Every time I make a video with MiniMax H3, I wait a few minutes, sometimes tens of minutes, and re-running after a small prompt tweak starts to feel like a burden.
That's why Turbo LoRAs that cut the default 20 to 50 generation steps down to 4 to 8 steps are catching on.
✦ Four LoRAs, Four Step Counts
Published run steps for each
Trouble is, there are so many kinds that it's hard to tell which is which.
So I'm going to compare four LoRAs getting attention right now: Pika H3 Flex, LightX2V Turbo, Alibaba ACC and DMAD. All of them run MiniMax H3 fast, but they differ in which scenes they handle well and in how they cut the steps.
LightX2V's 4 Steps Are Fast
LightX2V's strength is that you can switch between 4 and 8 steps depending on the situation.
One Japanese creator measured average times while making 5.17-second videos on an RTX 4070 12GB: 273 seconds for default H3 at 20 steps, 131 seconds for LightX2V at 8 steps, and 79 seconds at 4 steps.
The 4-step number comes from running the 8-step v1.0 LoRA with only the generation steps lowered to 4. Still, under the same conditions, the 4-step run was about 39% faster than 8 steps, and generation time dropped to roughly a third of the default setup.
✦ LightX2V: When to Use 4 or 8 Steps
Pick by the job
But 4-step results are a bit disappointing as final videos. In a clip of a person getting up and walking, afterimages trail behind the motion, and in a weapon-swinging scene the hands and the weapon look shaky.
In the end, 4 steps is for quickly testing composition and prompts, while 8 steps suits final output. New 4-step versions keep coming out, but some say the results got better and others say the older version was better.
There was also a singing-and-guitar test, and both LightX2V 4-step and 8-step had problems with music overlapping or falling apart.
Of course, the common complaint from people who've tried Turbo LoRAs is that audio performance drops, so it isn't unique to LightX2V. Still, the audio quality is often a bit disappointing.
DMAD Goes After Both Speed and Quality
If you tried LightX2V and felt it doesn't suit the videos you make, DMAD is worth considering. DMAD is a speed-up technique that won higher preference than DMD2, even at 4-step generation on MiniMax H3.
The Pika H3 Flex released recently uses a technique called PDMD, so the names look alike. Both aim for the same goal, but the way they train the model to generate fast is different.
✦ DMD, DMAD and PDMD: Who Trains on What?
Same goal, different training methods
The original DMD works by taking the tendencies of videos the original model would plausibly make as the reference, measuring which way the fast model's results drift, and training to shrink that gap.
You also have to keep training a separate helper model that judges the fast model's results. That adds compute cost, and if the judge gets it wrong, the training direction can wobble too.
DMAD changed this process. Where DMD only followed videos made by the original MiniMax H3, DMAD used real video datasets for training. It compares against both original H3 videos and real footage, so the fast model learns both traits.
That way it keeps the original model's expression while gaining room to produce results closer to real footage.
Let's look at the PDMD that Pika uses too. The AI makes a video, gets rated on how well it turned out, and learns from that rating how to make the next one. But if the rating is wrong, it can learn to change even parts that didn't need fixing.
PDMD reduces the effect of those wrong ratings on training, which cuts down on learning drifting the wrong way.
So what differences showed up in actually generated videos?
✦ DMAD vs. PDMD Generation Time
RTX 5080 16GB, 4 steps each, same prompt and seed
One creator ran DMAD and PDMD at 4 steps each, with the same prompt and seed, on an RTX 5080 16GB. For 5-second videos at 832×480, DMAD took about 30.8 seconds and PDMD about 29.6 seconds, and at 720p it was about 99.8 seconds and 98.4 seconds.
Generation speed is basically the same.
The difference you could see was color. In a night-street food stall scene, DMAD kept dark tones close to default H3's cinematic mood and captured the puddle reflections, rising steam and lantern light naturally.
PDMD, on the other hand, came out with brighter, more vivid signs and surrounding colors. If you want a calm, realistic night scene, DMAD might be the one you like better.
In evaluations with people, interesting results came out too. With everything run at 4 steps on MiniMax H3 and ties excluded, DMAD got 79.1% of the choices against DMD2, which got 20.9%. Against rCM, DMAD got 84.6% and rCM got 15.4%.
DMAD led in the detailed categories too. Visual quality was DMAD 84.2% to DMD2 15.8%, and DMAD 91.3% to rCM 8.7%.
Audio quality was DMAD 87.9% to DMD2 12.1%, and DMAD 86.0% to rCM 14.0%. Motion quality was DMAD 60.3% to DMD2 39.7%, and DMAD 73.8% to rCM 26.2%, while prompt adherence was DMAD 54.3% to DMD2 45.7%, and DMAD 61.1% to rCM 38.9%.
✦ Night Food Stall Test: Where the Prompt Slipped
Good color doesn't mean good prompt following
Of course, how well it follows real prompts is a separate matter. In the night food stall test from earlier, the prompt asked for a cat getting up and stretching, but the 720p result showed the sitting cat lying down instead.
A shop owner nobody asked for even showed up in the background. The same problems appeared in the PDMD result. Even with good color and overall polish, detailed motion and background composition don't always come out as prompted.
✦ DMAD 4-Step vs. 8-Step
The 8-step version was added October 29
On October 29, an 8-step version of DMAD was added too. It isn't just the existing 4-step LoRA switched to run eight times, but a model trained separately for 8-step generation from the start.
The dev team says this version noticeably improves video and audio quality over 4 steps. So for quick repeated generation, go with 4 steps, and for work where quality matters more, 8 steps.
In the end, DMAD's appeal is that it earned good reviews for color, visual polish and audio quality even at 4 steps. If you need fast motion or complex actions rendered accurately, use 8 steps.
Pika Is Strong in Low-Motion Videos
Pika H3 Flex, built on PDMD, is the one drawing the most attention lately, and it's getting good reviews for scenes showing a person's face, skin, and a product in hand.
Its default is 6 steps, so it has a clear speed edge over existing LoRAs, most of which default to 8 steps.
In a comparison test making an 8-second product demo video on an RTX 5090, with the same prompt and reference photo, Pika at 6 steps took 116 seconds, LightX2V at 8 steps took 148 seconds, and default H3 at 20 steps took 327 seconds.
Pika's result had natural skin, and in the scene of opening a cosmetics bottle and pouring it into a palm, there was no error of extra fingers. The LightX2V video in the same test had plastic-looking skin and added fingers.
✦ Pika H3 Flex: Where It Shines and Slips
Reported results, by scene
Judging from this alone, Pika looks good, but when the video length and scenes change, the verdict changes too. In another experiment making 12 videos of 15 seconds each on an RTX 5090, 233 frozen frames were found in the Pika videos.
Six of the 12 had no freezes, but problems turned up in every other video. Faces held up well and overall quality is good, but with long videos and lots of motion, the results are hit or miss.
There were differences in a test generating a 10-second video of five characters speaking in turn, too.
In the results made on an RTX 5060 Ti 16GB with the same prompt and seed, Pika at 6 steps had scene changes that didn't line up with each speaker, and the characters' voices came out all alike.
LightX2V at 8 steps, on the other hand, told the male and female voices apart and matched the order of lines and the on-screen placement better. It took 826 seconds, longer than Pika's 638 seconds, but its lines and scene composition were more accurate.
So Pika is a good fit for videos that center on one or two people and show faces and products. It should work especially well for low-motion videos.
If you often make scenes where several people take turns speaking and there's a lot of motion, a different LoRA is probably the better choice.
Leave Action Scenes to Alibaba ACC
In a comparison making a sword-duel scene on an RTX 3090 with the same prompt and seed, Alibaba ACC at 8 steps took about 1.5 times as long to generate as LightX2V at 4 steps. But on quality, Alibaba ACC clearly wins.
In the LightX2V video, the crossed-swords pose lasted almost unchanged for a few seconds, but in the ACC video the two characters moved apart and went on to the next move. In the fight scene too, the progression of the motion was clearer.
In action scenes, what matters isn't one cool frame of someone holding a sword but hits and dodges flowing crisply over time, right? ACC especially shows its strength when you're making scenes like that.
✦ Alibaba ACC: The Trade-offs
What comes with the action strength
ACC also needs a dedicated ComfyUI node, and the setup is fussier than the other LoRAs. It doesn't work properly with the regular LoRA loading method (though there are versions tuned so you can use it that way).
Still, if the videos you want to make have lots of intense motion or many action cuts, ACC is worth considering.
Wrapping Up
If it's a video of a person holding a product and explaining it to the camera, try Pika first. To generate fast, pick LightX2V 4-step or DMAD 4-step. And for action like swordplay and fighting, ACC has the edge.
DMAD 8-step is worth a try too. It's one of the LoRAs with the best overall reviews.
That's all for this post. See you next time.
Related posts
- Photoshop Price Hike? Try Free GIMP + OpenFill AI Editing
- MAI Image 2.7 Catches Up to GPT Image 2.5 in Image Editing!
- Qwen Keeps Reaching for Alibaba's Servers: Is Local AI Safe?
🌐 Korean version: Pika·LightX2V·Alibaba ACC·DMAD... 뭐가 다른 거야? MiniMax H3 고속화 LoRA 4종 비교
No comments:
No comments yet. Want to leave the first one?
Post a Comment