Source: Four AI Search videos, YouTube auto-captions fetched 2026-10-30: raw/New_BEST_AI_video_generator_is_here_Seedance_2.5_vs_Minimax_H3.md (Seedance 2.5 vs MiniMax H3, https://www.youtube.com/watch?v=Vpvne2istKg, uploaded 2026-08-03; sponsored by ChatLLM / Abacus AI), raw/The_BEST_local_AI_video_generator_is_here_Minimax_H3_tutorial.md (H3 install tutorial, https://www.youtube.com/watch?v=8HutJ9W5pTg, 2026-08-05), raw/The_BEST_local_AI_video_generator_just_got_BETTER.md (H3 advanced tutorial, https://www.youtube.com/watch?v=G3YHSvXZP_g, 2026-08-12; sponsored by Higgsfield) and raw/The_FASTEST_local_AI_video_generator.md (LTX 2.5 tutorial, https://www.youtube.com/watch?v=ig3PUfSow5Y, 2026-08-18). Caption fixes: “Min Max” → MiniMax, “high low” → Hailuo, “1 to GP” → Wan2GP, “Austras AI toolkit” → ostris ai-toolkit, “Light XTV” → LightX2V, “Abby Ray” → Abiray, “ERSD” kept as captioned (a sampler name), “A-Step” → ACE-Step, “LCX” → LTX.

How to run the two leading open video models of August 2026 on your own GPU in ComfyUI, from one creator’s install walkthroughs. MiniMax H3 is the flexible one: native audio, strong world knowledge, and a reference-to-video mode that takes images, video and audio. LTX 2.5 (Lightricks) is the fast one: about 20 seconds per 5-second clip on the presenter’s laptop, up to 4K and 50 fps, and compatible with existing LTX 2 LoRAs. A side-by-side test against hosted Seedance 2.5 shows where each is stronger. For news context and other creators’ H3 reports, see the August 2026 video-model wave.

Key Takeaways

  • H3 locally runs at about one minute of generation per second of video (FP8/INT8, 480p, presenter’s 16 GB laptop GPU). SageAttention adds 20–30%, EasyCache another 5–10%, and a turbo LoRA cuts 20 steps to 4–6 for up to five times faster.
  • LTX 2.5 is two to three times faster than H3 on the same machine: about 20 s for a 5-second text- or image-to-video clip, 30 s for first/last frame, because it renders a low-res pass and then runs a spatial upscaler.
  • VRAM floors (reported, not all tested by him): H3 pruned models 21 GB, reported working on 12 GB; Kijai’s W4A8 builds (about 12 GB) for roughly 8 GB; Wan2GP claims 5 GB at 480p. LTX 2.5’s official minimum is 16 GB; a 12.6 GB GGUF should fit 12 GB.
  • H3’s licence is a community licence with two catches (as he reads it): free commercial use only below $20M annual revenue, and the EU, UK, Korea and the US are excluded unless you apply for a separate licence, which he says is a simple form. This is the licence detail the wave article lacked.
  • Seedance 2.5 vs H3 (hosted, his tests): Seedance wins fight scenes, 3D-animation and sketch-to-final restyling, and long instruction-heavy prompts (up to 30 s); H3 wins music videos cut to the beat and text-overlay effects, and costs about half as much per clip. Both handled storyboard-to-ad and app-screenshot ads; both failed a Pythagorean-theorem diagram and a world-map motion graphic.
  • H3 can also be a sound generator. Use the image-to-video workflow at 0.1 MP, bypass the video nodes, and save only the audio: orchestral music, or a line of shouted dialogue over rock music.
  • Use MiniMax’s official prompt guide. Text-to-video prompts should have three parts: an integrated description (visuals, action, camera, dialogue), the soundscape, and background music. His shortcut: paste the guide and the input image into ChatGPT and ask for a prompt.

Seedance 2.5 vs MiniMax H3 (hosted, before H3’s open release)

Tested through BytePlus’s Lumina platform (Seedance) and MiniMax’s Hailuo site (H3), one run per prompt.

TestWinner (his call)
Fight scene from a rough 3D animation plus two character imagesSeedance. H3 ignored the 3D reference and had noise and artefacts.
Long multi-action Pixar chase promptSeedance got every beat in 30 s; H3 fitted almost all of it into 15 s, missing one jump.
Sketch animation to finished cinematic clipSeedance; H3 kept some sketch elements.
Luxury-handbag ad from a storyboard and logo, with voice-overTie.
Motion-graphic app ad from UI screenshots, British female voice-overTie; minor letter errors in both.
K-pop music video from an ACE-Step track, typography references, beat-synced hard cutsH3: same song, cuts on the beat, good typography. Seedance’s was “a bit weird.”
Nine languagesNot judged (he does not speak most).
Violinist playing Vivaldi’s SummerSeedance’s bowing looked realistic but neither knew the piece.
Professor explaining Pythagoras on a whiteboardBoth wrote the formula; neither drew a correct labelled triangle.
Peru highlighted on a world map with Indian-accented history voice-overBoth failed.

Specs and price he showed (August 2026):

  • Seedance 2.5: up to 720p (1080p and 4K planned), up to 30 s per generation, extendable by feeding a clip back in; up to 50 references (30 images, 10 videos, 10 audio). On Lumina a 10-second clip cost 460 credits, 570 with a reference video; 1,000 credits ≈ 5–6 per 10 s**. “Definitely the most expensive video model out there.” API not yet open.
  • H3 (Hailuo): 2K output by default, up to 15 s. A 10-second clip cost 120 credits, 252 with a reference video, so about $1–3. API live at about 13¢ per second at 2K.

He points to Mixamo (free character animations) and an Nvidia text-to-motion tool for making the rough 3D reference animations.

MiniMax H3: install and basic workflows

  1. Update ComfyUI (update/update_comfyui.bat on the Windows portable build), then Templates → search “MiniMax”. Ignore templates tagged API, which run in the cloud. Use text-to-video, image-to-video and reference-to-video.
  2. Download the models (linked from the ComfyUI MiniMax page):
    • Diffusion model, one for text/image-to-video and a separate one for reference-to-video. Sizes: full 66 GB, INT8 34 GB, pruned FP8 or pruned INT8 21 GB. He uses FP8 on a recent Nvidia card. → models/diffusion_models
    • Text encoder Qwen3-VL-32B (three compression levels; it can sit in system RAM rather than VRAM) → models/text_encoders
    • Both the video VAE and the audio VAE → models/vae
  3. Settings: aspect ratio, megapixels (0.4 MP ≈ 480p), duration; inside the subgraph, sampler, scheduler and steps (default 20).
  4. Reference-to-video is the powerful one. Plug in images and refer to them in the prompt as “picture one”, “picture two”. For a video reference use the Load Video (Upload) node, not the Comfy Essentials Load Video node; for audio, Load Audio (Upload), and set the duration to match the track. His examples: a woman from one image getting into a car from another; replacing a green-screen background with a bamboo forest; a 14-second music video of a singer on a cliff.

What H3 is good at (his demos): cartoons and known characters, TV shows and films from text alone, unusual actions, claymation and anime styles, multi-character consistency, ads from product references, short drama with dialogue, animating a website screenshot, gameplay footage from a game screenshot, motion transfer from a reference video, and adding, removing or cloning characters in existing footage. He calls it “basically like an open-source Gemini Omni.”

MiniMax H3: speed, VRAM and quality add-ons

Insert each of these after the Load Diffusion Model node, feeding Basic Guider and Basic Scheduler:

  • Patch SageAttention (from KJNodes), set to auto: 20–30% faster. You must first install a prebuilt SageAttention wheel matching your PyTorch, CUDA and Python versions (check them from ComfyUI’s python_embeded folder; his were torch 2.6, CUDA 12.4, Python 3.11) with python.exe -m pip install <wheel>, then git clone KJNodes into custom_nodes and install its requirements.
  • EasyCache (built into ComfyUI): another 5–10%, stackable.
  • Spectrum Apply MiniMax H3 (xmarre/ComfyUI-Spectrum-MiniMax-H3, updated with git pull to a version with less quality loss). He pairs it with the res_multistep sampler and simple scheduler. It does not work with the ERSD sampler.
  • Live preview: update KJNodes, download the tiny MiniMax-H3-TAE (under 10 MB, from Kijai/MiniMax-H3-TAE) into models/vae_approx (not vae), add Model Preview Override, and set preview frames to cover the clip (124 for 5 s at 24 fps). Cancel a bad generation halfway.
  • Turbo LoRAs (4–8 steps): larryvrh’s (an early, robust one), LightX2V’s 4- and 8-step ComfyUI versions, joyfox’s (for the 40 GB BF16 base, better on action), and Kijai’s resized LightX2V at 315 MB instead of about 2 GB. Kijai suggests 4 steps at 0.75 strength with the ERSD or SA solver sampler and a beta (or beta57) scheduler; he uses 6 steps. He personally prefers not to use turbo LoRAs.
  • Smaller weights for about 8 GB: Kijai’s experimental W4A8 diffusion models (about 12 GB each) plus his INT8 video VAE (3 GB against the official 5.2 GB), with SageAttention and Spectrum, and 16 steps instead of 20.
  • GGUFs exist (Abiray Q3 15.6 GB; Unsloth down to 6.7 GB), but he says ComfyUI no longer recommends GGUFs here: quantized ComfyUI-optimized builds give better quality at similar VRAM with the right settings.
  • Realism LoRA: fal’s MiniMax-H3-Realism-People-LoRA (131 MB) for casual, amateur-looking footage. Load it at about 0.8 and include the trigger word “realism”.
  • One-workflow alternative: ComfyUI-MiniMax-H3-Easy (by NKXX188) puts text-, image-, first/last-frame- and reference-to-video in one clean graph, with plain resolution choices (360/480/1080p) and @ mentions of attached media in the prompt.
  • Train LoRAs with ostris ai-toolkit, which added H3 support.
  • Below 12 GB entirely: Wan2GP, whose author says H3 runs on 5 GB at 480p.

Odd uses: audio-only generation (described above) and a community pseudo-image workflow that renders a single frame. He would not use it as an image model; Krea 2 and Ideogram are “much better.”

LTX 2.5 (Lightricks)

What changed from LTX 2.3 (vendor claims he relays): “diffusion fidelity rendering,” which spends more compute on complex parts of a scene; multi-shot generations with consistent characters; cleaner motion and better prompt understanding; up to 4K and 50 fps; up to 20 s “safely,” longer with enough VRAM; works with existing LTX 2 LoRAs.

Install:

  1. Update ComfyUI. Templates → “LTX 2.5”. Skip the crown-icon Pro workflows (paid, cloud); use the free text-to-video, image-to-video and first/last-frame ones.
  2. Download from huggingface.co/Lightricks/LTX-2.5:
    • Distilled diffusion model (4–6 steps), not dev (20–30 steps, meant for LoRA training). BF16 42 GB, INT8 22 GB (fits about 16 GB with optimization), FP4 19 GB on supported GPUs. → models/diffusion_models
    • Spatial upscaler (about 1 GB) → models/latent_upscale_models
    • Gemma 4 text encoder (16 GB compressed version) → models/text_encoders
    • Audio VAE and a video VAE → models/vae
  3. Turn the prompt enhancer off; it costs compute and he does not find it necessary.

Speed on his machine: 20 s for 5 s text-to-video, 20 s for image-to-video (a cat rock band with singing), 30 s for first/last frame (no upscaler stage in that workflow).

LoRAs: Load LoRA after the diffusion model, connected to both model inputs downstream; he ran a retro-’90s anime LoRA at 0.9 with its “Akira” trigger word.

12 GB GPUs: Abiray’s LTX 2.5 GGUFs, smallest Q3 small at 12.6 GB, into models/unet, loaded with Unet Loader (GGUF) (shift-drag the existing connections onto the new node).

Extras: the LTX upscaler can sharpen MiniMax H3 output (community workflow; “hit or miss,” better on slow shots), and LTX Director (WhatDreamsCost/WhatDreamsCost-ComfyUI), a mini video editor inside ComfyUI that chains text-, image- and first/last-frame clips with custom audio.

Models built on MiniMax H3 (added 2026-11-09)

Sources: raw/Kimi_K3.1_weights_OUT_open_source_computer_use_SOTA_image_model_science_model_Pika_H3_-_AI_NEWS.md (Matthew Berman) and raw/AI_News_-_New_ChatGPT_Google_Models_14_Updates_Worth_Knowing.md (Matt Wolfe). Figures are vendor claims as the two creators read them.

  • Pika H3 Flex (Pika Labs). Berman says “this is trained on MiniMax H3, but it is faster and better overall”. Both creators say it debuted as the top open-weight model on Design Arena’s text-to-video board.
    • Output: up to 4K, exportable audio stems, faster than real-time generation, and a 15-second clip with a soundtrack in as little as 7 seconds.
    • Price: API pricing starts at one cent per second of video (Berman).
    • Weights: on Hugging Face at 220 GB (Wolfe), so running it locally needs a high-end GPU.
  • PH3 Image (Pruna AI) turns MiniMax H3 into an open image generation and editing model.
    • Speed: 8 steps instead of 50, and a 4-megapixel image in 2.6 seconds instead of 12.4.
    • Size: 7x smaller than H3, about 55 GB at BF16. An 8-megapixel image needs about 56 GB of VRAM at 16-bit and 21 GB at 4-bit, so the 4-bit model fits some consumer GPUs (Wolfe).
    • Strongest at photorealism, documentary and editorial photography, film stills, and graphic styles with short quoted text (Berman, reading Pruna’s post).

Try It

  1. Pick by job. Short social clips, B-roll drafts and anything where iteration speed matters: start with LTX 2.5. Ads with dialogue, music-synced edits, character consistency from references: H3. Fight scenes or 30-second single takes: hosted Seedance 2.5. ^[inferred] This is a reading of his tests, not a benchmark.
  2. Install H3 with the speed stack first (SageAttention + EasyCache + Spectrum) and the live-preview node, so bad generations die early.
  3. On 8–12 GB, go straight to Kijai’s W4A8 H3 builds or the LTX 2.5 Q3 GGUF instead of the pruned 21–22 GB models.
  4. Write H3 prompts to MiniMax’s three-part format (description, soundscape, background music), or have ChatGPT or Claude draft them from the guide.
  5. Check H3’s licence for your country and revenue before client work. In the US, UK, EU or Korea you need the separate licence; above $20M annual revenue you need an agreement.
  6. Let an agent do the install if the wheel-matching steps look daunting; the wave article reports creators handing the GitHub links to Codex.

Implementation

Tool/Service: ComfyUI; MiniMax H3 open weights (plus the Hailuo hosted app and API); LTX 2.5 (Lightricks); KJNodes, SageAttention, Spectrum, ComfyUI-GGUF, ComfyUI-MiniMax-H3-Easy, LTX Director; Wan2GP for very low VRAM. Setup: ComfyUI Windows portable, updated; models into diffusion_models, text_encoders, vae, vae_approx, latent_upscale_models, loras, unet; custom nodes cloned into custom_nodes and their requirements installed with the embedded Python. Cost: free locally. Hosted H3: about 5–6 per 10-second clip. Integration notes: H3’s text encoder can live in system RAM. SageAttention needs a wheel built for your exact PyTorch, CUDA and Python versions. Spectrum is incompatible with the ERSD sampler. LTX first/last-frame has no upscaler stage, so it is slower than the other LTX workflows.

From the ComfyUI Windows portable python_embeded folder, then from ComfyUI\custom_nodes:

python.exe -c "import torch; print(torch.__version__, torch.version.cuda)"
python.exe --version
python.exe -m pip install C:\path\to\sageattention-<matching>.whl
git clone <KJNodes repo URL>
git clone <ComfyUI-Spectrum-MiniMax-H3 repo URL>

Open Questions

  • H3’s licence text is paraphrased by the presenter; the $20M cap and the excluded-territory list should be confirmed on the model card before use.
  • The 12 GB and 5 GB H3 floors are user and Wan2GP-author reports; he ran everything on a 16 GB laptop GPU.
  • The tests are single runs with one creator judging; Seedance and H3 were compared on hosted versions, and the open H3 weights may differ from Hailuo’s 2K service.
  • LTX 2.5’s licence is not discussed in the video.
  • The reference-to-video model’s caption name (“Ref2VA”) and the other model’s (“FL2VA”) are as captioned; check the exact filenames on the download page.