Everything on ComfyUI
10 mentions, newest first.
Reddit user makes a 2.5-minute sci-fi pilot with MiniMax H3 on one RTX 5090
A Reddit user reports a 2.5-minute hard sci-fi pilot where every shot is MiniMax H3 running locally in ComfyUI on a single RTX 5090. The setup is hybrid fl2va + ref2va in int8, with no cloud and no credits.
Usable todayComfyUI node Geiru Keystone Attention targets darkening in long H3 chains
A Reddit user posts Geiru Keystone Attention, a ComfyUI node that keeps long H3 chains from slowly darkening. It makes clip 1's clean first frame pull harder on each new clip than the frames carried over from the previous clip. It runs only on the 13 blocks the author found carry brightness (31-33, 36-44, 47), so motion is mostly untouched. The update is described as brighter chains and half the VRAM.
Usable todaycomfyui-krea2edit-advanced fork aims to stop drift when chaining Krea 2 Edit edits
A Reddit user built comfyui-krea2edit-advanced, a fork of lbouaraba's comfyui-krea2edit, to stop drift when chaining several edits. Without it, the parts you didn't touch slowly shift, darken and get grainy. The original Krea 2 Edit LoRAs, nodes and training code come from lbouaraba / conradlocke, with the identity-edit LoRA on Hugging Face (conradlocke/krea2-identity-edit).
Usable todayReddit users share MiniMax H3-Image settings, render times and a 2-step LoRA
A Reddit user reports the open-weight MiniMax H3-Image takes up to 16 reference images. Another reports a native 100 MP render in about 240 seconds on an RTX 5090 with visible issues, and a cleaner 61 MP render in about 110 seconds. A Reddit post reports a 2-step LoRA is out on Hugging Face.
Usable todayReddit user shares a MiniMax H3 Image 20MP ComfyUI workflow with 6-step LCM
A Reddit user shares a MiniMax H3 Image 20MP ComfyUI workflow, saying the official ComfyUI workflows for MiniMax Image do a poor job of showing what the model can do. The user says the settings come from Kijai's MiniMax testing in Discord: SamplerLCM, the Simple scheduler, 6 steps, s_noise: 1.
Usable todayMiniMax H3-Image arrives as an open-weight image model with ComfyUI weights
Reddit posts say MiniMax H3-Image was released as an open-weight image model built on MiniMax H3, originally a video generation model. Weights are linked on Hugging Face from MiniMaxAI and as Comfy-Org (ComfyUI) weights. A Reddit user reports 8.41 MP (2368 x 3552) images in 35.03-42.53 seconds on an AMD Radeon RX 9060 XT with 16 GB VRAM and 64 GB DDR4 system RAM.
Usable todayVideo Rebirth releases HyperFlow H3 Flash, video and audio in 4 to 8 steps
HyperFlow H3 Flash is Video Rebirth's compressed, few-step version of MiniMax-H3. It generates 5.2-15 s of 24 fps video with a 32 kHz stereo soundtrack, sampling in 4, 6 (default) or 8 steps without CFG. A 5.2 s clip at 6 steps takes 8.7 s on 4x B200, 36 s on 1x H200 and 92 s on 1x RTX 5090.
Partly usableReddit user builds a video with the MCP bundled in Viggle Remake
A Reddit user reports making a video with the MCP included with Viggle Remake. The user says it currently runs with diffusers, so no consumer GPU works, even though Viggle also offers ComfyUI-compatible loras and custom nodes. Claude set it up to use a custom Comfy worker, and the user had Claude use the Codex CLI to generate the edited frames.
Partly usableReddit user shares ComfyUI settings for 15-second H3 videos in about 2 minutes
A Reddit user shares settings that generate 15-second videos in about 2 minutes with H3. The user says video quality is great and audio quality stays very high, and that the settings are for well-made videos at a good speed, not maximum realistic showcase videos. The post links the model and a Veda Sparse Attention Node, with the workflow and a sample in the comments.
Partly usableH3 Continuity Studio brings back earlier scenes in H3 videos
A Reddit user built H3 Continuity Studio, a ComfyUI tool to bring back earlier scenes in H3 videos, and submitted it to Comfy's Open Source Higgsfield contest. The demo generates a detailed room, cuts outside the house, returns inside, moves the camera to a different corner and continues.
Partly usable