I just tested it with CivitAI's own generator... this is amazing.. just I can't post it for some reason.. definitely I will try this on ComfyUI too.

There are some, like Ace Step 1.5, and it's kind of .. okay. I tried to make it make something good. It always flubs something along the way. Fingers crossed for this, since I've seen enthusiastic responses so far.

add a still image and make it a video

@Cocosoy29 That's more work.

It is little hard but I kinda figured out how to use this.. for me I just gave it a Hindi Lyrics generated by Gemini.. the trick is to keep things simple.. just give it simple instructions inside [ ] brackets and the lyrics.. the result turned out better than I expected.. see my example in the Gallery section of this. CivitAI currently does not allow direct .mp3 posting (allows you to download mp3) so I just used Soundtools IO visualizer ( from here https://soundtools.io/music-visualizer/ ) to convert it to mp4 and post it.

@sarcastictofu 

I used very simple prompt like: "techno, raw, kicks", expanded that by LLM, and for lyrics I just kept [Intro][Instrumental][Outro], because if empty or there is something different, I do get vocals, and I just want instrumental piece and ... it produces jpop (exaggerating here but You do get an idea).

There is absolutely no control for BPM, DURATION, any timing etc. Simple fact that it cut music to whatever it wants is just non starter to me (and it happens almost always, requesting 60s got 19s @.@).

If someone know how to get techno / electro without vocals with BPM / tempo and duration control please do let me know :).

res_2s @ 4 cfg @ 40 steps will get it listening to prompts, even short dumb ones

Practically recreates the demos from the H3 website with these settings with the int8 models

@makiaeveli Can You actually create techno music using mmH3?

I tried Your suggested setting and yes it gets better but output is nowhere near what normal person would call "techno".

Some questions:

- cfg - is it sampler or MiniMax Music3 Text Encode cfg_scale? Should these 2 cfgs be the same (I do assume they should be same)

- what is scheduler (simple) ?

For techno what do You set in prompt and, most importantly, in lyrics field?

@N0n4m3 My bad. I only changed the KSampler settings: the scheduler from simple -> beta57; and the Ksampler cfg. I didn't even see that text encode CFG -- which honestly, increasing it may improve short/stubborn prompts. But no they are different. I'd presume the encoder cfg will result in crazed hallucinations the higher its pushed.

And actually yes, it is easier to make music with the video model. I have made much more prompt-adhering songs with the actual H3 video model, but obviously only like 10 seconds.

I haven't made techno, but have made some songs with long instrumental EDM breaks (of over 30 seconds in a 2 minute song). It doesn't work every time, but I have a few prompts that try and organize the song in the description box (i.e. they describe the pace of each individual verse and chorus and bridge). If you wanted more instrumental, I'd try it like that. If you raise the encode CFG as well, it might listen more to how you describe having longer instrumental sections. The lyrics could be kept blank, or maybe just with a few "Yeah"s or whatever. Maybe a "This is DJ Name" at the beginning, but you should probably have a callout for all the song stages (like intro verse chrous outro) so the description can describe how the song goes and it actually creates length (instead of just being a 10 second song of two 'yeahs').

I wouldnt be surprised if you never get "actual" BPM control though. it'd be like asking an image model for pixel art -- its not designed to output on a perfect grid. It could likely be trained -- like a lora bpm slider would be an idea.

@makiaeveli 

Well:

1. using mmH3 (sorry in my previous post I meant to write mmM3) I can, without any trouble, create techno music. It works so well. That's why I'm so surprised that mmM3 has so much problem with it.

2. vocal free, bpm, temp, prompt adherence, genre etc. is known problem, and confirmed by model creators, apparently:

https://huggingface.co/MiniMaxAI/MiniMax-Music3/discussions/23

https://github.com/MiniMax-AI/MiniMax-Music3/issues/7

so this explains my, and others, problem with this model.

3.

for the text encoder:

cfg - higher the more accurate it fallows the prompt, tested up to 8, didn't noticed anything "weird" beyond what it's currently known. Increasing this also allow to better match song length requested (it still breaks but now it's like 50/50 so "much better")

top_k - creativity basically higher the crazier it gets

sampler:

cfg - don't affect song per say, higher (4-7) result is more prompt adherence

steps - higher the more "richer" output is, 9 will not produce noise it will be just less rich with stage sound output (and crappier), 16 does the job for me

scheduler and sampler - just change the output and the way it is created, one that I tried, lcm, euler, res_2s does produce decent output.

Just my 2 cents after another few hour attack on mmM3.

@N0n4m3 I'm just saying you're probably not going to get good control over those "vocal free, bpm, temp, prompt adherence, genre etc." without loras that very specifically target the weights necessary -- cause their little text at the beginning of prompt one time isn't gonna cut it. Like even in one song the BPM could change many times, and genre, etc. Like you can tell cause the prompting isn't time-based (and come to think of it, H3 is based on time-based prompting). Think of how ideogram you put little json bounding boxes all over the image. you aren't going to get the fine-tuned control without some form of parallel

@makiaeveli I forgot to mention that these little prompts are expanded using LLMS and examples provided in mmM3 git repo so my prompt is what is recommended.

But I do get what You mean.

Anyway there was another model HeartMuLa that had / has exactly same issue as mmM3 and ... it is basically dead now.

Anyhow not much we can do about it. Hope there will be turbo lora for this so at least we could gen faster.

Try using @ 14 [Instrumental] and [Solo] tags ( depending on how long you want the song).

I found that specifying [Instrumental] in the text box where you would actually lay out the lyrics seems to work.