thaakeno46
/

ComfyUI-Workflows

@@ -1,48 +1,89 @@
----
-license: apache-2.0
-language:
-- en
-base_model:
-- Wan-AI/Wan2.2-T2V-A14B
-- Wan-AI/Wan2.2-I2V-A14B
-tags:
-- art
----
-# thaakeno's Wan Video Workflows
-Just a collection of my go-to workflows for Wan 2.2, VACE, and other GGUF video models.
-This is stuff that's been tested and actually works. No extra junk nodes, just what you need to get things done. Mostly focused on GGUF and getting decent speed without a 4090.
-# **[My WAN 2.2 Prompter](https://prompter.on.websim.com/)**
-### How To Use
-Download the `.json` file you want and just drag it onto your ComfyUI canvas.
-If you get red errors about missing nodes, use the ComfyUI Manager to install them. It's pretty good about finding what's missing.
-You'll obviously need the models, LoRAs, and VAEs too. The filenames are in the loader nodes, so just match them up with what you have.
-### Required Custom Nodes
-You'll need these installed for most of these to work.
-*   `ComfyUI-WanVideoWrapper` (Kijai's nodes are the core of most of these)
-*   `ComfyUI-VideoHelperSuite` (for all the VHS stuff, combining, loading)
-*   `comfyui-kjnodes` (lots of useful utilities and reroutes)
-*   A GGUF loader (I use city96's, but others should work)
-### Workflows Included
-**Wan 2.2 - GGUF V2V (2-Pass).json**
-My standard two-pass setup for video-to-video. It uses a high noise model for the first few steps, then switches to a low noise model to refine and add detail. Good balance of quality and keeping the original motion.
----
-These are just starting points. Tweak the steps, CFG, and LoRA weights to get what you want. Hope it saves you some time.

+---
+language:
+- en
+license: apache-2.0
+tags:
+- comfyui
+- workflows
+- image-generation
+- video-generation
+- minimax
+- minimax-h3
+- wan
+- wan2.2
+- generative-ai
+---
+# thaakeno's ComfyUI Workflows
+This is my personal collection of **ComfyUI workflows** that I actually use, test and keep updating.
+The repo originally started around my **Wan 2.2 video workflows**, but I'm expanding it with workflows for other models as I use them. The newest addition is my **MiniMax H3 Studio** workflow for image generation and editing.
+I generally try to keep the workflows practical rather than filling the graph with unnecessary nodes. Most of them are built around the model setups, quantizations and acceleration methods I actually use.
+If something changes upstream or I find a noticeably better setup, I'll update the workflows here.
+---
+# MiniMax H3 Studio
+My newest workflow is a full image-generation workspace built around **MiniMax H3**.
+Instead of using separate graphs for every image task, the workflow handles **text-to-image, image editing and multi-reference generation from one unified setup**.
+It uses my custom node:
+### [ComfyUI-MiniMax-H3-Studio](https://github.com/thaakeno/ComfyUI-MiniMax-H3-Studio)
+Current release:
+### [v0.1.0-alpha.13](https://github.com/thaakeno/ComfyUI-MiniMax-H3-Studio/releases/tag/v0.1.0-alpha.13)
+Direct custom-node ZIP:
+### [Download H3 Studio](https://github.com/thaakeno/ComfyUI-MiniMax-H3-Studio/releases/download/v0.1.0-alpha.13/ComfyUI-MiniMax-H3-Studio-v0.1.0-alpha.13.zip)
+## What the H3 workflow supports
+- **Text-to-image**
+- **FL2VA image editing**
+- **REF2VA multi-reference generation/editing**
+- Up to **9 ordered reference images**
+- Reference mentions using `@Image1` through `@Image9`
+- Per-image reference roles and retention controls
+- Integrated image upload and thumbnails
+- Automatic H3 route selection
+- Optional **Qwen3-VL image analysis**
+- Optional **Qwen3-VL prompt enhancement**
+- MiniMax H3 32B text conditioning
+- Base H3 sampling
+- **LightX v0.1 4-step acceleration**
+- Optional **PDD REF2VA acceleration**
+- Conservative and direct resolution modes
+- Direct high-resolution generation
+- 5 / 9 / 13 / 20-frame H3 image profiles
+- Experimental **T=1 Image VAE**
+- Smart final-frame selection
+- Live **TAEH3 previews while sampling**
+- Seed reruns and exact-seed reproduction
+- Conditioning, reference, VAE and latent caching
+- Built-in benchmark tools for comparing profiles and resolutions
+The goal is basically to make H3 feel less like manually assembling a video-model graph every time and more like an actual **image-generation workspace inside ComfyUI**.
+## H3 workflow
+Download/open:
+`MiniMax H3/H3_Studio_Unified_Image.json`
+The workflow starts with **zero reference images**, so it can immediately be used for normal T2I.
+For reference editing, add images directly inside the Image Director and reference them naturally in the prompt:
+```text
+Keep the identity, face and pose from @Image1.
+Use only the jacket from @Image2.
+Preserve the background and composition from @Image1.
+```