automatic icon indicating copy to clipboard operation
automatic copied to clipboard

[Issue]: WAN 2.2 still images: brown/blue “negative” grids + dtype mismatch on latest dev

Open damianmaz1992-gif opened this issue 1 month ago • 1 comments

Issue Description

What happened

Using WAN 2.2 T2V A14B SDNQ on SD.Next (latest dev branch) I still get the same broken brown/blue "negative" looking grids for still images. Other models (e.g. SDXL) look normal on the same setup.

Vlado on Discord mentioned that this should be fixed in dev, but I'm still hitting the issue + now I also get a dtype mismatch error when running txt2img with WAN.

Environment

  • SD.Next: fresh install, branch dev
  • OS: Windows 10
  • GPU: RTX 3060 12 GB
  • Torch: torch==2.8.0+cu128, torchvision==0.23.0+cu128 (installed from official PyTorch cu128 index)
  • Diffusers WAN model: Disty0/Wan2.2-T2V-A14B-SDNQ-uint4-svd-r32
  • Backend: Diffusers
  • VAE: diffusion_pytorch_model.safetensors (autoloaded)
  • Resolution: 1024x1024, batch size 1×1 (single still image)
  • No extra nodes, no ControlNet, standard sampler settings

Steps to reproduce

  1. Start SD.Next (dev branch).
  2. Load WAN model: Disty0/Wan2.2-T2V-A14B-SDNQ-uint4-svd-r32.
  3. Go to txt2img.
  4. Set size to 1024x1024, batch 1x1.
  5. Use any simple portrait prompt (no LoRAs/styles needed).
  6. Click Generate.

Expected behavior

  • Normal WAN still image output (as in official examples or when using the model in other frontends).

Actual behavior

  • Output is a 4×4 grid of images that look like "negatives":
    • everything is tinted brown/cyan/blue
    • faces barely visible, very noisy
  • When running txt2img I also get a runtime error:

RuntimeError: expected mat1 and mat2 to have the same dtype, but got: struct c10::BFloat16 != float

Full log

(paste full console log from SD.Next here)

[wan2_sdnext_log.txt](https://github.com/user-attachments/files/23696428/wan2_sdnext_log.txt)

### Version Platform Description

_No response_

### Relevant log output

```shell

Backend

Diffusers

Compute

nVidia CUDA

Interface

Standard

Branch

Master

Model

StableDiffusion 1.5

Acknowledgements

  • [x] I have read the above and searched for existing issues
  • [x] I confirm that this is classified correctly and its not an extension issue

damianmaz1992-gif avatar Nov 23 '25 14:11 damianmaz1992-gif

please upload full log

vladmandic avatar Nov 23 '25 14:11 vladmandic

@damianmaz1992-gif i'm still waiting for log?

vladmandic avatar Dec 01 '25 09:12 vladmandic

Hi, here is the full sdnext.log from a fresh run with WAN 2.2 on SD.Next (negative brown/blue image grid issue). I also attached one of the sample output grids from that run.

Let me know if you need any other tests or settings.

sdnext.log

Image

damianmaz1992-gif avatar Dec 03 '25 16:12 damianmaz1992-gif

you said you're using latest DEV branch, but your log shows you are on MASTER

2025-12-03 16:53:41,992 | santa | sd | INFO | installer | Version: app=sd.next latest=2025-11-14T14:52:21Z hash=e38a711c branch=master

this issue should already be resolved in dev branch. please switch to dev and confirm and if issue still exists, upload new log using dev branch.

vladmandic avatar Dec 03 '25 18:12 vladmandic

Switched SD.Next to the dev branch as suggested. WAN 2.2 now works correctly – colors are fine and the negative grid issue is gone. Thanks for the fix!

damianmaz1992-gif avatar Dec 05 '25 18:12 damianmaz1992-gif