Live data from Hacker News

H3-metal – Native MiniMax-H3 inference for Apple Silicon

github.com

91–100 of 108 posts

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#91

On my 128GB M4 Max Mac Studio, generating a 15s 480p video with MiniMax H3 in ComfyUI takes an hour and a half. Put Codex to work on deploying it now, hoping the speed can improve quite a lot :-) Thanks anyway

> Put Codex to work on deploying it now

Which codex?

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#92
post #71

I've been using MiniMax H3 on my M5 Pro 64GB MacBook Pro through ComfyUI. It works extremely well. I had to modify the default ComfyUI workflows to use a GGUF quant (city96's ComfyUI-GGUF custom node, UnetLoaderGGUF in place of the stock loader) [0]. I use the model labeled Q5_K_M. There is Q8_0 available as well, which is 34GB and fits fine in 64GB unified memory if you keep resolution modest. The main issue is spee…

GGUF is outdated in the latest versions of Comfy-UI. If you want a good balance of size, speed and quality you should use the int8_convrot model from the official Comfy Org Repo https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/diffus...

This is good advice if you have nvidia, but for Mac does not apply currently.

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#93
post #71

I've been using MiniMax H3 on my M5 Pro 64GB MacBook Pro through ComfyUI. It works extremely well. I had to modify the default ComfyUI workflows to use a GGUF quant (city96's ComfyUI-GGUF custom node, UnetLoaderGGUF in place of the stock loader) [0]. I use the model labeled Q5_K_M. There is Q8_0 available as well, which is 34GB and fits fine in 64GB unified memory if you keep resolution modest. The main issue is spee…

GGUF is outdated in the latest versions of Comfy-UI. If you want a good balance of size, speed and quality you should use the int8_convrot model from the official Comfy Org Repo https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/diffus...

GGUF is unsupported by ComfyUI’s memory management system that enables running models much larger than fit in VRAM with tolerable efficiency via weight streaming, but for unified memory systems that system is less relevant (unless using models too big to run in unified memory AND having fast enough mass storage to benefit from direct-from-disk weight streaming.)

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#94
post #35

Earlier quoted context omitted.

This implementation is much faster on my M5 Max, like a few minutes for the same video, but on an M5 Max with 128GB, didn't test on M5 Pro. About memory, could be executed on 64GB with a few changes.

Memory bandwidtih between pro and max is double. 300gb/s vs. 600gb/s btw.

My understanding is that that tends to be more critical with LLMs than image/video gen models, which are relatively more compute vs. memory transfer intensive than LLMs

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#95
post #54

Earlier quoted context omitted.

H3 is quite uncensored, but was not trained on p0rn, so it has no anatomy clues needed to generate that kind of stuff. For softer adult content it is reported to be fine on Reddit.

loras have been fixing that for years

Other than for pure t2v usage you probably don't need a LoRA for much, I have seen evidence that it is knowledgeable enough that it can handle a fair amount of anatomy looking and behaving reasonably with just relevant cues in control (for the fl2va model) or reference (for the r2va model) images.

And the r2va model can also use video input for motion reference.

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#96

Alright I’ve been afraid to ask but have been having trouble finding What are some adult entertainment workflows in comfyui, I need best loras, best prompts to start with and the communities, are they on telegram or something?

While you will find plenty of people willing to scam you to pay for “adult entertainment workflows”, the built in templates in ComfyUI for the model—perhaps dropping in a Lora Loader node for the a Turbo lora for speed—handle running the model, the subject matter adaptation isn’t really a workflow issue but one of reference/control images/audio/videos and prompting.

For Minimax H3, more than most models, you should read (and, if you are using an LLM for prompt assistance, make it sure it has access to) the official prompt guidelines, as each of the main models (fl2va that handles text-to-video and first- and/or last-frame-to-video and r2va that handles more complex reference cases) has its own structured prompt format (with many common features).

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#97

Earlier quoted context omitted.

> I’d personally steer clear of messaging platforms for this - who knows what one might stumble into there Personally I have no interest, but sometime browse stuff out of curiosity. But this got more of my curiosity, what kind of "stuff" are you implying they might stumble upon on the open, public internet? Sure, some NSFW, horror and otherwise weird stuff is there, especially around AI generation, but hardly somethi…

One thing I read on this topic on reddit is never EVER use the word "girl" when prompting H3. So CSAM probably.

Avoiding use of “girl” when you aren’t asking for a feminine child is good advice for any model that isn't specifically trained on a specialized prompting vocabulary (e.g., danbooru tags) where “girl” has a different meaning.

But I haven't seen anything reliable about H3 being particularly special in ther regard.

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#99

Earlier quoted context omitted.

Memory bandwidtih between pro and max is double. 300gb/s vs. 600gb/s btw.

My understanding is that that tends to be more critical with LLMs than image/video gen models, which are relatively more compute vs. memory transfer intensive than LLMs

Performance might still end up being bounded by data transfer speed if SSD streaming is heavily used to make up for limited RAM. By comparison, it doesn't take many parallel-batched sessions to make LLM decode compute-bound on typical hardware (hence seeing very limited gains from even wider batching), but this just doesn't apply when streaming weights from disk, the setting is completely different.

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#100

Earlier quoted context omitted.

> I’d personally steer clear of messaging platforms for this - who knows what one might stumble into there Personally I have no interest, but sometime browse stuff out of curiosity. But this got more of my curiosity, what kind of "stuff" are you implying they might stumble upon on the open, public internet? Sure, some NSFW, horror and otherwise weird stuff is there, especially around AI generation, but hardly somethi…

One thing I read on this topic on reddit is never EVER use the word "girl" when prompting H3. So CSAM probably.

are these still 2026 problems? a simple vector index would associate girl as a synonym for adult woman as is used by both men and women in the English lexicon

even a Chinese model that thinks its Claude when asked would have inherited this association

Post reply on HN