A standalone INT8 MiniMax H3 Remix for live-action video, reference-to-video and synchronized generated audio. No separate LoRA loading is required.
This Remix used Singularity, Turbo, Realism and Motion models/LoRAs, with additional fine-tuning on selected real-person YouTube videos.
Colab G4 package: Open the inference notebook
Models and runtime: Public Google Drive folder
If your Colab plan provides G4 access: save a notebook copy, select G4, run Setup, upload reference images, paste a complete six-section Ref2VA prompt, then Generate. Preview and download the MP4 before deleting the runtime. The package includes our optimized nodes and operators, pinned dependencies and verified downloads. Default settings: 960×544 → 1632×928, 8+3 steps, native 24 fps and generated audio.
Trigger: r34l1sm, added automatically by the notebook.
Powered by MiniMax H3. The MiniMax H3 Community License and upstream component terms apply. License and attribution notices are included in the public package.
如认为本模型涉及侵权，请留言说明。
Description
Standalone INT8 release for MiniMax H3 video generation with synchronized audio. No separate LoRA loading is required. Colab G4 setup, download links and usage instructions are in the model description.
FAQ
Comments (15)
你放这么多辣椒不怕痔疮爆了？
𝓱𝓱𝓱𝓱𝓱𝓱
绷不住
这个模型在ns方面有帮助么
ns 没做特别的，都是提到的几个模型和 LoRA 的能力，我加的训练的微调只强化了口型和一些镜头跟随上的
No ComfyUI Workflow, no Remix, this is just a mystery box atm. (on top some trigger word?)
I tried stuff up but without luck.
workflow I used the second mining workflow, the sample video workflow I put in the notebook，What is the problem you encounter? Normally, he has nothing to do with other turbo ckpts.
TLDR: A working Remix and even embedded Workflow would do the trick.
If you're talking about ... H3-Colab-G4-api.json ... not a challenge... took minutes to patch up that API json into a working Workflow. But again results is nowhere near your samples... and looking at the evidences in that API, I'm not sure if its the right one anyway.
TLDR: euler / beta / pass 1 - 8 steps / pass 2 - 3 steps / 1280x768 @ 30 fps (???).
Gen Time on a RTX 5xxx 16GB - 13m45s.
Yeah good luck ppl, that file is deep into that .tag.gz on the google drive.
PS: I can understand something being lost in translation but I tried my best anyway.
@jdoe_55513 Hey bro, here's an easy way - just send me the workflow you're using and I can help you tune it. For 50-series GPUs, I can get the latency down to under 5 minutes
@Coffmeteor Thank you for the offer, I was mostly thinking about the other folks.
For now, I'll be fine for testing this model.
Keep up the good work.
@jdoe_55513 its based on singularity. no worth losing time, singularity is completely broken.
So its based on singularity? No point to download then, singularity is completely messed and worthless crap. Never generated anything correctly without washing out and blurring.
It works exceptionally well for me and delivers very satisfying performance. While individual environments vary—so others might not get the same results—this is, at the very least, the best-performing model I’ve used recently. Although it’s not entirely without its shortcomings, I have tried several other models (including singularity v 1.3 and v2), and personally, yours yields the best results for me. Thank you for sharing it.
Thank you so much for the kind words! Really glad to hear it’s working so well for you.
And you’re absolutely right — compatibility and generalizability are definitely areas where there’s room for improvement. This is really just a hobby project, and I’m mostly sharing what I’ve been using myself rather than trying to make something that works perfectly for everyone.
If I get more time in the future, I’ll definitely try to put more effort into making it work well across different setups. Thanks again for the support and feedback!
Been using this for a few days. It is actually pretty damn good at 11 steps. I stopped using those speed loras a while ago and just did 25 steps, but this model is fast and giving nice quality at 720p