From f4b912749da62f6ed9a9fdfc15dd7bd87a991fd5 Mon Sep 17 00:00:00 2001
From: Frank <frank@dankware>
Date: Mon, 21 Sep 2026 20:19:57 -0700
Subject: [PATCH] comfyui: VAE 4D-to-5D latent expansion for video VAEs

The Qwen-Image 2.1 VAE is a video VAE (5D convs, time_conv, depth-3).
ComfyUI detects it as WanVAE (latent_dim=3), but EmptyLatentImage produces
a 4D latent. The decode function's memory_used_decode lambda accesses
shape[4] which is out of range for 4D.

Fix: expand 4D latents to 5D in the decode function when latent_dim==3.
Added to the patch script for persistence across container recreations.
---
 dankclaw/Containerfile.comfyui | 1 +
 1 file changed, 1 insertion(+)

diff --git a/dankclaw/Containerfile.comfyui b/dankclaw/Containerfile.comfyui
index 153d341..38eb82d 100644
--- a/dankclaw/Containerfile.comfyui
+++ b/dankclaw/Containerfile.comfyui
@@ -36,3 +36,4 @@ RUN chmod +x /opt/ComfyUI/comfyui-start.sh
 
 CMD ["/opt/ComfyUI/comfyui-start.sh"]
 EXPOSE 8188
+# v2: VAE 4D-to-5D latent expansion for video VAEs (qwen_image21)