beclab · Terminus-Gitbot · Aug 14, 2026 · Aug 14, 2026
diff --git a/comfyuisharev3/Chart.yaml b/comfyuisharev3/Chart.yaml
@@ -3,4 +3,4 @@ appVersion: '0.28.0'
 description: description
 name: comfyuisharev3
 type: application
-version: 1.1.20
+version: 1.1.23
diff --git a/comfyuisharev3/OlaresManifest.yaml b/comfyuisharev3/OlaresManifest.yaml
@@ -6,7 +6,7 @@ metadata:
   description: The most powerful and modular stable diffusion GUI and backend.
   icon: https://app.cdn.olares.com/appstore/comfyui/icon2.png
   appid: comfyuisharev3
-  version: '1.1.20'
+  version: '1.1.23'
   title: ComfyUI
   categories:
   - Creativity
@@ -47,7 +47,7 @@ permission:
   - Home
 spec:
   onlyAdmin: true
-  versionName: '0.30.0'
+  versionName: '0.31.0'
   featuredImage: https://app.cdn.olares.com/appstore/comfyui/1.png
   promoteImage:
   - https://app.cdn.olares.com/appstore/comfyui/1.png
@@ -135,19 +135,20 @@ spec:
     - Input / output: `appCommon/comfyui/input` and `appCommon/comfyui/output`
 
   upgradeDescription: |
-    Upgrade ComfyUI to 0.30.0
+    Upgrade ComfyUI to 0.32.0
+    Original release notes: https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.32.0
 
     **Highlights**
-    - **Expanded model and node support**: Adds support for int8 ConvRot embedding lookup, PrunaVAED (a faster LTX 2.3 decoder), MiniMax-H3 including 768p output, Ideogram P-Image, and updated xAI nodes for Grok Imagine Video 1.5.
-    - **More resilient performance and memory handling**: Introduces process-RAM weight loading with an MRU pinning policy, improves dynamic pin cleanup and pin-registration priority, disables diffusion-model gradients, and improves attention fallbacks—including cuDNN fallback on Linux when Flash Attention fails. 
-    - **Better video and audio workflows**: Adds a CRF setting to the Save Video node, moves MP4 metadata toward the beginning of files when possible, supports nested audio decoding and nested-latent previews, and fixes an LTXAV crash when no audio latent is supplied.
-    - **Security and file-access hardening**: Adds a dedicated dataset folder to reduce arbitrary folder access, uses a more reliable path-within-directory check, and restores SVG previews affected by stored-XSS download protections.
-    - **Usability and observability improvements**: Adds a configurable DETAIL logging side channel—disabled by default—fixes custom user.css loading, improves job previews by preferring media output, refreshes cached CurveEditor histograms after page reloads, and lets the Save Text node write CSV files.
-    - **Compatibility and stability fixes**: Includes AMD support in comfy-kitchen, resolves a Flux KV issue, handles Windows swap-memory query failures, updates the frontend package and workflow templates, and addresses several partner-node pricing and media-processing adjustments. 
-    - **Practical takeaway**: This release is chiefly a capability, reliability, and media-pipeline update, with notable gains for LTX/MiniMax users, video export workflows, memory-constrained environments, and safer dataset handling
-
-    Full changelog: https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.30.0
-
+    - Improved debugging for nested tensors and made memory/peak memory handling more robust, especially for H3 and MiniMax-H3 models.
+    - Updated workflow templates to the latest versions.
+    - Raised required minimum PyTorch version to 2.7.
+    - Enhanced user experience by making certain nodes and flags more discoverable and self-explanatory.
+    - Fixed issues related to upscale models on low VRAM, tiled audio decode, and CLIP vision regressions.
+    - Improved the backend with new fields, filter options, and tag filtering for assets APIs.
+    - Added support for new models and nodes, including Qwen-Image 3.0 Pro, LTX 2.5, and Grok-Imagine-Image-2.0.
+    - Optimized VAE and attention mechanisms for better performance.
+    - Increased security and clarity, including more visible hardware warnings, safer method removals, and reduced reliance on external tokenizers for Mistral and Llama models.
+    
   developer: Comfy Org
   website: https://www.comfy.org/
   sourceCode: https://github.com/Comfy-Org/ComfyUI

diff --git a/comfyuisharev3/i18n/en-US/OlaresManifest.yaml b/comfyuisharev3/i18n/en-US/OlaresManifest.yaml
@@ -82,15 +82,16 @@ spec:
     - Input / output: `appCommon/comfyui/input` and `appCommon/comfyui/output`
 
   upgradeDescription: |
-    Upgrade ComfyUI to 0.30.0
+    Upgrade ComfyUI to 0.32.0
+    Original release notes: https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.32.0
 
     **Highlights**
-    - **Expanded model and node support**: Adds support for int8 ConvRot embedding lookup, PrunaVAED (a faster LTX 2.3 decoder), MiniMax-H3 including 768p output, Ideogram P-Image, and updated xAI nodes for Grok Imagine Video 1.5.
-    - **More resilient performance and memory handling**: Introduces process-RAM weight loading with an MRU pinning policy, improves dynamic pin cleanup and pin-registration priority, disables diffusion-model gradients, and improves attention fallbacks—including cuDNN fallback on Linux when Flash Attention fails. 
-    - **Better video and audio workflows**: Adds a CRF setting to the Save Video node, moves MP4 metadata toward the beginning of files when possible, supports nested audio decoding and nested-latent previews, and fixes an LTXAV crash when no audio latent is supplied.
-    - **Security and file-access hardening**: Adds a dedicated dataset folder to reduce arbitrary folder access, uses a more reliable path-within-directory check, and restores SVG previews affected by stored-XSS download protections.
-    - **Usability and observability improvements**: Adds a configurable DETAIL logging side channel—disabled by default—fixes custom user.css loading, improves job previews by preferring media output, refreshes cached CurveEditor histograms after page reloads, and lets the Save Text node write CSV files.
-    - **Compatibility and stability fixes**: Includes AMD support in comfy-kitchen, resolves a Flux KV issue, handles Windows swap-memory query failures, updates the frontend package and workflow templates, and addresses several partner-node pricing and media-processing adjustments. 
-    - **Practical takeaway**: This release is chiefly a capability, reliability, and media-pipeline update, with notable gains for LTX/MiniMax users, video export workflows, memory-constrained environments, and safer dataset handling
-
-    Full changelog: https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.30.0
+    - Improved debugging for nested tensors and made memory/peak memory handling more robust, especially for H3 and MiniMax-H3 models.
+    - Updated workflow templates to the latest versions.
+    - Raised required minimum PyTorch version to 2.7.
+    - Enhanced user experience by making certain nodes and flags more discoverable and self-explanatory.
+    - Fixed issues related to upscale models on low VRAM, tiled audio decode, and CLIP vision regressions.
+    - Improved the backend with new fields, filter options, and tag filtering for assets APIs.
+    - Added support for new models and nodes, including Qwen-Image 3.0 Pro, LTX 2.5, and Grok-Imagine-Image-2.0.
+    - Optimized VAE and attention mechanisms for better performance.
+    - Increased security and clarity, including more visible hardware warnings, safer method removals, and reduced reliance on external tokenizers for Mistral and Llama models.
diff --git a/comfyuisharev3/i18n/zh-CN/OlaresManifest.yaml b/comfyuisharev3/i18n/zh-CN/OlaresManifest.yaml
@@ -81,15 +81,16 @@ spec:
     - 输入 / 输出：`appCommon/comfyui/input` 与 `appCommon/comfyui/output`
 
   upgradeDescription: |
-    升级 ComfyUI 至 0.30.0
+    升级 ComfyUI 至 0.32.0  
+    原始发行说明: https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.32.0
 
-    **重点更新**
-    - 新增支持 int8 ConvRot embedding lookup, PrunaVAED (a faster LTX 2.3 decoder), MiniMax-H3 包括 768p 输出, Ideogram P-Image, 以及更新后的 xAI 节点用于 Grok Imagine Video 1.5.
-    - 更健壮的性能和内存处理：引入进程 RAM 权重加载策略，改进动态 pin 清理和注册优先级，禁用扩散模型梯度，并改进注意力回退——包括 Linux 上的 cuDNN 回退当 Flash Attention 失败时。
-    - 更好的视频和音频工作流：添加 Save Video 节点的 CRF 设置，将 MP4 元数据移向文件开头（当可能时），支持嵌套音频解码和嵌套 latent 预览，并修复 LTXAV 当没有音频 latent 时崩溃的问题。
-    - 安全性增强：添加 dedicated dataset folder 以减少任意文件夹访问，使用更可靠的路径检查，并恢复受存储 XSS 下载保护影响的 SVG 预览。
-    - 可用性和可观察性改进：添加可配置的 DETAIL 日志侧通道——默认禁用——修复自定义 user.css 加载，改进作业预览，优先选择媒体输出，刷新缓存的 CurveEditor 直方图，并让 Save Text 节点写 CSV 文件。
-    - 兼容性和稳定性修复：包括 AMD 支持 in comfy-kitchen，解决 Flux KV 问题，处理 Windows 交换内存查询失败，更新前端包和工作流模板，并调整合作伙伴节点定价和媒体处理调整。
-    - 实际收获：这个版本主要是能力、可靠性和媒体流水线更新，对 LTX/MiniMax 用户、视频导出工作流、内存受限环境和更安全的 dataset 处理有显著收益。
-
-    完整更新日志：https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.30.0    
+    **主要更新**
+    - 改进了对嵌套张量的调试功能，并使内存/峰值内存处理更健壮，尤其适用于 H3 和 MiniMax-H3 等模型。
+    - 工作流模板更新至最新版。
+    - PyTorch 最低版本要求提高至 2.7。
+    - 提升了用户体验，使部分节点和标志更易被发现和理解。
+    - 修复了低显存环境下的超分模型、音频解码和 CLIP vision 的相关问题。
+    - 后端新增字段、筛选项，资产 API 支持标签过滤。
+    - 增加支持如 Qwen-Image 3.0 Pro、LTX 2.5、Grok-Imagine-Image-2.0 等新模型和节点。
+    - 优化了 VAE 及注意力机制，提升性能表现。
+    - 增强安全性和提示，例如更明显的硬件警告、更安全的方法移除，以及对 Mistral 和 Llama 模型不再依赖外部分词器。
diff --git a/comfyuisharev3/templates/server.yaml b/comfyuisharev3/templates/server.yaml
@@ -18,9 +18,9 @@
 {{- if $isGb10 -}}
 {{- $memRequest = "22Gi" -}}
 {{- end -}}
-{{- $image := "beclab/comfyui:v0.30.0-fe1.50.0-launcher0.2.50" -}}
+{{- $image := "beclab/comfyui:v0.32.0-fe1.51.3-launcher0.2.51" -}}
 {{- if $isGb10 -}}
-{{- $image = "beclab/comfyui_launcher-arm64:v0.30.0-feunknown-launcher0.2.50-arm64" -}}
+{{- $image = "beclab/comfyui_launcher-arm64:v0.32.0-fe1.51.3-launcher0.2.51-arm64" -}}
 {{- end -}}
 
 ---
@@ -437,7 +437,7 @@ spec:
             - name: TZ
               value: Etc/UTC
             - name: CLI_ARGS
-              value: '--disable-xformers --disable-smart-memory --disable-cuda-malloc --disable-dynamic-vram --front-end-version Comfy-Org/ComfyUI_frontend@v1.50.0'
+              value: '--disable-xformers --disable-smart-memory --disable-cuda-malloc --disable-dynamic-vram --front-end-version Comfy-Org/ComfyUI_frontend@v1.51.3'
             - name: HOME
               value: /root
             - name: PIP_CACHE_DIR