```

Negative prompt must include:

```text
readable text, words, subtitles, captions, numbers, logos, watermark, brand names, UI text, fake letters
```

## Composition rule

Final assembly should happen outside Wan:

```text
Wan text-free background → deterministic 1080x1920 renderer overlays → TTS audio → final MP4
```

If the single Wan clip is shorter than the final video, looping/upscaling under deterministic overlays is acceptable for a controlled assessment, but disclose it clearly as a limitation in the review package.

## Review package

Include:

- Wan background path and prompt plan/history
- final MP4 path, SHA256, duration, resolution, fps
- audio path and duration
- caption/overlay manifest
- whether forced alignment was used
- exact note that this was one Wan call/render, not a variant series
- no posting/platform API performed

## Pitfalls

- Do not let Wan generate text, labels, UI, captions, or fake numbers; those often become unreadable and unreviewable.
- Do not hide that a single short Wan clip was looped if the final video is longer.
- Do not spin up extra seeds/variants after an “exactly one final render” approval; ask for a second-render approval first.