ComfyUI¶

Development line: project:comfyui · thread comfyui-development
Last event: 2026-09-03 · 21 dated since 2023-11-24 · Researched: 2026-09-05 · confidence: medium

What it is¶

ComfyUI-H3VAE_TRT is a third-party ComfyUI extension by lihaoyun6. It compiles MiniMax-H3 VAE ONNX models into TensorRT engines and loads them as a VAE. This is a custom-node integration, not a release of the MiniMax-H3 model or a core ComfyUI update.

Development line¶

• 2023-11-24 — ComfyUI Publishes Stable Video Diffusion Workflows. On 2023-11-24, ComfyUI published first-party Stable Video Diffusion example workflows for image-to-video generation. The examples documented model placement and controls for frame count, motion, frame rate, and classifier-free guidance.
• 2024-06-12 — ComfyUI Core Adds SD3 Support. On 2024-06-12, ComfyUI added SD3 support to its core engine in commit 8c4a9be. The change established native engine support for building SD3 workflows.
• 2024-09-16 — ComfyUI Adds SD3 and Flux Inpainting ControlNet Support. On 2024-09-16, ComfyUI added support for AliMama ControlNet models for SD3 and Flux inpainting in commit f48e390. The change expanded structurally conditioned inpainting workflows for those model families.
• 2024-10-21 — ComfyUI V1 Introduces a New Interface and Desktop Beta. On 2024-10-21, ComfyUI introduced its V1 interface and opened closed-beta Desktop packaging. The release added a bundled Manager, tabs, settings, model and workflow browsers, and model-download assistance. The original interface moved to a maintenance path.
• 2024-12-06 — ComfyUI Responds to the Ultralytics Package Compromise. On 2024-12-06, ComfyUI documented its response to malicious Ultralytics 8.3.41 and 8.3.42 packages encountered through some custom-node dependency chains. The project distinguished the compromised dependency from ComfyUI core and added Manager warnings and version pinning.
• 2025-07-28 — ComfyUI Adds Day-Zero Wan 2.2 Support. On 2025-07-28, ComfyUI added native day-zero support for Wan 2.2. The release provided official mixture-of-experts text-to-video and image-to-video workflows and packaged model files.
• 2025-08-13 — ComfyUI Expands Wan 2.2 Fun InP and Control Workflows. On 2025-08-13, ComfyUI added Wan 2.2 Fun InP, Control, and LightX2V support. The workflows enabled first-and-last-frame generation, structural controls, and faster four-step generation with LoRA acceleration.
• 2025-08-26 — ComfyUI Adds Native Qwen-Image DiffSynth ControlNet Support. On 2025-08-26, ComfyUI added native Qwen-Image DiffSynth ControlNet support. The update expanded controllable image generation and documented optional cache-based workflow acceleration.
• 2025-08-27 — ComfyUI Adds Hosted Gemini and Local Qwen-Image Control Paths. On 2025-08-27, ComfyUI added two distinct execution paths: Google Gemini 2.5 Flash Image through credit-billed Partner Nodes and local InstantX Qwen-Image Unified ControlNet workflows. The local workflows supported canny, depth, pose, and soft-edge controls.
• 2025-08-28 — Second Comfy Challenge Expands the Community Workflow Program. On 2025-08-28, ComfyUI expanded its community showcase and workflow-learning program through the second Comfy Challenge. This was an ecosystem and education milestone rather than a core runtime release.
• 2025-09-24 — ComfyUI Adds Wan 2.2 Animate and Qwen Image Editing Workflows. On 2025-09-24, ComfyUI added native workflows for Wan 2.2 Animate and Qwen-Image-Edit-2509. They supported performer-driven character animation and replacement as well as multi-image editing.
• 2025-10-07 — ComfyUI 0.3.63 Makes Subgraphs Publishable. On 2025-10-07, ComfyUI 0.3.63 made subgraphs publishable as reusable blueprint nodes. The release also redesigned selection tooling for graph editing.
• 2025-10-09 — ComfyUI Adds Sora 2 Partner Nodes. On 2025-10-09, ComfyUI made Sora 2 and Sora 2 Pro available through paid Partner Nodes. The integration supported text-to-video and image-to-video generation with sound.
• 2025-10-22 — ComfyUI 0.3.66 Improves Subgraph and Template Tooling. On 2025-10-22, ComfyUI 0.3.66 added a parameter panel for subgraphs. The release also redesigned workflow-template discovery and filtering.
• 2025-11-26 — Comfy Cloud Expands Infrastructure and Feature Allowances. On 2025-11-26, Comfy Cloud announced a move to RTX 6000 Pro infrastructure and support for using personal LoRAs. It also described longer Pro runs and unified credits across Cloud GPU execution and Partner Nodes.
• 2025-12-30 — ComfyUI Begins Moving Its Canonical Repository to Comfy-Org. On 2025-12-30, ComfyUI began moving its canonical GitHub repository from comfyanonymous/ComfyUI to Comfy-Org/ComfyUI. GitHub redirects were retained, and users were advised to update their Git remotes.
• 2026-01-27 — ComfyUI Adds Day-Zero Z-Image Support. On 2026-01-27, ComfyUI added day-zero support for the non-distilled Z-Image model. The release included an official workflow and model-specific sampling guidance.
• 2026-02-05 — ComfyUI Ecosystem Announces an Open-Model Grant Initiative. On 2026-02-05, the ComfyUI ecosystem announced an open-model grant initiative described as a $1 million program alongside the Anima launch. The initiative was presented as funding for open model development.
• 2026-03-26 — ComfyUI Introduces Stable Dynamic VRAM Management. On 2026-03-26, ComfyUI documented Dynamic VRAM as a stable mechanism for reducing memory pressure and out-of-memory failures on supported Nvidia systems. Its guidance emphasized measuring total workflow time instead of isolated iteration speed.
• 2026-09-02 — Launch of ComfyUI H3VAE TensorRT integration. The maintainer announced a TensorRT MiniMax-H3 VAE version for ComfyUI that can increase speed up to 1.7x.
• The extension provides separate Compiler and Loader nodes. The checked source revision is 1b72e7439e1fc306b9dfbfd3d485866553b5960e, committed later on September 2; the original announcement did not identify a release tag. (2026-09-02) — https://github.com/lihaoyun6/ComfyUI-H3VAE_TRT/blob/1b72e7439e1fc306b9dfbfd3d485866553b5960e/README.md
• 2026-09-03 — ComfyUI update. An independent tester reported 48.8s versus 36.6s for the same 2.0MP, six-second VAE decode on an RTX 4090 with 24GB, ComfyUI 0.34.2 and TensorRT 10.12. At 768p the result was 19.6s versus 18.0s. This is one workload-specific decode test, not an overall generation-speed claim.

What changed¶

The maintainer reports up to 1.7x faster MiniMax-H3 VAE processing in a third-party ComfyUI integration. The claim concerns VAE encode/decode, not end-to-end video generation. The source announcement date is retained separately from the subsequent research.

How to use this¶

1. Clone the extension into ComfyUI/custom_nodes and install its requirements using the Python environment that runs ComfyUI. — https://github.com/lihaoyun6/ComfyUI-H3VAE_TRT/blob/1b72e7439e1fc306b9dfbfd3d485866553b5960e/README.md
2. Download the encoder and decoder ONNX files from the maintainer's linked model repository, including accompanying .data files when supplied. Put them under ComfyUI/models/vae. — https://github.com/lihaoyun6/ComfyUI-H3VAE_TRT/blob/1b72e7439e1fc306b9dfbfd3d485866553b5960e/README.md
3. Build the engines with MiniMax-H3 TRT VAE Compiler, then load them with MiniMax-H3 TRT VAE Loader. — https://github.com/lihaoyun6/ComfyUI-H3VAE_TRT/blob/1b72e7439e1fc306b9dfbfd3d485866553b5960e/README.md

Best practices¶

• Below 12GB VRAM, the maintainer suggests the w4a16_awq decoder. Treat this as an author recommendation, not a guarantee that the complete video workflow will fit your GPU. — https://github.com/lihaoyun6/ComfyUI-H3VAE_TRT/blob/1b72e7439e1fc306b9dfbfd3d485866553b5960e/README_zh.md
• Compare the same latent at your target resolution and measure VAE decode separately from total generation time. The independent test found a much smaller gain at 768p than at 2.0MP. — https://note.com/seal309midorin/n/nb75b14bbab4e

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The original announcement does not identify an immutable release tag or the installed version.
• The maintainer's maximum 1.7x claim is not an independent benchmark of complete video generation.
• A Chinese community guide was located, but its page could not be retrieved in the root verification pass. Chinese coverage here is first-party documentation; no independent Chinese benchmark is claimed.
• No workflow was executed on our GPU for this article. Hardware measurements are attributed to their authors.

Sources¶

Agent brief¶

• Subject: project:comfyui, thread comfyui-development, 21 dated events 2023-11-24 → 2026-09-03.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.