Activity
Mon
Wed
Fri
Sun
Nov
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
What is this?
Less
More
Muse Collective
648
Free
5 contributions to Muse Collective
4d •
How to know when you are bottle necking?
Some good points came out of Mike's post, which raised a question that I'm hoping others can shed some more light on. @Wojciech Solarczyk made a good point about live monitoring "finding a bottleneck requires live monitoring: are you running out of VRAM? RAM?" The question I have is at what benchmark should you be looking for to know you are bottle necking? Based on a current render, my stat's are: Ram 76%, GPU 100%, VRAM 95%. Note that I'm only using KitchenAtt and Ref2VA_int8_pruned_convrot. I know this is not the fastes for a 5070ti, but i believe it keeps the quality decent, albeit this is still an area that is a soft spot for me as there are so many models and lora variations out there. Similar to Mike, I did discover that I need to keep the chunks bellow 13s, otherwise render times turn into full days vs 15-20min.
1 like • 2d
There is realistically only one main thing to keep an eye on, preventing that one thing saves TONS of generation time. And that is Disk swap. Very few people are going to avoid offloading on to RAM .. for minimax you'd need huuuuge amounts of VRAM to prevent that. Multiple 5090's, or extremely expensive commercial grade GPU's. We're all realistically going to be offloading in to RAM. The problem is offloading in to your disk. That happens when you run out of VRAM space and RAM space. And that can be fairly costly depending on your disk type. If you are offloading to a HDD .. forget about it. You're adding enormous time to generations. Offloading to swap isn't as bad if you are offloading to an NVME drive with good speeds though. In my case my NVME is actually not far off the memory bandwidth of my 3090, so it adds only seconds to generation times. Whereas when i used to offload to a HDD ... you're talking an extra 30 minutes for a 10 second clip. Other things matter but are way more simple .. GPU speed is obvious. I'm on a 3090, if i swapped to a 4090 or 5090 speeds would increase a lot. RAM speed does matter .. but likely not as much as you'd think. RAM capacity is way more important to prevent the earlier discussed disk swap offloading. The slowest ram any of you have is likely faster than the fastest drive any of you have... And on that point actually VRAM vs gpu speed is harder to state clearly. Because a very fast GPU with little VRAM that is offloading to RAM may well outperform a slower GPU with tons of VRAM and no offloading. It depends on a lot of variables. If you can avoid disk swap offloading by making sure you have enough RAM, and get the fastest GPU you can afford, you're about as good you're gonna get. Of course you can use attention types like comfy-kitchen or sage to speed things up, as well as clevel workflow tricks and turbo lora's too.
1 like • 2d
@Rob Bertrand I just bought some 64gb DDR .. the prices are insane! It's a very annoying time for those in to A.I for sure! Im sat over here with 7TB of storage space and mine is almost totally full too .. If you have the space you WILL use it! Just make sure you are using the fastest drives for the things that need them.. the OS, models etc, slowest drives get boring storage duties (But not for models unless you dont mind longer loading times!)
5d •
📣 UPDATES: New Nodes, Models & Workflows
This is where I'll post everything new, in one place that always stays at the top: node updates, new models, new and improved workflows, and changes to the classroom. How to use it: 🔔 Hit the bell on this post and you'll be notified every time I add something. 📌 The newest updates are listed first below. Each one links to the full details or the download. 💬 Questions about an update? Ask in the comments here, or post in Support/Help. Andy
0 likes • 4d
You sir are a gentleman. Seriously looking forward to the V2 node.
1 like • 2d
@Andy Whittaker With chunk regeneration issues like this are a lot easier to fix though right. Find the chunk that has the issue, change the prompt, try again. That's one of many reasons why the V2 node is going to be amazing.
Sep 7 •
Speedup Help Needed
I've run 2 generations, using 3 characters and a location image with an output of 15 seconds. - The first run I used the Muse MiniMax H3 Unified Loader (POC) node and it took 24 minutes - The next run I used the New Subgraph node and it took 36 minutes Any suggestions on how I can speed up the workflow? It would take me a considerable amount of time to create a 90 minute series at this rate. Any suggestions are greatly appreciated! I have an RTX 3090, 24 GB VRAM and 64 GB installed ram, Windows, below is an image of the models used.
0 likes • 8d
@Mike C My experience is the same as Andy's. Ive used many minimax workflows and any time generation speed goes down so does quality. However there are things you can do. Like having separate low res / upscale stages with a fixed seed so you can iterate low res generations very quickly (2 minutes on my 3090 for 10 seconds of video at 0.4mp) then when a generation comes out how you want you unbypass the upscale section and run it with the fixed seed. I use comfy kitchen attention with minimax, that helps a bit. You can use a 4 step turbo lora instead of an 8 step, you just need to make sure you're using the correct settings for the 4 step lora. And make sure you are avoiding disk swap thrashing by having enough ram. 32gb of ram with my 3090 was not enough to avoid thrashing with minimax. After I upgraded to 64gb that issue went away.
0 likes • 4d
@Mike C Tbh i wrote this before messing with seed hunt which essentially is what that is is Andy's nodes. But i think i do prefer doing it manually, it just feels easier to do it that way. But as an idea it saves a lot of time regardless of whether you do it manually or use seed hunt!
7d •
Muse Director V2 Update
One of the community members suggested possibly having the chunks in tabs. I have looked into and come up with this. I just wanted to get some feedback on it before I decide to integrate it. Take a look at the video below 👇 ( I promise it's a short one).
Poll
20 members have voted
2 likes • 6d
Yeah this would be a definite upgrade. I can see myself setting everything up in the tabbed arrangement and maybe switching to the other view during generation. But yeah .. the tabbed arrangement is a fantastic idea.
May 19 •
Welcome to Muse Collective (Start Here ✨)
AI can now do things that used to need a studio, a crew and a big budget. Create photorealistic images. Build characters that look the same every time. Clone a voice and make it speak. Turn a still image into video, lip-sync it, even put it to music. And do it all on your own computer, without a stack of monthly subscriptions. Muse Collective is the place to learn how, whether you've never opened ComfyUI or you're already building advanced video and character systems. Not hype. Not theory. We build here. How it works: 🟢 Free for everyone: the community, the classroom lessons, and a lot of workflows, including free video, lip sync and MiniMax H3 workflows. You can go a long way without paying anything. 🟠 Paid tier: where all my advanced nodes and workflows are kept. Upgrade to get the advanced nodes I've built myself, for video, images and more, plus the ready-to-run workflows that use them. If you're in, you're ready to: ✅ Create a realistic AI twin of yourself or an AI Character ✅ Generate professional AI images and video ✅ Clone your voice and bring your character to life ✅ Make your digital twin/character work for you! ✅ Create Music Videos with perfect Lipsync ✅ Bring a story to life with MiniMax H3 workflows Step 1: Introduce yourself below 👇 (copy and paste this) 👤 Name & where you're from? 🎯 What's your goal from AI in the next six months? 🤖 Which level are you starting at? (Beginner / Intermediate / Advanced) 🔧 What do you need most help with right now? How this community works: 💡 Share wins, workflows, and experiments, even the rough early ones 🤝 Drop a like and comment on others' posts, it costs nothing and means everything 🧠 Ask questions openly, post here for community feedback or DM me for support 📈 Share your AI twin, or video, or images as you build them. Progress posts are celebrated here ❌ Zero self-promotion. Zero. This isn't a platform for self-promotion. You will be removed without warning. Step 2: Head to the Classroom tab. Start with Welcome & Roadmap, then work through the levels in order.
1 like • 8d
👤 Name & where you're from? Ben from southwest England. 🎯 What's your goal from AI by end of 2026? To have high quality A.I influencer content uploaded and beginning to draw attention. 🤖 Where are you in your journey? (Total beginner / some experience / advanced) Im pretty advanced with local A.I and comfyui in general. I have custom Llm software, custom comfy nodes etc etc. I work in cybersecurity too. 🔧 What do you need most help with right now? Effective high quality pipelines for real content in minimax and krea2. And to continue to develop some more advanced ideas to enhance quality and efficiency.
1-5 of 5
