Activity
Mon
Wed
Fri
Sun
Oct
Nov
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
What is this?
Less
More
Muse Collective
486 members • Free
5 contributions to Muse Collective
4d •
Long-Form H3 Render Finished All 20 Chunks but Failed at Final Assembly
Hi Andy, I wanted to share something I ran into while testing the new Muse MiniMax Director with a longer-form generation. I ran a 300-second / 5-minute video with 20 chunks at 0.4 MP, single-stage, on my RTX 5080 system with 32 GB RAM. The really good news is that the workflow successfully generated all 20 chunks. It ran for about 10 hours and 47 minutes, and the generation itself appeared to work all the way through. The problem happened only at the very end when the Director tried to combine everything into the final output. I received this error: DefaultCPUAllocator: not enough memory — tried to allocate 37,165,547,520 bytes The stack trace points to: torch.cat(all_images, dim=0) So, it looks like all of the completed chunks were being combined into one large tensor in system RAM at the end, and my 32 GB machine couldn't handle that final allocation. Is there a way the Director could eventually handle longer videos by saving/assembling completed chunks progressively to disk, rather than holding all of the decoded frames in RAM until the end? Or is there already a setting/workflow method you recommend for longer videos, so this final RAM issue doesn't happen? I'm asking because the actual 20-chunk generation worked, which was really encouraging. It was only the final assembly that failed. Thanks — I'm really enjoying testing what this workflow can do.
1 like • 2d
@Andy Whittaker Any idea what the price tag will be for version two? I've had similar issues that @Mona Leggett is having.
2d •
Isolate Chunk
I'm really liking this director node! Currently I have a 203.7 second video (14 chunks) generating with "Lip-sync drive dialogue from this recording" set with an audio file, if part of this video needs a certain chunk redone is it possible to redo only one chunk that's based off of the audio file that's uploaded? I hope the question makes sense, Id rather run all chunks and fix what's needed after the generation because my computer isn't the best and it takes a while to generate (the unified loader node does help speed things up).
1 like • 2d
Hi Andy! I do believe I will love version 2 as well :) This would be phenomenal! "You can create a project, run as many chunks as you like then stop it, save project and resume where you left off the next day, while still keeping it seamless. Each chunkncan use its own seedhunt or reroll. If you don't like a result, you can try again with a new seedz and it still joins seamlessly to the prior chunks." Totally understandable about not being able to change the middle. P.S. How did you know I was working on a music video 😁
0 likes • 2d
@Andy Whittaker Really looking forward to it because I cant get through 14 chunks yet 🤪
4d •
Charater sheet
making some great character sheets now with chatgpt, shame its limited to so many gens but i have a few gmail accounts so i get a few shots, it does seem to be very good at doing these compared to other ways ive tried and never really locked in a go to workflow for this.
1 like • 2d
Hi @Thomas Miller I used z-image turbo and got pretty good results with the prompt below (I change the upper portion if I want different characters and leave the last 4 paragraphs the same): Photorealistic adult female late 30s, mixed-race, international pop-idol presence, striking naturally balanced facial features, realistic skin texture, subtle natural freckles and facial details, sophisticated editorial makeup, confident but slightly mischievous expression. A few controlled loose strands frame the face. The hairstyle should feel like a futuristic warrior hairstyle, elegant enough for a luxury fashion editorial but rebellious. physically accurate materials, realistic skin, individual hair strands, detailed jewelry, natural human anatomy, cinematic depth of field, sophisticated composition, premium fashion editorial photography combined with gritty American realism. orthographic turnaround character sheet, wide shot composition, layout: wide character turnaround sheet on a solid clean white background, margin space above heads and below feet, strictly three separate views: (far-left: complete full-body figure from head to toe, strict front view facing camera, symmetrical neutral A-pose, both arms hanging straight down slightly angled away from torso, feet flat on floor shoulder-width apart, entire body and shoes fully visible inside the frame:1.4), (center: complete full-body figure from head to toe, strict rear back view turned exactly 180 degrees, identical matching A-pose posture, exact same arm angle and identical leg stance as the front view, mirror silhouette, back of head visible, entire body fully visible inside the frame:1.5), (right: separate big portrait, direct close-up detail of the front-facing head and neck, ending at collarbones, neutral eye-level angle:1.3)
2d •
Abandoned Heart Full Video
Ok, so this is a 3 minute (ish) music video created with MiniMax H3 V2 workflow. This took 2 hours 20 minutes in almost 1 run and was all stitched together in my workflow. I had to regen 3 chunks which were 10 minutes each, otherwise could have been under 2 hours. It's only rendered at 1MP as it was a test to see how far I could take it, but I think it is pretty good. I still need to improve the skin on long renders, I have a couple of Lora's. Might be better with higher MP.
1 like • 2d
I stumbled across this LoRA recently but haven't had time to test it yet https://huggingface.co/fal/MiniMax-H3-Realism-People-LoRA Realism from fal
9d •
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.
1 like • 9d
Thank you @Andy Whittaker! That helped a lot, I'll keep playing with the numbers to find a happy medium so generations don't take forever ;)
1-5 of 5
Active 11h ago
Joined Sep 7, 2026