# Wan2.2 Native + Image-to-Video on RTX A5000 16 GB — Quality Lessons

Session context: remote ComfyUI over Tailscale on an RTX A5000 Laptop GPU 16 GB, using native ComfyUI Wan2.2 nodes for short vertical video tests.

## Core correction

Do not treat a technically successful Wan run as a usable clip. If the user asks for quality, inspect the output visually and be blunt. In this session, the initial T2V car/house clips technically ran, but the user correctly called them unusable: dark flickering colors, blurred objects, unreadable text/plates, and poor semantic control. Report technical success separately from creative usability.

Suggested wording:

- "The workflow completed successfully, but the clip is not production-usable."
- "This proves the stack runs; it does not prove quality."
- "T2V is still GPU lottery for concrete subject/action control; I2V gives more direction."

## Native Wan2.2 vs Kijai WanVideoWrapper

For Wan2.2 5B quality tests, prefer the official/native ComfyUI workflow over `ComfyUI-WanVideoWrapper`.

Native ComfyUI Wan2.2 5B model set:

```text
models/diffusion_models/wan2.2_ti2v_5B_fp16.safetensors
models/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors
models/vae/wan2.2_vae.safetensors
```

Kijai WanVideoWrapper note:

- `LoadWanVideoT5TextEncoder` rejected `umt5_xxl_fp8_e4m3fn_scaled.safetensors` with `fp8 scaled is not supported`.
- Use the unscaled/legacy T5 only for wrapper workflows, or avoid the wrapper for Wan2.2 quality baselines.

Native node probe that worked:

```text
UNETLoader: wan2.2_ti2v_5B_fp16.safetensors
CLIPLoader: umt5_xxl_fp8_e4m3fn_scaled.safetensors, type=wan
VAELoader: wan2.2_vae.safetensors
Wan22ImageToVideoLatent
KSampler
VAEDecode
CreateVideo
SaveVideo
```

## A5000 16 GB run profile

Known-good native Wan2.2 I2V test profile:

```text
resolution: 416x736
frames/length: 49
fps: 10
duration: 4.9s
steps: 16
cfg: 5.0
sampler: uni_pc
scheduler: simple
```

Observed resource envelope:

```text
VRAM peak: ~14.1 GB
RAM free during job: ~14-16 GB
runtime: roughly 2-3 minutes for 49 frames / 16 steps
```

Use `/free` before experiments when switching models:

```json
{"unload_models": true, "free_memory": true}
```

For repeated runs of the same stack, `unload_models=false` can save reload time.

## RAM/offload and safety

ComfyUI/native Wan can offload model parts between VRAM and system RAM. This can prevent OOM but does not improve image quality. If RAM exhausts and Linux swaps heavily, ComfyUI appears hung.

Recommended safety for a 16 GB VRAM / ~32 GB RAM host:

- start ComfyUI without `--lowvram` for quality baseline;
- use `--reserve-vram 1.0` or `--reserve-vram 2.0`;
- add a 16-32 GB swapfile if disk allows;
- keep initial tests at 416x736 / 25-49 frames before scaling.

## I2V is better than T2V, but keyframe quality dominates

The I2V rabbit test was much better than pure T2V for subject identity, but a bad start pose limited motion. A sitting frontal rabbit mostly morphed; it did not convincingly hop across the field.

For action clips, generate/select keyframes with:

- full body visible;
- side view or three-quarter view;
- action pose matching requested motion, e.g. mid-hop;
- enough empty space in travel direction;
- no embedded text/logos/plates;
- no composited inset boxes or blurred-background rectangles unless intentionally stylistic.

Bad keyframe pattern discovered:

```text
blurred 9:16 background + smaller pasted photo rectangle
```

Wan preserved the rectangle as a visible box in the output. Prefer a true full-frame crop or outpainted full-frame keyframe.

## Prompt pattern for photoreal rabbit I2V

Positive:

```text
Photorealistic vertical nature video of a cute rabbit hopping across a sunny green meadow. The rabbit makes small realistic hops forward through grass, ears bouncing naturally, fur detailed and realistic, gentle camera tracking, bright daylight, shallow depth of field, natural colors, high quality wildlife footage, no text, no logo, no watermark. Keep the rabbit anatomically correct and preserve the same rabbit from the starting image.
```

Negative:

```text
cartoon, animation, deformed rabbit, extra ears, extra legs, melting body, distorted face, blurry, low detail, flickering colors, abstract, static image, text, subtitles, watermark, logo, bad anatomy, duplicated rabbit, disappearing rabbit, unrealistic jump, camera shake
```

## Cleanup after experiments

When disk is tight and switching from wrapper to native Wan2.2, keep only the native stack above. Candidates to remove if no longer using wrapper/Fast Wan2.1 fallback:

```text
models/diffusion_models/TI2V/Wan2_2-TI2V-5B_fp8_e4m3fn_scaled_KJ.safetensors
models/vae/Wan2_2_VAE_bf16.safetensors
models/diffusion_models/Wan2_1-T2V-1_3B_fp8_e4m3fn.safetensors
models/vae/Wan2_1_VAE_bf16.safetensors
models/text_encoders/umt5-xxl-enc-fp8_e4m3fn.safetensors
/tmp/wan22_native
~/.cache/huggingface/hub
ComfyUI/output/* test files
```

Before deleting cache, verify needed model files exist under `ComfyUI/models/...`.