> We found that the model's modulation weights (~40% of the total parameters) could be pruned and replaced with a functionally equivalent lookup table, dramatically shrinking the memory footprint with no loss in output quality.
Is this a common approach to reducing weights with "no loss in output quality", assuming this is true? Seems almost too simple to work. If this is doable, would this be applicable to LLMs as well?
Neat with native frame-to-frame generation, but wonder how easy it is to "link" together clips at the intersection, typically the models kind of lose the "momentum" across these stiches, being able to merge things with frame-to-frame between clips might help with this it feels like.
It is a well-known trick, given that the timestep is between 0 to 1, you can slicing them at any resolution (1000, or 10000, give or take), and then keep a look-up table for modulation scale / bias etc for each. It is quite different from quantization and it is indeed lossless.
It is also only applicable to diffusion models as only these operates at per-timestep.
They ship a complete checkpoint for easily management (inference & training) in their own infrastructure. Moving to a LUT would make training on these layers impossible. BTW, these are not useful for lightweight fine-tuning, but might still be useful if you do serious post-training work.
Of course, these are also not an issue for things like FLUX.2 which adopts DiT-Air arch, that doesn't have this wasted space issue.
It may or may not be true. The people who made this modification and the other commenters didn't do anything rigorous to verify what they did. They just eyeball it. They could very well make some other error - this has happened frequently - that developing on prod, not knowing what they are doing, has and hopefully will again solve.
"Modulation weights" here refers to weights used to adjust layer normalization depending on the task (adaLN). General-purpose LLMs generally don't have those in the first place.
One thing similar would be projecting both the head.weight and the final LLM activations into a smaller vector space, since that is basically just cosine similarity ranking step (so that would reduce the head.weight size). But again, it must be tried many times and just not working as well. LLM space is pretty saturated with tricks.
Also begs the question whether this is applicable for high-throughput applications on FPGAs, which are to my novice mind basically LUTs, right?
I remember a paper which was posted on HN a few weeks ago where somebody implemented KAN networks in FPGAs, since those can readily be approximated as LUTs.
Whoah, could this mean we can treat layers like a jpg, where we come up with a formula that estimates the weight values of a layer instead of storing all of the weights?
FWIW on a 6000 Pro it takes 68 seconds for 10 seconds 480p video, (cold) same demo workflow as you used. Set "megapixels" to 2.0 (1920x1088) and same video seems to take 5+ minutes, not sure if everything is right/correct at the moment.
As far as I can tell, the current ComfyUI nodes don't even do compilation, and I haven't looked into what attention mechanism they're using, but I'm sure with time these durations will come down even more.
People on reddit have definitely pointed out that sageattention will speed up the renders.
And it's literally the first day. Someone will make a distilled 4-8 step LoRA and we're off to the races.
Edit: did a couple of 10 second long 864/480 i2v videos on my RTX Pro 6000: sageattention bumps them up 33%, that is to say, 140.89 seconds without sageattention becomes 105.69 with sageattention on (if using the KJ Sageattention node, "allow_compile" doesn't seem to affect it, just "sage_attention" set to "auto" works fine).
EasyCache also appears to work, but does affect quality, at least with the default threshold or even down to 0.10. Still, at 0.10 threshold the same render above, with sageattention, is down to 71.33 seconds, so depending on your use case the quality hit might be worth it. Also it seems that with EasyCache the video still matches the un-EasyCached video (with the same seed), so you could use it to do seed hunting.
I'm on an RTX 5090. I told it to make a 5 second 864x480 video, it's been running for over 30 minutes and is only 35% done in the SamplerCustomAdvanced step.
EDIT: Oh, I'm an idiot. Forgot I had a llama.cpp webserver running with a model loaded. Killed it and it finished very fast.
I am particularly curious how multimodal models will work with types of knowledge that are inherently non-text. For example, SOTA LLMs really suck at electronics, especially analog electronics.
Is MiniMax H3 capable of logical / technical reasoning, or is it purely art oriented?
Suck at what aspect of (analog) electronics specifically? Not contradicting the claim, just want to understand it.
I have not tested yet, but I suspect that LLMs with a harness that can execute code can do SPICE simulations rather ok these days? I have seen MCPs for measurement equipment also, maybe they can even close the physical loop?
spice is a very bad simulation. it's not like a unit test or a VM. it works very differently to the real thing, only approximating it in some highly controlled situations.
> work with types of knowledge that are inherently non-text
What are those things exactly? AFAIK, everything we can "know" can be written down, one way or another, even analog circuits.
Also, what SOTA LLMs are you referring to? GPTs been handling analog circuits fine for quite some time, I want to say for at least one year? I've been "pair programming" a bunch of working circuits with GPT models since probably GPT 5 or so.
> AFAIK, everything we can "know" can be written down, one way or another, even analog circuits.
writing down circuit diagrams is like cooking about music.
> what SOTA LLMs are you referring to?
I have done a survey among analog electronics designers just a couple weeks ago and they all said that their forays into LLMs were great for digital electronics, code, and firmware, but for analog they were pretty terrible, with a variety of LLMs, according to everyone.
The mouse render is surprisingly good. Several of those clips stood out to be a pretty big leap in terms of current SOTA models.
The only one that looks "off" is the beverage ad video during the can opening clip, it still has that "AI smoothening" effect. Good thing this can be done pretty well using traditional rendering.
I feel like for a good while now we'll transition into a process that uses traditional "close-up" rendering/shots + AI generated wide-shots or quick cuts.
Exciting, but also troubling. This being open-weights is a massive win for the community though.
right before that, there's the part with the person hiking up the dish and 'breathing', and the clouds of water vapor coming out of their mouth don't line up with their breaths
This is still about a year and a half behind Seedance 2.0/ Seedance 2.5
But it represents a coming price pressure that will face the leading foundation models. Open source will prevent runaway costs.
Moreover, it prevents the hair-trigger platform safety checkers from shutting down creative work. Video models are notoriously bad at shutting down a huge number of requests.
Creatives will prefer to work on cloud or private GPU clusters. Waiting 10 minutes for a few seconds of 480p is unacceptable. Hobbyists will have fun, but most actual production work is happening in the cloud.
Artist's time is worth money, and they like to spin up dozens of concurrent generations at a time to more quickly explore the generation state space and make progress on completing work.
I haven't seen any user preference comparisons between Seedance 2.5 and MiniMax H3. As an upper bound, H3 cannot be more than 6 months behind Seedance 2.5 since H3 is already ahead of where Seedance was 6 months ago.
> A year and a half behind Seedance 2.0? That is a bold claim that needs evidence.
SOTA a year ago was Kling 2.5, and H3 does not look or perform at that level.
The Artificial Analysis rankings are whack. They rank Omni first, which is incredulously wrong. Google's models broadly suck, and there they all are - right at the top.
Artificial Analysis has notoriously ranked models such as Grok Imagine highly and continues to rank Happy Horse as a good model despite the model being absolute garbage.
Could it be because they are subject to broad based statistical attacks? It's easy to encode information about the origin of media in either its metadata or output frames. Or maybe there's simply no overlap between creatives and people who click on ELO scores.
I've spent thousands upon thousands of dollars generating video. I will stand by the claim that nothing touches Seedance 2.0 / 2.5.
It's good that we're getting better open weights. It puts price pressure on the foundation model companies. But these open models are not a substitute for Kling or Seedance yet. Not even close.
> The result gives a total memory footprint reduced by 66%, from 123.6 GB in full precision to 42.5 GB with the smallest models variants. Combining this with our dynamic VRAM offloading enables a next-generation 2K video model to run locally on a GPU like the RTX 3060.
Pretty cool.
But assuming you have a 16GB 3060, how long would it take to generate a 15 second clip?
I remember reading a report where people running AI-model instagram account were using insanely long and detailed prompts about the setting, lighting, makeup, pose, disposition, clothing, etc. about their models. Presumably with some reference image of the face / body to remain consistent across images.
It‘s not clear to me whether a sufficiently detailed prompt can generate actually interesting video with a natural ”texture” (for lack of a better word).
That would be the prompt. With the right assistance from Qwen3.5/Ornith I was able to achieve some amazing results. Unfortunately due to their licensing I'm not allowed to use it in the USA, so I had to halt testing.
I mean, even the demo prompts on the ComfyUI page aren't adhered to by the model. From the first prompt, one of the four lines:
> TRANSITION: a violent WHIP PAN off the rooftop that SMEARS the floating words away with it, motion-streaked —
And the video just didn't do any of that transition at all, it just replaced it with a cut. If you look at the rest of the prompts, you'll find similar lines that are just totally ignored. Except maybe the mouse one, I didn't see anything wrong with that off the bat.
Reference-to-video mode seems like all that was missing to enable completely independent cinematography as right now one couldn't stitch different scenes together properly without altering substantial portions of the scene.
"Regions such as the EU, UK, South Korea, and the US are currently developing or enforcing AI-related regulations that may have specific implications for generative video models"
I've said it before and I'll say it again, human directors are still valuable, as they use AI video editing tools to generate the shots they want and put them together in a cohesive way. Previously they might've used film and actors but if they can just prompt the AI (or create workflows as seen with ComfyUI) then they arrange them together just like how an EDM producer doesn't actually play the instruments but instead the creativity is in the arrangement.
I suspect it'll be quite a while until AI gets a good enough aesthetic sense to do this, as even with static HTML websites humans can easily see that it's AI slop.
The example video just looks like the highly produced art (TV, commercials, games) other people have created. I find it impossible to believe this wasn't trained on other people's work, and there is no protection for it. Terribly sad. A lack of original thinking is coming.
It is coming up with Seinfeld episodes even with voice overs. Unless it is trained to do so, don't think that would happen. This model is just around the corner from getting banned due to this copyright issue. (which i hope not)
As with the arts, 99.9% of people can't use these models to express vision, get attention, or achieve distribution.
The game is the same as it has always been. You still need hard work, taste, something important to say, the ability to articulate it, good timing, and luck.
Nothing has changed. We can just build faster.
What this does enable is for more to be created that caters to a wider variety of interests. It disrupts existing structures of capital allocation, production, and distribution and gives new players a chance to reshape the game.
The bar will rise and people will still be running at the same pace on the treadmill. There will be more to see, but less time to see it.
Eh, the technical examples are impressive. But normal people really don't like AI. And I don't think that's going to improve if we really start to see job displacement.
I see a future here for digital advertising, for design mock-ups, for political propaganda, for robotics model training/world simulation, for fan movies, etc.
But the film industry is struggling to get eyeballs and butts in seats already; using a tool that most people inside-and-outside the industry loathe isn't the answer to fixing that.
At most, I could see it used for pre-viz/pre-production work in Hollywood.
Is this a common approach to reducing weights with "no loss in output quality", assuming this is true? Seems almost too simple to work. If this is doable, would this be applicable to LLMs as well?
Neat with native frame-to-frame generation, but wonder how easy it is to "link" together clips at the intersection, typically the models kind of lose the "momentum" across these stiches, being able to merge things with frame-to-frame between clips might help with this it feels like.
It is also only applicable to diffusion models as only these operates at per-timestep.
Of course, these are also not an issue for things like FLUX.2 which adopts DiT-Air arch, that doesn't have this wasted space issue.
I remember a paper which was posted on HN a few weeks ago where somebody implemented KAN networks in FPGAs, since those can readily be approximated as LUTs.
As far as I can tell, the current ComfyUI nodes don't even do compilation, and I haven't looked into what attention mechanism they're using, but I'm sure with time these durations will come down even more.
And it's literally the first day. Someone will make a distilled 4-8 step LoRA and we're off to the races.
Edit: did a couple of 10 second long 864/480 i2v videos on my RTX Pro 6000: sageattention bumps them up 33%, that is to say, 140.89 seconds without sageattention becomes 105.69 with sageattention on (if using the KJ Sageattention node, "allow_compile" doesn't seem to affect it, just "sage_attention" set to "auto" works fine).
EasyCache also appears to work, but does affect quality, at least with the default threshold or even down to 0.10. Still, at 0.10 threshold the same render above, with sageattention, is down to 71.33 seconds, so depending on your use case the quality hit might be worth it. Also it seems that with EasyCache the video still matches the un-EasyCached video (with the same seed), so you could use it to do seed hunting.
I'm on an RTX 5090. I told it to make a 5 second 864x480 video, it's been running for over 30 minutes and is only 35% done in the SamplerCustomAdvanced step.
EDIT: Oh, I'm an idiot. Forgot I had a llama.cpp webserver running with a model loaded. Killed it and it finished very fast.
Is MiniMax H3 capable of logical / technical reasoning, or is it purely art oriented?
I have not tested yet, but I suspect that LLMs with a harness that can execute code can do SPICE simulations rather ok these days? I have seen MCPs for measurement equipment also, maybe they can even close the physical loop?
What are those things exactly? AFAIK, everything we can "know" can be written down, one way or another, even analog circuits.
Also, what SOTA LLMs are you referring to? GPTs been handling analog circuits fine for quite some time, I want to say for at least one year? I've been "pair programming" a bunch of working circuits with GPT models since probably GPT 5 or so.
writing down circuit diagrams is like cooking about music.
> what SOTA LLMs are you referring to?
I have done a survey among analog electronics designers just a couple weeks ago and they all said that their forays into LLMs were great for digital electronics, code, and firmware, but for analog they were pretty terrible, with a variety of LLMs, according to everyone.
Do you know how to ride a bike?
The only one that looks "off" is the beverage ad video during the can opening clip, it still has that "AI smoothening" effect. Good thing this can be done pretty well using traditional rendering.
I feel like for a good while now we'll transition into a process that uses traditional "close-up" rendering/shots + AI generated wide-shots or quick cuts.
Exciting, but also troubling. This being open-weights is a massive win for the community though.
devs pls fix
Really poor. In the last shot, looks more like smoke.
The fact the director even included this item in the reel speaks volumes.
But it represents a coming price pressure that will face the leading foundation models. Open source will prevent runaway costs.
Moreover, it prevents the hair-trigger platform safety checkers from shutting down creative work. Video models are notoriously bad at shutting down a huge number of requests.
Creatives will prefer to work on cloud or private GPU clusters. Waiting 10 minutes for a few seconds of 480p is unacceptable. Hobbyists will have fun, but most actual production work is happening in the cloud.
Artist's time is worth money, and they like to spin up dozens of concurrent generations at a time to more quickly explore the generation state space and make progress on completing work.
According to one user preference leaderboard, MiniMax H3 is already ahead of Seedance 2.0 based on thousands of A/B votes: https://artificialanalysis.ai/video/leaderboard/text-to-vide...
I haven't seen any user preference comparisons between Seedance 2.5 and MiniMax H3. As an upper bound, H3 cannot be more than 6 months behind Seedance 2.5 since H3 is already ahead of where Seedance was 6 months ago.
SOTA a year ago was Kling 2.5, and H3 does not look or perform at that level.
The Artificial Analysis rankings are whack. They rank Omni first, which is incredulously wrong. Google's models broadly suck, and there they all are - right at the top.
Artificial Analysis has notoriously ranked models such as Grok Imagine highly and continues to rank Happy Horse as a good model despite the model being absolute garbage.
Could it be because they are subject to broad based statistical attacks? It's easy to encode information about the origin of media in either its metadata or output frames. Or maybe there's simply no overlap between creatives and people who click on ELO scores.
I've spent thousands upon thousands of dollars generating video. I will stand by the claim that nothing touches Seedance 2.0 / 2.5.
It's good that we're getting better open weights. It puts price pressure on the foundation model companies. But these open models are not a substitute for Kling or Seedance yet. Not even close.
Pretty cool.
But assuming you have a 16GB 3060, how long would it take to generate a 15 second clip?
I remember reading a report where people running AI-model instagram account were using insanely long and detailed prompts about the setting, lighting, makeup, pose, disposition, clothing, etc. about their models. Presumably with some reference image of the face / body to remain consistent across images.
It‘s not clear to me whether a sufficiently detailed prompt can generate actually interesting video with a natural ”texture” (for lack of a better word).
> TRANSITION: a violent WHIP PAN off the rooftop that SMEARS the floating words away with it, motion-streaked —
And the video just didn't do any of that transition at all, it just replaced it with a cut. If you look at the rest of the prompts, you'll find similar lines that are just totally ignored. Except maybe the mouse one, I didn't see anything wrong with that off the bat.
There is some debate on the license for those in the US, UK, EU, plus… no comment other than whew those samples though!
You just have to pinkie promise you won't make disney mad and they will send you a licence https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/Q...
I suspect it'll be quite a while until AI gets a good enough aesthetic sense to do this, as even with static HTML websites humans can easily see that it's AI slop.
But as far as your composition tools, they are already available.
This is AGI.
As with the arts, 99.9% of people can't use these models to express vision, get attention, or achieve distribution.
The game is the same as it has always been. You still need hard work, taste, something important to say, the ability to articulate it, good timing, and luck.
Nothing has changed. We can just build faster.
What this does enable is for more to be created that caters to a wider variety of interests. It disrupts existing structures of capital allocation, production, and distribution and gives new players a chance to reshape the game.
The bar will rise and people will still be running at the same pace on the treadmill. There will be more to see, but less time to see it.
I see a future here for digital advertising, for design mock-ups, for political propaganda, for robotics model training/world simulation, for fan movies, etc.
But the film industry is struggling to get eyeballs and butts in seats already; using a tool that most people inside-and-outside the industry loathe isn't the answer to fixing that.
At most, I could see it used for pre-viz/pre-production work in Hollywood.