# Wan2.2 I2V Workflow Reset on RTX A5000 16 GB

Use this when a user rejects blind T2V/keyframe morphing and asks to repair the production method before generating more content visuals.

## Stop conditions

If a content visual proof is rejected because it is abstract, semantically unclear, stylistically inconsistent, or not I2V-ready:

- mark the proof rejected and not eligible for full-final progression;
- do not generate another content proof immediately;
- do not substitute a fallback animatic for a failed Wan/I2V proof;
- run a technical I2V sanity test first;
- only after the sanity test passes generate styleframes for direction review.

Animatics are allowed only as internal composition-review artifacts and must be labelled `not_wan` / `composition-review-only`. They do not satisfy a visual-proof gate.

## A5000 hardware policy

Treat Laptop RTX A5000 16 GB VRAM as capable but constrained:

- no 14B/high-VRAM workflows as default;
- short technical clips before content clips;
- low/moderate resolution for smoke tests;
- use FP8/offload/tiling where possible;
- never run a 38s/full-final render before accepted scene proofs.

## Node audit checklist

Always query `/object_info` and record available models/nodes before a long job. For Wan2.2 TI2V/I2V 5B inspect at least:

- Native-ish ComfyUI nodes: `UNETLoader`, `CLIPLoader`, `VAELoader`, `Wan22ImageToVideoLatent`, `WanImageToVideo`, `WanFirstLastFrameToVideo`, `KSampler`, `VAEDecodeTiled`, `CreateVideo`, `SaveVideo`, `LoadImage`.
- Kijai WanVideoWrapper nodes: `WanVideoModelLoader`, `LoadWanVideoT5TextEncoder`, `WanVideoVAELoader`, `WanVideoImageToVideoEncode`.

Distinguish native and wrapper stacks. Do not mix wrapper model/VAE outputs with native conditioning assumptions unless the workflow is known-good.

## Known-good start-frame I2V sanity stack

A minimal native Wan2.2 start-frame I2V test can be:

```text
LoadImage RGB start frame
UNETLoader: wan2.2_ti2v_5B_fp16.safetensors, weight_dtype=fp8_e4m3fn
CLIPLoader: umt5_xxl_fp8_e4m3fn_scaled.safetensors, type=wan
CLIPTextEncode positive/negative
VAELoader: wan2.2_vae.safetensors
Wan22ImageToVideoLatent: start_image + VAE, small width/height, short length
KSampler: MODEL + conditioning + latent, few steps for smoke test
VAEDecodeTiled
CreateVideo
SaveVideo
```

Example smoke parameters:

```text
resolution: 256x448 or 416x736
length: 25 frames
fps: 10
steps: 4-8 for technical sanity only
workflow_type: i2v
prompt_only_generation: false
has_start_frame: true
has_end_frame: false
```

A valid sanity result must report model names, text encoder, VAE, input frame path(s), output MP4 path, duration, FPS, exit code, and parser status.

## 12-channel vs 3-channel wrapper error

If `WanVideoImageToVideoEncode` fails with:

```text
RuntimeError: expected input to have 12 channels, but got 3 channels
```

Treat it as a wrapper workflow/model-mode mismatch. The practical cause is that an ordinary RGB image tensor (3 channels) was routed into a FLF2V/Fun-style encoder path expecting concatenated multi-channel latent input. Typical contributing mistakes:

- using wrapper `WanVideoImageToVideoEncode` with start+end images while the loaded TI2V model/mode is not compatible with that end-frame path;
- wrong `fun_or_fl2v_model` setting for the model family;
- assuming `wan2.2_ti2v_5B_fp16.safetensors` can be driven through the wrapper FLF2V path without a proven workflow;
- mixing Kijai wrapper nodes and native ComfyUI assumptions.

Repair ladder:

1. Prove native start-frame I2V with `Wan22ImageToVideoLatent` first.
2. Only then test native `WanImageToVideo` if needed.
3. Only after that test `WanFirstLastFrameToVideo` with same-camera coherent start/end frames.
4. Do not proceed to content keyframes if the technical sanity test fails.

## Styleframe direction gate

After technical I2V sanity passes, generate at most 3 styleframes for direction review — no endframes and no video yet. For mechanism-reveal shorts, styleframes should be:

- 9:16;
- same visual concept across variants;
- clean high-end 3D/CGI business-tech motion design;
- text-free information tiles/cards;
- a clear funnel/compressor;
- a visible unconnected decision-target ring;
- no fake UI text, logos, toy figures, photoreal office clutter, dark sci-fi tunnels, or abstract fantasy forms.

Only after the user sends `APPROVED_STYLEFRAME_DIRECTION <candidate_id> <variant_id>` should you create a matching endframe and run one short content I2V proof.