Matt

ORB AI Community - X page for -

[Matt @AbyssHorror]

Will always be one of my favourites. #eerie #horror

[Matt @AbyssHorror]

@nikgeneburn is there something wrong with Labs loot? There's almost nothing in any rooms recently. Playing PvE with friends and it's just empty every time. After last small patch it was good for a day then back to nothing at all since

[Henry Daubrez 🌸💀 @henrydaubrez]

AI video doesn’t need another infinite slop machine. It needs a showrunner.

Using MiniMax H3 Max, thanks @fal, I built "THIS WAY", a live film engine that remembers characters, arcs and consequences.

The audience votes live. Two futures generate ahead. The winner becomes canon and the story keeps unfolding.

It’s not perfect yet, but this feels like an important step: not infinite generation, but AI serving an actual plot through audience decisions.

[Stable Diffusion Tutorials @SD_Tutorial]

Minimax H3 🤩 LongVideos Node Pack

Make long (up to ~120s) MiniMax-H3 video + synchronised audio from a single prompt, in ComfyUI. Self-contained core's H3 support.

H3 renders one shot at a time. This node turns a written scene into a chain of shots

👇
https://t.co/ZQzzVsxnul

[cocktail peanut @cocktailpeanut]

There's too much misinformation online about which machine to buy for running AI models locally. Most aren't intentional, people are just biased toward their own use cases. Before spending money, always consult the most powerful AI model you can ask about your exact use case.

[Matt @AbyssHorror]

@cocktailpeanut One thing I would highly recommend to people that are on 32GB ram that can't afford another 32GB is to look into a cheap Gen5 NVMe for pagefile overflow. Aslong as your GPU is enough for whatever model you are using. Amazing upgrade to have at a fraction of ram pricing.

[cocktail peanut @cocktailpeanut]

MiniMax-Music3 JAM Local

I ported @victormustar's MiniMax-Music3 JAM to run on any LOW VRAM PC, locally.

Just enter anything, like "synthpop about spicy food" and get your song!

- Mac, Linux, Windows
- Generate up to 5 full minutes without restriction.

1-click install 👇

[Matt @AbyssHorror]

Well that's me now optimised a decent portion and producing 5s 768p 24fps videos in 5~ min with Minimax H3 via Comfyui.
Could definitely get this much lower if I change the resolution to something smaller and upscale.

Specs:
32GB RAM
AMD RX 9070XT

[Enze Xie @xieenze_jr]

🚀 MiniMax H3 Super Acceleration in Sol-Engine🤩

We pushed H3 far beyond our previous 3–4× optimization regime — reaching 22.2× speedup for 5s video and 27.7× for 10s video vs. the published SGLang baseline.

On a single NVIDIA GB200:
• 5s 768p: 152.3s → 6.85s (22.2×)
• 10s 768p: 414.1s → 14.93s (27.7×)

But the more interesting part may be what this means economically.

Using MiniMax’s published H3 API price as a reference, we translate inference speed directly into production economics. Under an ideal fully utilized GB200 scenario, Sol-Super can serve about 525 five-second videos/hour, corresponding to roughly $210/hour of output value at the reference API price.

Assuming $5.50/GPU-hour, that implies a 97%+ GPU-only gross margin in the idealized model.

At full utilization, one GB200 could produce:
• 12.6K 5s videos/day
• 378K videos/month
• equivalent to 525 hours of finished video per month

For us, this is the bigger point of inference optimization: a 20×+ speedup does not just reduce latency — it can fundamentally change the unit economics, serving capacity, and viable business models of video generation.

🔗 https://t.co/dI7uBD1dSo

[Matt @AbyssHorror]

Cooking up something fun for a change! #theoffice Threat Level Midnight vibes..

[Matt @AbyssHorror]

@06ERpUoJJUZvOkS No problem! Feel free to throw any questions over that you have!
If you have a newer AMD GPU that uses RDNA 4 architecture then make sure you're also using ROCm 7.2. By default it's usually an older version that installs which fails to recognise newer AMD hardware

[Matt @AbyssHorror]

Lots of optimisation still to go but so far producing 5s 768p 24fps videos in 10~ min with Minimax H3 via Comfyui.

Specs:

32GB RAM
AMD RX 9070XT

Turbo LoRa with 8 Steps

Now to start optimising and see how much it can be pushed on AMD.

[Matt @AbyssHorror]

@06ERpUoJJUZvOkS Ofcourse, happy to help as much as I can!

For AMD I use minimax_h3_fl2va_pruned_int8_convrot.safetensors

However you can actually also use the non pruned int8 model aswell!
I run ComfyUI via WSL2 on a Gen5 NVMe that also has 64GB - 96GB pagefile for overflow.

[DΞV @junwatu]

H3 generates 10s in under 3 minutes 🤯

After adding LoRA, the generation time dropped from 14 minutes to just under 3 minutes. God damn!

Stack:
- H100, 80GB
- ComfyUI latest
- minimax_h3_fl2va_pruned_bf16 (40GB)
- Kijai/MiniMax-H3_comfy on Hugging Face
- PyTorch 2.13 + CUDA 13.0

The video quality is great, but the audio really needs some processing. All video resolution is 768p.

[Matt @AbyssHorror]

@blizaine Any chance of an AMD supported version? Feels like such a longshot asking for such a thing haha

[Matt @AbyssHorror]

@cocktailpeanut @blizaine 5/5. If the above was done it would significantly improve performance on AMD hardware and unlock a much higher ceiling for potential.

[Matt @AbyssHorror]

@cocktailpeanut @blizaine 4/5 ​Ditch Triton dependencies for Windows users: Offer a pure ONNX/DirectML execution path for Windows users, or pre-compile the ROCm kernels so Windows doesn't try (and fail) to compile them on the fly.