Загрузка видео...

Не удалось загрузить видео

Возникла проблема при загрузке этого видео. Это может быть связано с временными проблемами сети или видео может быть недоступно.

На главную

MiniMax H3 🧐 Comfy Tested -Minimum requirements for 480p video on this model is a 3060 with 12GB vram, 32GB of system ram and a good nvme SSD. -generating a 5 second (124 frames) 480p (864x480) video on this system and it took a bit less than 9 minutes... show more

Stable Diffusion Tutorials

12,084 subscribers

102,325 просмотров • 1 месяц назад •via X (Twitter)

Наука и технологии

Anya Rossi• Live Now

Private livecam show

Комментарии: 30

Фото профиля Prince does AI

Prince does AI1 месяц назад

5 seconds in 9 minutes is rough

Фото профиля JuwalBose

JuwalBose1 месяц назад

We should get lightning or fusionx style acceleration loras in no time ☺️ Ltx gives 1080p 15s on a perfect run for me but takes like 5 gens to get a good one. Assuming minimax gets it right in single run, math is still bad at 5s 480p. Let's hope community brings solutions 🤓

Фото профиля TRGNRD

TRGNRD1 месяц назад

Could you clarify what about image generation rather than video? How long does it take to generate a 1K–2K image on an RTX 5090?

Фото профиля Stable Diffusion Tutorials

Stable Diffusion Tutorials1 месяц назад

community is obsessed with videoGen...not found anything for ImaGens

Фото профиля Intern

Intern1 месяц назад

Releasing in 6 hrs

Фото профиля Senex

Senex1 месяц назад

Any time estimates for better / newer cards? Isn't a 3060 pretty modest?

Фото профиля Stefan Larsen

Stefan Larsen1 месяц назад

Why didn't they test on bigger GPUs? Like 5090 and H100... Does everyone have access to do this test?

Фото профиля Stable Diffusion Tutorials

Stable Diffusion Tutorials1 месяц назад

models can run smoothly on fat boys...real deal is how they behave on lower grade as everyone can't afford

Фото профиля Stefan Larsen

Stefan Larsen1 месяц назад

Smoothly? I mean, we need to know how it compares to Seedance, LTX, etc. So of course we would need to know how all GPUs perform!

Фото профиля @TheINAOG

@TheINAOG1 месяц назад

12vram????? 🙀

Фото профиля kkiran

kkiran1 месяц назад

Will DGX Spark fly with this?

Фото профиля .

.1 месяц назад

just hope it takes a similar time compared to ltx2 it would be perfect omg

Фото профиля Scribblewick

Scribblewick1 месяц назад

And this is nascent! Give it time, and we'll have true seedance quality locally.

Фото профиля Rowelski

Rowelski1 месяц назад

Maybe with rented gpu can generate in less than 3 mins

Фото профиля king gore

king gore1 месяц назад

It's 20 steps. With DMD Lora or Nvidia PDD it should be much faster. Nice I'll try it tonight!

Фото профиля Jikku Jose

Jikku Jose1 месяц назад

Great! Thats precisely my system spec! Will try this!

Фото профиля Movan

Movan1 месяц назад

Whats requirement for 2k 15s multimodal video with audio ?

Фото профиля Darrick

Darrick1 месяц назад

Really interesting. Been awhile since I used LTX. Looking forward to testing this out later.

Фото профиля Agentsfy

Agentsfy1 месяц назад

int8 version would be faster. Waiting for the release

Фото профиля Patrik Číhal

Patrik Číhal1 месяц назад

how do you have access when weights aren't released yet?

Фото профиля Stable Diffusion Tutorials

Stable Diffusion Tutorials1 месяц назад

its Comfy team tested results.

Фото профиля Emircan ERKUL

Emircan ERKUL1 месяц назад

Looking forward to see if someone play with 6800xt or rdna2 amd gpus with that and figure out faster than 9m. 9m for 5s is huge... But due to quality of that, we indeed could utilise our local gpus more

Фото профиля Pandoras Parabox

Pandoras Parabox1 месяц назад

Really looking forward to this :)

Фото профиля jeff chen

jeff chen1 месяц назад

Is it free? If you got the requirements?

Фото профиля dexteam3

dexteam31 месяц назад

@grok 已經開放了嗎

Фото профиля NiiCola 【FlaVR】

NiiCola 【FlaVR】1 месяц назад

With or without quantization?

Фото профиля Eric Kang

Eric Kang1 месяц назад

Useful hardware report. Add peak RAM/VRAM, load time, sampling settings, and seconds generated per output second to estimate throughput. H3 prompt/output references:

Фото профиля AI Mastery Guide

AI Mastery Guide1 месяц назад

A 3060 handling this is genuinely solid news.

Фото профиля Aivoxy

Aivoxy1 месяц назад

I have 16 GB and 32 GB RAM not working smoothly

Фото профиля @vectormotionlab

@vectormotionlab1 месяц назад

great

Похожие видео

(1/5) 5 seconds of video. 1.8s seconds of generation. One NVIDIA GeForce RTX 5090 on FastVideo. 🤯🚀 - FastWan-QAD, a new family of video generation models - Trained with FastVideo's Quantization-Aware Distillation (QAD) recipe. - Powered by FastVideo, we push a single NVIDIA GeForce RTX 5090 to its absolute limit: generating a 5-second 480P video in 1.8s end-to-end! 📜 Blog: 💻 Code: 💽 Model:

(1/5) 5 seconds of video. 1.8s seconds of generation. One NVIDIA GeForce RTX 5090 on FastVideo. 🤯🚀 - FastWan-QAD, a new family of video generation models - Trained with FastVideo's Quantization-Aware Distillation (QAD) recipe. - Powered by FastVideo, we push a single NVIDIA GeForce RTX 5090 to its absolute limit: generating a 5-second 480P video in 1.8s end-to-end! 📜 Blog: 💻 Code: 💽 Model:

Hao AI Lab

13,312 просмотров • 3 месяцев назад

MiniMax H3 is open weights and native in ComfyUI. MiniMax (official) H3 has text, images, video, and audio in one model, video and stereo audio out in a single pass, running locally or on Comfy Cloud. To try this workflow, link below 👇

MiniMax H3 is open weights and native in ComfyUI. MiniMax (official) H3 has text, images, video, and audio in one model, video and stereo audio out in a single pass, running locally or on Comfy Cloud. To try this workflow, link below 👇

ComfyUI

12,740 просмотров • 1 месяц назад

MiniMax H3 is live on OpenArt 🔊 MiniMax (official) • 2K video with native audio built right in • 5 to 15 second clips • Text, image, and audio to video • Start frame, end frame, and reference image support • Wonder users get unlimited H3 until August 05. Terms & conditions apply. Try it now:

MiniMax H3 is live on OpenArt 🔊 MiniMax (official) • 2K video with native audio built right in • 5 to 15 second clips • Text, image, and audio to video • Start frame, end frame, and reference image support • Wonder users get unlimited H3 until August 05. Terms & conditions apply. Try it now:

OpenArt

24,650 просмотров • 1 месяц назад

🫵You should generate every shot in Seedance 2.5 at 480p. And not because it saves credits. 480p → Topaz upscale to 4K gives you something 720p and 1080p can't: the upscaler has to reconstruct detail, and it outputs images that are softer at the edges. Which means with 480p there's even less clinical, digital sharpness. Native 1080p out of a video model looks like a render for this reason. Pair it with Dehancer for the halation and film grain and it looks more authentic. AI video is at its realest when its purposely imperfect. Cheaper, and closer to film(imo 🤷). That's the whole trick. Check out this example 👇

🫵You should generate every shot in Seedance 2.5 at 480p. And not because it saves credits. 480p → Topaz upscale to 4K gives you something 720p and 1080p can't: the upscaler has to reconstruct detail, and it outputs images that are softer at the edges. Which means with 480p there's even less clinical, digital sharpness. Native 1080p out of a video model looks like a render for this reason. Pair it with Dehancer for the halation and film grain and it looks more authentic. AI video is at its realest when its purposely imperfect. Cheaper, and closer to film(imo 🤷). That's the whole trick. Check out this example 👇

God's Way Foundation

87,896 просмотров • 1 месяц назад

Local. Open weights. Native 4K. LTX-2 is now a 100% Open Source AI video model, and I tested it on my rig! Installation, VRAM usage, and the prompts I used for this video, below 👇

Local. Open weights. Native 4K. LTX-2 is now a 100% Open Source AI video model, and I tested it on my rig! Installation, VRAM usage, and the prompts I used for this video, below 👇

TechHalla

35,380 просмотров • 8 месяцев назад

this is what 12gb of vram builds in 2026, absolute magic > rtx 3060 12gb, #1 gpu on steam > bonsai 2 27b + mtp, 5.95 gb of weights > hermes agent, 5 hours, 328k tokens written > 8 js files, 2,368 lines, zero hand written code > 50 tok/s fresh, 22 tok/s average, 125k context watch the full video, 5 hours in 12 minutes of pure dance of a local ai model on rtx 3060 12gb vram, and stay till the end for the full gameplay. this entire game was built by bonsai2, a qwen 3.8 27b dense compressed to ternary, and this small model is punching way above its weight. it built multi file engineering work using hermes agent, sure it's not fast but perfect for overnights and routine work and the quality is insane, and context holding is another best one, it does not lose the thread. i ran PrismML bonsai 1 made from qwen 3.6 27b dense and i built things with it, but this time with the latest base model qwen 3.8 these results are insane, and because i loved building with it a lot i thought many more of you would run it because this gpu exists in almost every home. so i packaged mtp, doubled the speed from 26 tok/s to 50 tok/s, packed the prefill fix in and released it on huggingface, almost 4,000 downloads in 3 days. i'll leave a link below.

this is what 12gb of vram builds in 2026, absolute magic > rtx 3060 12gb, #1 gpu on steam > bonsai 2 27b + mtp, 5.95 gb of weights > hermes agent, 5 hours, 328k tokens written > 8 js files, 2,368 lines, zero hand written code > 50 tok/s fresh, 22 tok/s average, 125k context watch the full video, 5 hours in 12 minutes of pure dance of a local ai model on rtx 3060 12gb vram, and stay till the end for the full gameplay. this entire game was built by bonsai2, a qwen 3.8 27b dense compressed to ternary, and this small model is punching way above its weight. it built multi file engineering work using hermes agent, sure it's not fast but perfect for overnights and routine work and the quality is insane, and context holding is another best one, it does not lose the thread. i ran PrismML bonsai 1 made from qwen 3.6 27b dense and i built things with it, but this time with the latest base model qwen 3.8 these results are insane, and because i loved building with it a lot i thought many more of you would run it because this gpu exists in almost every home. so i packaged mtp, doubled the speed from 26 tok/s to 50 tok/s, packed the prefill fix in and released it on huggingface, almost 4,000 downloads in 3 days. i'll leave a link below.

Sudo su

33,665 просмотров • 5 дней назад

fal head of engineering batuhan the fal guy reveals how reinforcement learning turned H3 Max into a frontier video model that generates 5 seconds of video in under 3 seconds, down from nearly 2 minutes: "It's probably the only frontier video model from a quality and output and adherence perspective that can generate videos in real time." "We took this open source base MiniMax H3 video model, what we wanted to do was not just make it faster, but improve its quality ahead of anything else by implementing reinforcement learning for verifiable concepts like video editing and prompt adherence." "We were able to get the model from 120 seconds for a 5-second video generation, almost 2 minutes, to under 3 seconds for a 5-second video generation. Almost double real time, while improving its quality ahead of the open source version and ahead of many frontier closed source models." fal

fal head of engineering batuhan the fal guy reveals how reinforcement learning turned H3 Max into a frontier video model that generates 5 seconds of video in under 3 seconds, down from nearly 2 minutes: "It's probably the only frontier video model from a quality and output and adherence perspective that can generate videos in real time." "We took this open source base MiniMax H3 video model, what we wanted to do was not just make it faster, but improve its quality ahead of anything else by implementing reinforcement learning for verifiable concepts like video editing and prompt adherence." "We were able to get the model from 120 seconds for a 5-second video generation, almost 2 minutes, to under 3 seconds for a 5-second video generation. Almost double real time, while improving its quality ahead of the open source version and ahead of many frontier closed source models." fal

MTS

31,308 просмотров • 13 дней назад

Anyone with 8GB or 12GB VRAM setups needs to understand that "-ncmoe" is the key flag to boost performance on llama.cpp Here are my results for Qwen3.6 35B A3B, with 64k q8_0 context on a 8GB RTX 3070Ti: ⚪️ no flag → 8.7 tok/s RAM: 13.6GB & VRAM: 7.8GB 🔴 -ncmoe 35 → 27.5 tok/s RAM: 12.1GB & VRAM: 4.3GB 🟢 -ncmoe 30 → 32.5 tok/s RAM: 12GB & VRAM: 5.6GB 🔵 -ncmoe 25 → 40.9 tok/s RAM: 12GB & VRAM: 6.9GB Please note the ram and vram usage you see are total usage of a windows pc, with the model running. My friend's setup: 8GB VRAM and 16GB RAM. You can boost performance by switching to Linux, just something to keep in mind. Basically, this flag keeps the MoE experts in the first X layers on your CPU + RAM, instead of eating all your VRAM straight away. This is a smart hybrid offload way that lets you run bigger models without OOM while keeping the rest on your GPU for speed. As we can see on the data, there's a sweet spot. When we lower it from 35 to 25, speed bumps +50% because there are more layers on your GPU (look at the VRAM usage). The key here is to play around with the number and fit as much as possible on your VRAM, goal is to have 1GB/800MB headroom to avoid stress. ↓ server flags below

Anyone with 8GB or 12GB VRAM setups needs to understand that "-ncmoe" is the key flag to boost performance on llama.cpp Here are my results for Qwen3.6 35B A3B, with 64k q8_0 context on a 8GB RTX 3070Ti: ⚪️ no flag → 8.7 tok/s RAM: 13.6GB & VRAM: 7.8GB 🔴 -ncmoe 35 → 27.5 tok/s RAM: 12.1GB & VRAM: 4.3GB 🟢 -ncmoe 30 → 32.5 tok/s RAM: 12GB & VRAM: 5.6GB 🔵 -ncmoe 25 → 40.9 tok/s RAM: 12GB & VRAM: 6.9GB Please note the ram and vram usage you see are total usage of a windows pc, with the model running. My friend's setup: 8GB VRAM and 16GB RAM. You can boost performance by switching to Linux, just something to keep in mind. Basically, this flag keeps the MoE experts in the first X layers on your CPU + RAM, instead of eating all your VRAM straight away. This is a smart hybrid offload way that lets you run bigger models without OOM while keeping the rest on your GPU for speed. As we can see on the data, there's a sweet spot. When we lower it from 35 to 25, speed bumps +50% because there are more layers on your GPU (look at the VRAM usage). The key here is to play around with the number and fit as much as possible on your VRAM, goal is to have 1GB/800MB headroom to avoid stress. ↓ server flags below

left curve dev

166,478 просмотров • 4 месяцев назад

zeroscope_v2 XL, A watermark-free Modelscope-based video model capable of generating high quality video at 1024 x 576 Model on Hugging Face : This model was trained with offset noise using 9,923 clips and 29,769 tagged frames at 24 frames, 1024x576 resolution. zeroscope_v2_XL is specifically designed for upscaling content made with zeroscope_v2_576w using vid2vid in the 1111 text2video extension by kabachuha. Leveraging this model as an upscaler allows for superior overall compositions at higher resolutions, permitting faster exploration in 576x320 (or 448x256) before transitioning to a high-resolution render. zeroscope_v2_XL uses 15.3gb of vram when rendering 30 frames at 1024x576

zeroscope_v2 XL, A watermark-free Modelscope-based video model capable of generating high quality video at 1024 x 576 Model on Hugging Face : This model was trained with offset noise using 9,923 clips and 29,769 tagged frames at 24 frames, 1024x576 resolution. zeroscope_v2_XL is specifically designed for upscaling content made with zeroscope_v2_576w using vid2vid in the 1111 text2video extension by kabachuha. Leveraging this model as an upscaler allows for superior overall compositions at higher resolutions, permitting faster exploration in 576x320 (or 448x256) before transitioning to a high-resolution render. zeroscope_v2_XL uses 15.3gb of vram when rendering 30 frames at 1024x576

AK

714,494 просмотров • 3 лет назад

Wan 3.0 is now live on Comfy! It generates what earlier versions couldn't: a full 30-second scene in one take, as low as 480p for speed and iteration, and up to 1080p for production quality. Click the link below to try it today 👇

Wan 3.0 is now live on Comfy! It generates what earlier versions couldn't: a full 30-second scene in one take, as low as 480p for speed and iteration, and up to 1080p for production quality. Click the link below to try it today 👇

ComfyUI

55,509 просмотров • 1 месяц назад

Found a pretty good deal for anyone using AI video regularly. MiniMax H3 Max is now available on an Unlimited plan through Pollo AI. If you’re generating videos every day, being able to use H3 Max without worrying about credits makes this one worth checking out:

Found a pretty good deal for anyone using AI video regularly. MiniMax H3 Max is now available on an Unlimited plan through Pollo AI. If you’re generating videos every day, being able to use H3 Max without worrying about credits makes this one worth checking out:

Ace

124,185 просмотров • 25 дней назад

AI video isn’t expensive. Bad prompting is. I built a JSON template for MiniMax H3 so you just drop your idea in and hit generate. 5 different use cases below, tested on magnific 👇

AI video isn’t expensive. Bad prompting is. I built a JSON template for MiniMax H3 so you just drop your idea in and hit generate. 5 different use cases below, tested on magnific 👇

TechHalla

59,326 просмотров • 1 месяц назад

WanVideo has been lot of fun, currently playing with applying sliding context windowing similar to AnimateDiff, the 1.3B model is especially suitable for this due to it's speed, here's 1025 frames in one go, under 5GB VRAM used, but took about 10 mins on a 5090 with 30 steps.

WanVideo has been lot of fun, currently playing with applying sliding context windowing similar to AnimateDiff, the 1.3B model is especially suitable for this due to it's speed, here's 1025 frames in one go, under 5GB VRAM used, but took about 10 mins on a 5090 with 30 steps.

Jukka Seppänen

14,778 просмотров • 1 год назад

MiniMax H3 is the SOTA open weight model... but the large model comes with slow gens (depending on your hardware). You can be waiting for minutes just for the generation to miss the mark Here's a ComfyUI workflow to get the most out of H3 without the wait. → H3 Turbo LoRA (4–8 steps) → Model Preview Override + Tiny AutoEncoder H3 for a fast composition preview → Bad preview? stop it, tweak the prompt or seed → Save your precious gpu time Workflow and model links below 👇

MiniMax H3 is the SOTA open weight model... but the large model comes with slow gens (depending on your hardware). You can be waiting for minutes just for the generation to miss the mark Here's a ComfyUI workflow to get the most out of H3 without the wait. → H3 Turbo LoRA (4–8 steps) → Model Preview Override + Tiny AutoEncoder H3 for a fast composition preview → Bad preview? stop it, tweak the prompt or seed → Save your precious gpu time Workflow and model links below 👇

rob - comfyui

26,990 просмотров • 1 месяц назад

I enjoyed every second of this. It gives all insight into Amorim's system too. If you've got 10 mins of your time to spare, it'd be absolutely worth it. If just 5 people see this to the end, I'm good. PS: this isn't my video and I can't find the OP here on X.

I enjoyed every second of this. It gives all insight into Amorim's system too. If you've got 10 mins of your time to spare, it'd be absolutely worth it. If just 5 people see this to the end, I'm good. PS: this isn't my video and I can't find the OP here on X.

Michael

60,866 просмотров • 9 месяцев назад

MiniMax Design unlocked a new creative world. Guess how long this video took? 10 minutes! And the crazy part? I only used one prompt. It looks like something I’d spend a week making in AE Plus, annual members get 20% OFF H3 and image generation.

MiniMax Design unlocked a new creative world. Guess how long this video took? 10 minutes! And the crazy part? I only used one prompt. It looks like something I’d spend a week making in AE Plus, annual members get 20% OFF H3 and image generation.

Fathela Esq

491,144 просмотров • 1 месяц назад

I normally HATE clickbait like this. I hate it. BUT. I did it myself on Google Maps (iPhone app and took a screen record. I have a legitimate question… Is this a breadcrumb or is there a legitimate reason it says “PERMANENTLY CLOSED” on Google Maps (iPhone App) when you search for “Board of Governors of The Federal Reserve System?” (FYI Folks it’s ALOT easier to try and prove something with a video than a bunch of screen captures)

I normally HATE clickbait like this. I hate it. BUT. I did it myself on Google Maps (iPhone app and took a screen record. I have a legitimate question… Is this a breadcrumb or is there a legitimate reason it says “PERMANENTLY CLOSED” on Google Maps (iPhone App) when you search for “Board of Governors of The Federal Reserve System?” (FYI Folks it’s ALOT easier to try and prove something with a video than a bunch of screen captures)

NewsTreason Channel 17

23,997 просмотров • 1 год назад