Last week I watched a 10-second cooking video where an Asian chef knife-skilled his way through a stir-fry, wok flames licking up in sync with a sizzle track that nobody recorded separately. The video had audio. Real audio. Then I realized the whole thing was generated on a single graphics card, and the person who made it never opened a microphone.
We crossed a line in August 2026 and almost nobody outside the AI community noticed. Two open-weight models, LTX-2.5 from Lightricks and MiniMax H3 from Hailuo, dropped within days of each other. Both generate video and audio together. Both run locally. Both work with reference images, video clips, and audio samples. And both are good enough that I am genuinely uncertain whether some of the demo clips I saw this week were shot or synthesized.
I want to walk through what each model does, what the hardware bill looks like, the stuff nobody official wants to talk about, and where this is going next. So here we go LTX-2.5 vs MiniMax H3!
LTX-2.5: the speed demon with multishot
Lightricks released LTX-2.5 on August 11, 2026, calling it an “open world model.” The marketing language is dense, so let me translate what actually matters.
The headline feature is native multishot generation. Previous versions of LTX produced one continuous shot per run. LTX-2.5 can produce multiple connected shots in a single pass, and it holds character identity, environment, lighting, voice, and visual style across the cuts. That is a big deal for anyone who has tried to stitch together AI clips into something coherent. The continuity problem has been the bottleneck for months, and this addresses it directly.
Under the hood, LTX-2.5 uses a 22B parameter transformer with a custom Gemma 4 12B text encoder. Lightricks rebuilt the video decoder entirely, calling it “Diffusion Fidelity Rendering.” Instead of locking every scene to one compression rate, the model dynamically allocates compute based on scene complexity. The practical effect: sharper faces, legible on-screen text, fewer of those melted-texture artifacts that scream “AI video” from across the room.
The model also ships with a prompt enhancer that expands short prompts into detailed cinematic instructions, and an optional auto-duration predictor that figures out how long the clip should be based on the described action. Small touches, but they remove the most tedious parts of the prompt engineering workflow.
LTX-2.5 supports native 4K HDR at up to 50 FPS. And it generates synchronized audio alongside the video through a dedicated audio VAE. No separate audio generation step needed.
The distilled model is worth talking about separately. It uses a fixed 8-step schedule with CFG=1, which means it trades some quality for dramatic speed gains. On Lightricks’ own hardware (2x GB200), a 10-second image-to-video clip takes 6.8 seconds. That is faster than real-time. On a consumer card with quantized weights, you are looking at minutes, not hours.
MiniMax H3: the omni-modal reference engine
MiniMax H3 came out a few days earlier and takes a different philosophical approach. Where LTX-2.5 is a video model with audio, H3 is explicitly designed as a general-purpose omni-modal system from the ground up.
H3 is a 33B parameter model that treats text, images, video, and audio as one unified context. The way MiniMax describes it: you can hand it a reference video for camera movement, an image of a character, and an audio clip of someone singing, and tell it to combine all three. “Reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, with the vocals matching Audio 3.” That is not three separate pipeline steps. That is one prompt.
The architecture tells you why this works. H3 uses a Qwen3-VL-32B as its encoder, feeding hidden states from its 50th layer into what MiniMax calls the H3-Omni-Transformer. This transformer jointly predicts video and audio latents in a single pass. The visual VAE compresses at f16t4d24 (16x spatial, 4x temporal, 24 latent channels), and the audio VAE handles 32kHz stereo at 40Hz latent rate.
The reference system is where H3 pulls ahead of everything else right now. The “Ref2VA” mode accepts up to 9 images, 3 video clips, and 3 audio clips simultaneously, with a maximum of 12 mixed files. You can feed it a character sheet, a motion reference, and a voice sample in the same generation.
H3 outputs at 768p by default and can regenerate at 2K through what MiniMax calls “In-Context Regeneration.” Instead of a conventional super-resolution pass, the base model re-reads its own low-resolution output alongside the original context and regenerates at higher resolution. The advantage: it can recover details that traditional upscalers can only guess at, like small text and fine patterns.
Output is 24 FPS, up to 15 seconds, with native 32kHz stereo audio. H3 supports 11 languages for dialogue generation: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.
One important catch: H3’s preprocessing system (H3-Context-IR) is not open-sourced. It is a hosted service. You can use the API to reproduce the official workflow, or follow MiniMax’s prompting guidance to build your own preprocessing. For local-only users, this means the full official pipeline requires an internet call for context processing.
Here is an example of H3 generating an animated menu with native audio:
What native audio changes
I want to spend a minute on this because I think people are underrating it.
For two years, AI video was silent. You generated a clip, then ran it through a separate text-to-audio model, then tried to sync them. The results were always slightly off. Footsteps did not quite match footsteps. Lip movements drifted from speech. Ambient sound felt pasted on.
Both LTX-2.5 and H3 generate audio in the same pass as the video. The audio is structurally synchronized because it comes from the same latent space. When a glass breaks in the video, the sound of breaking glass is in the audio. When someone speaks, their mouth moves.
H3 goes further by producing stereo audio, not mono. The left-right spatial imaging is baked into the generation. For a market scene with a vendor on the left and a customer on the right, the voices come from the correct direction. This is the difference between “AI video that looks kind of real” and “AI video that feels like footage.” The audio sync is what tricks your brain.
Download size and VRAM: the real cost
Here is the comparison nobody puts in their press release.
For LTX-2.5, the practical ComfyUI download (int8 quantized, distilled transformer) is roughly 36 GB. That includes the DiT, text encoder, video VAE, audio VAE, and spatial upscaler. If you want the full bf16 dev transformer, the DiT alone is about 44 GB. An NVFP4 variant for Blackwell GPUs shrinks the DiT to around 11 GB.
For MiniMax H3, the ComfyUI pack from Comfy-Org totals about 42.5 GB. A pruned NVFP4 variant from the community sits around 43 GB. The full bf16 model with the Qwen3-VL-32B encoder would be well over 100 GB, which is why nobody runs that locally.
The VRAM situation is more nuanced than the spec sheets suggest. Lightricks officially recommends 32 GB VRAM (A100 or H100 territory), but their own comparison table claims 16 GB minimum, and community reports confirm usable runs on 12 GB cards with GGUF quantization. One Reddit user generated 2.5K (2560×1408) 10-second clips on an RTX 5070 Ti with 16 GB VRAM and 64 GB system RAM in 11 minutes.
H3 scales more aggressively at the low end. Users have generated clips on an RTX 2060 with 6 GB VRAM, taking 10 to 15 minutes for short degraded clips. A 12 to 16 GB card produces noticeably cleaner output in less time. 24 GB gives you headroom for LoRA training alongside generation. 32 GB and above (RTX 5090, RTX 6000) is where H3 hits its intended quality tier with 15-second clips at higher resolution.
On Mac, H3 runs via a 4-bit NF4 quantization through community tools, with as little as 8 GB of unified memory. Quality drops noticeably on complex scenes.
The bottom line: if you have a 16 GB consumer card from the last three years, you can run both models. If you have 24 GB or more, you can run them well.
Side-by-side comparison
| Spec | LTX-2.5 | MiniMax H3 |
|---|---|---|
| Parameters | 22B (DiT) | 33B (BF16) |
| Max resolution | 4K HDR, 50 FPS | 2K (via regeneration), 24 FPS |
| Max duration | 10s+ (auto-duration) | 15 seconds |
| Native audio | Yes (mono, via audio VAE) | Yes (32kHz stereo) |
| Multishot | Yes, native | Yes, native |
| Reference inputs | Image, video, audio | Up to 9 images, 3 videos, 3 audio (12 total) |
| ComfyUI download | ~36 GB (int8 distilled) | ~42.5 GB (Comfy pack) |
| Min VRAM (usable) | 12-16 GB | 6-12 GB (degraded) |
| Recommended VRAM | 32 GB+ (official), 16 GB (community) | 32 GB+ (RTX 5090 class) |
| Text encoder | Gemma 4 12B (custom) | Qwen3-VL-32B (50th layer) |
| License | LTX-2.x Community (under $10M revenue) | MiniMax H3 Community License |
| Model gated? | Yes (HuggingFace access request) | No (open download) |
The elephant in the room
I am not going to pretend this section does not exist, because anyone who has spent time in the ComfyUI and Stable Diffusion communities this week already knows.
Both models have open weights. Open weights mean anyone can fine-tune them. And the community has been moving fast.
For MiniMax H3 specifically, the situation is more intense because the model was designed with reference understanding. You can hand it a character image and a voice sample. Combine that with LoRA fine-tuning, and you get a system where someone can train a LoRA on a specific person’s appearance, clone their voice through the audio reference system, and generate video of that person in any scenario.
There are already LoRAs on Civitai and HuggingFace depicting explicit adult content in significant detail. The “heretic” community builds remove safety layers from the model stack. Voice cloning means the characters speak. And because this all runs locally, there is no API provider to enforce content policy.
MiniMax is aware of this. They have started issuing takedowns on decensor and explicit LoRAs uploaded to HuggingFace. Whether that works against a decentralized community that mirrors everything to personal servers and torrent networks is an open question. Based on how this played out with Stable Diffusion, the answer is probably no.
I do not have a clean moral position on this. The capability exists because the weights are open. Open weights are also what makes legitimate creative work possible, what lets researchers study these systems, what prevents a handful of companies from controlling all video generation. You cannot really have one without the other. But I also think it is worth being honest that “everyone can create adult videos with characters of their choice, including real people, with cloned voices, entirely offline” is a thing that is now true, and the social and legal frameworks for dealing with it do not exist yet.
For what it is worth, LTX-2.5 has been less affected by this so far, partly because Lightricks gated the model on HuggingFace (you have to request access) and partly because the reference system is less aggressively multimodal than H3’s. But the LoRA infrastructure is the same. It is only a matter of time.
What is coming next
Both companies telegraphed their roadmaps, and the community is already filling gaps.
MiniMax listed three priorities for future H3 versions: integrating capabilities from their M-series language models for stronger multimodal understanding, scaling up the model size (they explicitly said “scaling is a clear path forward”), and pushing toward higher resolutions and visual fidelity. They also promised to release their sparse-attention implementation, which would significantly reduce inference costs for long sequences. That is not available yet in the open-source release, which ships with full attention only.
The community has already built what amounts to a Turbo version. Multiple “MiniMax H3 Turbo LoRA” repos appeared on HuggingFace within days, enabling 4-step text-to-video and image-to-video generation. The quality tradeoff is real but the speed gain is substantial. Expect official Turbo variants from MiniMax themselves within weeks.
For LTX-2.5, the distilled model is already the Turbo equivalent. Lightricks also mentioned applicability to “physical AI” and robotics, which suggests they see this architecture extending beyond creative video into simulation territory. The LTX Trainer supports video, audio, and cross-modal conditioning in one framework, so fine-tuning custom versions is straightforward.
Looking at the broader field: Flux 3 delivers strong results but is API-only with no open weights announced. WAN 3.0 from Alibaba supports 30-second generations but is also API-only. Seedance 2.5 and Kling 3.0 Pro remain closed. The gap between open and closed models has narrowed dramatically, and for local-first creators, LTX-2.5 and H3 are the only games in town that you actually own.
Getting started: ComfyUI workflows
Both models have official ComfyUI workflow templates ready to go.
For LTX-2.5, ComfyUI ships three native workflows: Text-to-Video, Image-to-Video, and First-Last-Frame-to-Video. You can find them in the Template Library under Video, or follow the documentation at docs.comfy.org. The workflow hub also has dedicated pages on comfy.org for LTX-2.5 T2V, I2V, and FLF2V. Search for “LTX-2.5 T2V,” “LTX-2.5 I2V,” or “LTX-2.5 FLF2V” in the template browser.
You will need to request access to the Lightricks/LTX-2.5 repository on HuggingFace first, since the model is gated. Once approved, ComfyUI’s model downloader handles the rest.
For MiniMax H3, ComfyUI offers three workflows: Text-to-Video, Image-to-Video, and Reference-to-Video. The model files are in the Comfy-Org/MiniMax-H3 repository, which is not gated. Update ComfyUI to version 0.30.0 or later, open the Template Library, and search for “MiniMax H3.” Follow the official tutorial for setup details.
The reference-to-video workflow is where H3 shines. Feed it an image, a motion reference video, and an audio clip, then describe how they should combine. The workflow handles the multimodal context packing automatically.
Both models benefit from community LoRAs and IC-LoRAs. For LTX-2.5, the LTX Community Library on ltx.io collects user-trained adapters for style, editing, and audio conditioning. Most LoRAs trained on LTX-2.3 transfer to 2.5 without changes. For H3, browse the HuggingFace model tree, which already lists 49 fine-tunes and 9 adapter models, with new ones appearing daily.
My honest take after a week with both
These models are not perfect. H3’s 15-second limit is restrictive for anything beyond short-form content. LTX-2.5’s gated access is mildly annoying. Both still produce occasional artifacts that break the illusion. Neither handles complex multi-character interaction reliably.
But I have been following AI video since the first wobbly ModelScope clips, and this is the first time I have felt that the gap between “generated” and “filmed” is closeable. Not today. But the trajectory is clear. A year ago, generating a 5-second clip at 720p with no audio took minutes and looked wrong. Today, I can generate 10 seconds at 2K with synchronized stereo audio, multiple shots, and character consistency, on hardware that fits under my desk.
If you have been waiting for AI video to get good before paying attention, this is the moment. The videos generated with both models. are already rolling out on Youtube and Tiktok.