Skip to content

fix: scale Qwen Image 2.1 VAE convolutions - #2054

Merged
leejet merged 1 commit into
masterfrom
fix/qwen-image-21-vae-conv-scale
Sep 25, 2026
Merged

leejet merged 1 commit into
masterfrom
fix/qwen-image-21-vae-conv-scale

Conversation

@leejet

@leejet leejet commented Sep 25, 2026

Copy link
Copy Markdown
Owner

Summary

  • Reduce FP16 overflow in Qwen Image 2.1 VAE by scaling convolution inputs by 1/128 and restoring outputs before adding bias.

Thanks to @CarlGao4 for the earlier work #2043 on Qwen Image 2.1 VAE precision handling that helped inform this fix.

Related Issue / Discussion

Fix #2024, #2050.

Additional Information

N/A

Checklist

@leejet
leejet merged commit 0a9340c into master Sep 25, 2026
10 checks passed
@leejet
leejet deleted the fix/qwen-image-21-vae-conv-scale branch September 25, 2026 10:21
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[Bug] Qwen-Image-2.1 transparent (RGBA) output: most of the background decodes opaque white (Metal)

1 participant