Wan I2V Issues

[tiller]

Hello,

I'm trying to make Wan 14B I2V works, but I'm struggling with weird results and I can't seem to find the cause : what am I doing wrong?

I have compiled a series of tests, and none of them gave a good result.

Here is my setup:

• UI Preset: wan

• VAE: wan_2.1_vae.safetensors

• Text encoder: wan_umt5_xxl_fp8_e4m3fn_scaled.safetensors

• Diffusion in low bits: automatic

• img2img

• prompt: "man look at the camera, and smiles"

• 720x480, 1 batch, 33 frames

• Shift: 5

• Denoising Strength: 0.6

• Seed: -1

• Branch new SD Forge Neo installation from 4th of July

• Command line args: --cuda-malloc --listen --uv --enable-insecure-extension-access --disable-sage --reserve-vram 2

• Cuda 13.2

• python: 3.13.5

• torch: 2.11.0+cu130

• gradio: 4.40.0

• Nvidia 2080Ti 11GB VRAM

• i5-13600KF, 32GB Ram

• Debian

• Source image: https://docs.stooit.com/s/9zNHNmSRyiKEzTj

Vanilla experience: no sage attention, or I don't know what else

Here are my tests:

+----------------------------------+--------------------------------------------------+-------------------------------------------------+-----------------------------------------------------------+-------------+-----------------+---------------+----------------+-------------------+-----------+
|                #                 |                    Checkpoint                    |                     Refiner                     |                           Lora                            | ImageStitch | Sampling Method | Schedule Type | Sampling Steps | Switch at (steps) | CFG Scale |
+----------------------------------+--------------------------------------------------+-------------------------------------------------+-----------------------------------------------------------+-------------+-----------------+---------------+----------------+-------------------+-----------+
| 1                                | wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors | wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors |                                                           | No          | Euler           | Simple        |             20 |               0,9 |         1 |
| 2                                | wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors | wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors | wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors | No          | Euler           | Simple        |              4 |              0,75 |         1 |
| 3                                | wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors | wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors |                                                           | Yes         | Euler           | Simple        |             20 |               0,9 |         1 |
| 4                                | wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors | wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors | wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors | Yes         | Euler           | Simple        |              4 |              0,75 |         1 |
| 5                                | wan2.2-I2V-A14B-HighNoise-Q5_K_M.gguf            | wan2.2-I2V-A14B-LowNoise-Q5_K_M.gguf            |                                                           | No          | Euler           | Simple        |             20 |               0,9 |         1 |
| 6                                | wan2.2-I2V-A14B-HighNoise-Q5_K_M.gguf            | wan2.2-I2V-A14B-LowNoise-Q5_K_M.gguf            | wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors | No          | Euler           | Simple        |              5 |               0,6 |         1 |
| 7                                | wan2.2-I2V-A14B-HighNoise-Q5_K_M.gguf            | wan2.2-I2V-A14B-LowNoise-Q5_K_M.gguf            |                                                           | Yes         | Euler           | Simple        |             20 |               0,9 |         1 |
| 8                                | wan2.2-I2V-A14B-HighNoise-Q5_K_M.gguf            | wan2.2-I2V-A14B-LowNoise-Q5_K_M.gguf            | wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors | Yes         | Euler           | Simple        |              5 |               0,6 |         1 |
| 9                                | wan22I2VA14BHIGHQ4KM.Mw0y.gguf                   | wan22I2VA14BLOWQ6K.eRRD.gguf                    |                                                           | No          | Euler           | Simple        |              4 |              0,75 |         1 |
| 10                               | wan22I2VA14BHIGHQ4KM.Mw0y.gguf                   | wan22I2VA14BLOWQ6K.eRRD.gguf                    |                                                           | Yes         | Euler           | Simple        |              4 |              0,75 |         1 |
| 11                               | wan22I2VA14BHIGHQ4KM.Mw0y.gguf                   | wan22I2VA14BLOWQ6K.eRRD.gguf                    |                                                           | Yes         | Euler           | Simple        |              4 |               0,5 |         1 |
| 12                               | wan2.2-I2V-A14B-HighNoise-Q5_K_M.gguf            | wan2.2-I2V-A14B-LowNoise-Q5_K_M.gguf            | wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors | Yes         | Euler           | Simple        |              5 |               0,6 |         4 |
| 13                               | wan2.2-I2V-A14B-LowNoise-Q5_K_M.gguf             |                                                 |                                                           | Yes         | Euler           | Simple        |             20 |                   |         1 |
| 14                               | wan2.2-I2V-A14B-LowNoise-Q5_K_M.gguf             |                                                 | wan2.2_i2v_lightx2v_4steps_lora_v1_low_noise.safetensors  | Yes         | Euler           | Simple        |              4 |                   |         1 |
| 15                               | wan22I2VA14BLOWQ6K.eRRD.gguf                     |                                                 |                                                           | Yes         | Euler           | Simple        |              4 |                   |         1 |

All resulting videos: https://docs.stooit.com/s/f36wrkZ4xdfRCGi

To sum-up the video, when using the raw model, it's mainly:


The best I got was:

and using a random nsfw model from civitai gives better results (but not perfect):

So the main things I don't understand:

• why do I have these results?
• is it normal that without ImageStitch I don't get anything resembling my base image at all? (my understanding was the base image given to img2img was the first frame, or at least a very strong recommandation of what it should look like, and ImageStitch allows to define the last frame). But if it's normal, and I have to use ImageStitch, I don't understand how I specify a first frame?
• why a "random" nsfw model gives better results?

Notes:

• unet comes from Comfy-Org and QuantStack on Huggingface
• I have both the high_noise and low_noise loras for the 4step lightning
• all logs are available : https://docs.stooit.com/s/f36wrkZ4xdfRCGi
• I have tried other Sampling Method / Schedule Type outside of these tests, with no better results

Thank you!

[abzaloff]

• Denoising Strength: 0.6

Denoising Strength: 1