Jan Sverre Bauge

I Tested MiniMax H3 Locally – 18 Videos on One RTX 4090

I tested MiniMax H3 locally with 18 videos on one RTX 4090. Here are the setup, Turbo 8 measurements, strengths, limits and time-saving fixes.

MiniMax H3 local video testing on one RTX 4090
Eighteen local MiniMax H3 video tests revealed a practical Turbo 8 workflow for one RTX 4090.

Running MiniMax H3 locally on one RTX 4090 feels like using a small production tool, not merely a model that animates a single image. After 18 test videos, I came away with usable identity consistency, video references, audio, scene transformations and animation – often within the same workflow.

That does not mean one prompt gives you a finished film. H3 rewards direction, testing and editing: one clear physical task per clip, the right sampler, and a willingness to combine the best parts from several generations. When it works, it is tremendous fun. When it does not, you may get a hand movement that turns into soup or a line of dialogue from an alternate reality.

I tested the model from August 4 to 8, 2026 with 24 GB of VRAM. The practical winner was 12 to 13 seconds of vertical video using Turbo 8, SageAttention and RTX HIGH upscaling. Here is the full setup, the measurements and the failures actually worth knowing about.

What did I actually achieve with MiniMax H3?

I got MiniMax H3 to combine identity images, reference video with its accompanying audio, separate audio files, and first and last frames in real clips. The model can also generate native stereo audio. The official model card describes H3 as an omnimodal model, but what mattered to me was that these inputs could actually work together locally.

The first animated clip kept the character, clothes, cel style and office chair coherent for about ten seconds. The laundromat clip then tested identity, video and audio over a longer reference sequence with SageAttention and RTX upscaling. The Base run took 867.54 seconds, so the quality arrived before the speed did.

The Nordic Crime Runner was a harder stress test involving running, a backward glance, a car, a traffic cone and splashing water. Identity and anatomy held up surprisingly well while the subject moved through the scene. For music videos and narrative clips, that is far more useful than yet another person standing still and swaying in front of the camera.

Animation and office chair

Laundromat, Sage and RTX

Nordic Crime Runner

How I ran MiniMax H3 locally on one RTX 4090

The stable profile on my machine was 736 × 1312 at 24 fps, normally with 294 to 311 frames. That works out to roughly 12.25 to 12.96 seconds. I used the Turbo LoRA at strength 1.0, MiniMaxH3TurboSampler, the simple scheduler and 8 steps, then scaled the decoded frames to 1080 × 1920 with RTX Video Super Resolution set to HIGH.

The machine had an RTX 4090 with 24 GB of VRAM and 64 GB of system memory. The environment was ComfyUI v0.30.1, Python 3.12.10, PyTorch 2.10.0+cu130, SageAttention 2.2.0+cu130torch2.10.0andhigher.post5 and Triton Windows 3.6.0.post26. The H3 Turbo node was v1.1.0 at commit 96cc1ddc001617da132dd73f31cd43666bf1d8d4.

The tested model set occupied about 59.15 GiB, or 59.87 GiB with the Turbo LoRA. The key files were minimax_h3_ref2va_pruned_int8_convrot.safetensors and minimax_h3_fl2va_pruned_int8_convrot.safetensors in ComfyUI/models/diffusion_models, qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors in ComfyUI/models/text_encoders, the video and audio VAE files in ComfyUI/models/vae, and the Turbo LoRA in ComfyUI/models/loras/MiniMaxH3. Large downloads should be resumable and verified before a temporary .part file gets its final name.

Turbo must use the dedicated H3 Turbo sampler. Video and audio follow different flow schedules, and a regular sampler with very few steps can overdrive the audio stream until it distorts. The Turbo node and its use of MiniMaxH3TurboSampler are documented in the project’s GitHub repository.

Video, audio and identity in the same clip

The Sildajazz test combined one identity image, an entire video reference and the audio that belonged to that video. Sildajazz is a jazz festival in Haugesund, Norway. H3 picked up the festival setting and camera feel, placed me among the crowd and ended with a clean wave. This is generative reference use, not pixel-perfect compositing, but that is exactly why the clip shows how much the model can draw from several inputs at once.

A new variation used the same reference environment for a delayed and far more absurd reveal. The point was not the person who appeared, but that the model waited before revealing him and let the setting carry the scene first. That small detail matters when a clip is supposed to tell a story.

Sildajazz with an identity reference

Delayed Trump reveal

In the Norwegian Audio 0 test, I used an identity image and a 12-second excerpt from my own WAV file. The generated audio had a waveform correlation of 0.952816 and an energy-envelope correlation of 0.988222 at zero time offset. That shows extremely close timing, but not perfect phoneme-level lip sync. The mouth did not close during every measured pause, so I muxed the exact WAV file back into the finished demonstration.

The Nyra clip used a first and last frame across 8.7 seconds. The character walked through a portal of light and emerged as a human, with good overall continuity. A small clothing detail drifted and some extra speech had to be trimmed, but the transformation itself held together. The fact that the clip ended up on Nyra’s own channel feels rather fitting.

Norwegian Audio 0 test

Nyra first/last frame

Base 20 vs. Turbo 8 and Turbo 6

Reality Debugger in Haugesund was the cleanest A/B test. The prompt, seed, two max references, native resolution, frame count and RTX HIGH setting were identical. The scene required me to walk down a rain-soaked street, grab a cyan seam in reality, pull the street aside and step into a warm alternate world.

Wall-clock time comparison for MiniMax H3 Base 20, Turbo 8 and Turbo 6
Base 20 took 24:15, Turbo 8 took 10:45 and Turbo 6 took 8:30. Turbo 8 delivered the best practical balance in my test.
ProfileWall-clock timePractical verdict
Base 2024:15.53 – 1,455.53 secondsCleanest result with calmer texture, but too slow for heavy iteration
Turbo 810:45.28 – 645.28 secondsBest balance and my standard profile
Turbo 68:30.28 – 510.28 secondsFast preview, but visible grid and effect leakage

Base 20 preserved the hand movement, the fold, the transition and the final face, but took 24 minutes and 15.53 seconds. Turbo 8 performed the same job in 10 minutes and 45.28 seconds. That is 2.256 times as fast and a 55.7 percent reduction in wall-clock time. The skin texture became more aggressive and the final smile more intense, but the time saving was large enough to let me try more ideas.

Turbo 6 came in at 8 minutes and 30.28 seconds, 135 seconds faster than Turbo 8. The cost was obvious: the fold took on more of a laser and hologram look, while grid lines leaked across the street, jacket and parts of the face. I would use it for previews, not as my default profile for finished clips.

Base 20

Turbo 8 – best balance

Turbo 6 – fast preview

H3 works best when one visual idea has room to breathe

Running in ComfyUI gradually turned a rain-soaked street into a physical world of nodes as I went from walking to running. Identity held, and the transformation unfolded without a black or gray ending. Small pseudo-elements on the node surfaces still reveal the model, but an abstract idea became something physical and filmable.

Latent Trench, The Source Code and Caffeine Overclock try three different approaches: a physical environment opening up, a digital world taking over and an effect-driven transition. They share one clear visual idea with a person moving through it. H3 handles that better than a wish list containing five critical events.

Latent Trench

The Source Code

Caffeine Overclock

The anime pair gave me one simple rule I will carry forward. Anime Ink Dragon began with a semi-realistic image and ended up closer to Western animation and concept art, even though the prompt asked for anime. True Anime Train began with a genuine anime image and landed much closer to the intended style. The starting image speaks louder than the style words in the prompt.

Anime Ink Dragon

True Anime Train

The Haugesund time-lapse improved when the prompt described physical actions and changes of clothing in concrete terms. Walking, people and clothing changed, while the rain I asked for still failed to appear. When a scene already contains many tasks, a critical detail needs to become physically central or move into a separate clip.

Physical contact often has to be solved in the edit

The animated sequence featuring Inger Marie and me showed the real production method most clearly. Turbo delivered energy and pursuit, but grappling and handcuffs smeared together. Base 20 rendered a simpler sweep and pin much more clearly. The best video therefore became a hybrid cut from both.

That is not a defeat. One profile was best for pace, another for the moment of contact. When hands, grips or grappling merge together, the choreography should be reduced to one clear point of contact. Base 20 can handle the moment where precision matters, while Turbo carries the rest of the movement.

The same principle applies to audio. H3 generates native stereo at 32 kHz, and the soundscape makes drafts feel much more complete. But both Reality Debugger variations mispronounced the English line as something like “Reality is just the deafualt sesting,” and some clips added words. When the wording must be exact, I listen through the result and mux the master audio back in.

How long and sharp can the clips realistically be?

Roughly 12 to 13 seconds was the stable, practical range for vertical production. Around 15 seconds can work, but the tail may collapse into gray texture after the action has finished. The fix is to trim at the last clean frame, not discard the whole clip. A 20-second attempt was not a safe working profile, and one wrapper also hit a tensor conflict between modalities.

Native 1088 × 1920 worked for 124 frames, or 5.17 seconds, but took 11 minutes and 45.32 seconds with Turbo 8. A 311-frame attempt at the same resolution reset the NVIDIA driver with SageAttention. Without Sage, the job remained stable at about 23.6 GiB of VRAM, but reached only 13 percent after 1 hour, 25 minutes and 37.9 seconds. I stopped it.

That is why full local 2K is not the useful story on this machine. Running at 736 × 1312 and using RTX HIGH to reach 1080 × 1920 gives a far better balance of time and result. In a controlled test, HIGH produced about 81 percent more edge energy than Lanczos, with 8.6 percent more variation between neighboring frames. ULTRA produced only about 2 percent more edge energy than HIGH, but 7.6 percent more temporal variation.

The failures and fixes that save the most time

The first smoke test failed because PathchSageAttentionKJ was set to auto but still tried to import SageAttention when it was not installed. I bypassed the node in my workflow copy and ran the control test with standard Comfy attention. I then installed wheels that matched the Python, Torch and CUDA matrix and ran pip check. Do not blindly run pip install --upgrade -r requirements.txt; it can replace the correct CUDA build of Torch with an incompatible or CPU-only package.

A warmed-up short test with Sage took 48.223 seconds, compared with 70.15 seconds using standard attention – 31.3 percent less time and 1.455 times the throughput. The first cold Sage run took 123.08 seconds because of warm-up and compilation. One cold number is a poor benchmark.

The RTX node had another concrete trap. A nested resize_type object failed validation before the job began. The dimensions must be passed as flat fields:

{
  "resize_type": "target dimensions",
  "resize_type.width": 1080,
  "resize_type.height": 1920,
  "quality": "HIGH"
}

The other fixes follow the same pattern: use MiniMaxH3TurboSampler when running Turbo with very few steps, trim a gray tail at the last clean frame, use clean identity images when max references pull the background into the result, and reduce the number of simultaneous tasks when a critical detail disappears. It sounds almost too simple. It is often what works.

Who is MiniMax H3 useful for?

H3 is most compelling if you make music videos, small cinematic scenes, funny clips or anything else where identity, movement, setting and audio need to meet. A local model does not have to beat every cloud service on every individual image to be valuable. The ability to keep testing without a per-attempt charge and inspect the entire graph is a productivity gain of its own.

If you only need one quick clip without any setup, a cloud service is easier. Here you must download about 60 GiB of models, keep the Python, Torch and CUDA versions aligned, and accept about ten minutes for a good Turbo 8 attempt. In return, you get a workflow in which video, images and audio can guide the same generation.

MiniMax H3 has open weights, but it uses its own community license rather than a standard FOSS license. Norway is not among the excluded territories, but you should read the agreement before local or commercial use.

My verdict after 18 videos is that Turbo 8 moves H3 from an impressive demonstration to something you can actually iterate with. Ten minutes is still ten minutes, but it is short enough to try several ideas in one evening. Compared with my local test of LTX 2.3, H3 is especially exciting because multiple reference types and native audio come together in one system.

The best finished result may combine Turbo movement, a Base 20 contact moment and external master audio in the edit. “One generation” is a rather poor measure of real production value. This is the AI video model that kept me trying.

The workflow I started with

I usually build my own ComfyUI workflows, but this time I began with Nerdy Rodent’s tidy all-in-one workflow. I preserved the original before continuing in a copy. It combines text-to-video, a starting image, first and last frames, multiple image, video and audio references, model switching and export in one graph.

The walkthrough runs for 15:35 and is a good starting point if you want to understand the graph itself. My tests went further with SageAttention, RTX HIGH, Turbo 8 and 6, Norwegian Audio 0, A/B measurements and final editing.

Workflow and walkthrough: Nerdy Rodent

Frequently asked questions

Can MiniMax H3 run locally with 24 GB of VRAM?

Yes. In my test, an RTX 4090 with 24 GB of VRAM generated roughly 12 to 13 seconds of vertical video at 736 × 1312 before RTX HIGH upscaled it to 1080 × 1920. That is a documented profile for this machine, not a promise of identical speed on every 24 GB card.

Is Turbo 8 better than Base 20 in MiniMax H3?

Turbo 8 offered the best practical balance, not the highest absolute quality. It took 10:45.28 compared with 24:15.53 for Base 20 while preserving the main action. Base produced calmer texture and could perform better on difficult physical contact.

Can MiniMax H3 generate Norwegian speech and lip sync?

H3 can follow a Norwegian audio reference closely and generate speech-like mouth movement, but this test does not prove perfect phoneme-level lip sync. When the words must be exact, check the audio and mux the original WAV file back into the finished clip.

How much storage does MiniMax H3 need locally?

The model set in my test occupied about 59.15 GiB, or 59.87 GiB with the Turbo LoRA. ComfyUI, Python, Torch, custom nodes and generated videos require additional space. Keep plenty of free storage and use resumable downloads for the largest model files.

Leave a Reply

Your email address will not be published. Required fields are marked *