AI short film production: which GPU for each step?

Short answer: Most stages of an AI short film run on a 24–32 GB card: storyboard images, Wan 2.2 TI2V-5B, voice, lip sync, music, upscaling and interpolation. The 14B and larger video models are where you need 48 GB, 80 GB or more, or heavy offloading: Wan 2.2 A14B, MiniMax H3 and LTX-2.5. Rent each stage on the card it needs and destroy the instance between stages, because billing is per minute. Check each tool’s licence before commercial use: some are non-commercial or carry extra conditions.

The pipeline at a glance

VRAM figures and licence wording are from each project’s own README or model card, read 2026-11-04. “None published” means the project gives no GB figure. The Deploy page shows which cards you can start now.

Stage Tool VRAM guidance (published) Card class to start with Licence note
Storyboard images FLUX.1, SDXL See how much VRAM for FLUX and SDXL 24 GB FLUX.1 dev needs its licence accepted; FLUX.1 schnell is Apache 2.0
Video Wan 2.2 TI2V-5B ComfyUI: “should fit well on 8GB vram with the ComfyUI native offloading” 24–32 GB Apache 2.0
Video Wan 2.2 T2V/I2V-A14B README: “at least 80GB VRAM” for its single-GPU command 48–80 GB Apache 2.0
Video + audio MiniMax H3 diffusers: int8 + streaming on 24–32 GB; one 80 GB card with offload 32 GB with plenty of host RAM, or 80–96 GB Community licence, see below
Video + audio LTX-2.5 (22B) No GB figure; README: “consider --quantization fp8-cast --offload {cpu, disk}” under memory limits 48–96 GB Paid licence for entities with “annual revenues of at least $10,000,000”
Video HunyuanVideo 1.5 (8.3B) “Minimum GPU Memory: 14 GB (with model offloading enabled)” 24 GB Community licence, see below
Character consistency musubi-tuner LoRA “12GB or more recommended for image training, 24GB or more for video training”; “Main Memory: 64GB or more recommended” 24–48 GB Apache 2.0 (Wan code)
Character consistency ai-toolkit LoRA Example config train_lora_wan22_14b_24gb.yaml; “Training on and generating video is very VRAM intensive.” 24–48 GB MIT
Character consistency Wan-Animate-2 “tuned for 8× A800 GPUs”; 480P “tested … on 2× A800 GPUs” several 80 GB cards Apache 2.0
Voice (TTS) Fun-CosyVoice 3.0 (0.5B) None published 24 GB Apache 2.0
Voice (TTS) IndexTTS-2.5 None published (an RTF table on RTX 4090) 24 GB bilibili Model Use License (custom)
Lip sync LatentSync “Minimum VRAM for inference: 8 GB with LatentSync 1.5; 18 GB with LatentSync 1.6” 24 GB Apache 2.0
Lip sync MuseTalk Tested on a 4 GB laptop GPU: “generating an 8-second video takes approximately 5 minutes” 24 GB MIT (code)
Lip sync / dubbing InfiniteTalk No GB figure; low-VRAM mode: “set --num_persistent_param_in_dit 0” 24–48 GB Apache 2.0
Music ACE-Step 1.5 “XL … require ≥12GB VRAM (with offload + quantization) or ≥20GB (without offload)”; “under 10s on RTX 3090” per song 24 GB MIT
Music MiniMax Music 3 “The full precision fits under 24GB of VRAM” 24 GB Community licence: “MiniMax-Music3” UI notice; over $20M revenue needs authorization
Music YuE2 “24 GB VRAM” 24 GB Weights CC BY-NC 4.0: non-commercial
Upscale SeedVR2 “High VRAM (24GB+): Use FP16 models”; “Moderate VRAM (12-16GB): Use FP8 models with BlockSwap or VAE tiling” 24–32 GB Apache 2.0
Upscale Real-ESRGAN None published 24 GB BSD-3-Clause
Frame interpolation RIFE (Practical-RIFE) None published; “4.24+ is quite suitable for post-processing of some diffusion model generated videos” 24 GB MIT
Frame interpolation GIMM-VFI “[4K interpolation] … Memory-Usage: 10922MiB” on V100 24 GB S-Lab License 1.0: check before commercial use

Two tools you may still see recommended: Chatterbox is retired (“As of August 2026, Chatterbox is retired research and is no longer maintained.”), and Microsoft has “removed the VibeVoice-TTS code from this repository”.

Licences to read before commercial use

  • MiniMax H3: MiniMax H3 Community License Agreement, not an open-source licence. Read the LICENSE before use.
  • HunyuanVideo 1.5: Tencent Hunyuan Community License, not an open-source licence. Read the LICENSE before use.
  • YuE2: the weights are CC BY-NC 4.0, so companies must contact the authors before commercial use.
  • GIMM-VFI: S-Lab License 1.0. Read it before using GIMM-VFI in paid work.
  • LTX-2.5: free below the revenue threshold. “Entities with annual revenues of at least $10,000,000 … are required to obtain a paid license.”

For commercial work with the simplest licence terms, use the Apache 2.0 and MIT tools in the table: Wan 2.1 and 2.2, SeedVR2, RIFE, ACE-Step 1.5, LatentSync and CosyVoice.

Run each stage on the card it needs

A short film does not need a large card running the whole time. A typical split:

  1. 24 GB card (RTX 4090): storyboard frames, Wan 2.2 TI2V-5B drafts, voice, music.
  2. 48 GB or 80 GB card (RTX 4090 48G, A100 80G): final shots on a 14B model, or LoRA training.
  3. Back to a 24–32 GB card: lip sync, upscaling, frame interpolation.

Copy your outputs off the instance (rsync or scp) before you destroy it, because data on the instance is deleted. At 2026-11-05 07:55 (UTC+8), two hours on an RTX 4090 cost $0.78, and one hour of final shots on an RTX 4090 48G costs $0.69. Keeping an A100 80G for the same three hours would cost $4.77. Billing is per minute, rounded up, with a 1-minute minimum, and it stops when you destroy the instance. Current rates are on /pricing.

If you only need finished clips from a hosted model, CloudGPU’s model API also has video models, including wan-3.0. You create a job with POST /v1/videos (API docs).

Related: Which GPU for Wan video generation · Which GPU for MiniMax H3 · How to run Wan 2.2 on a cloud GPU · Cloud GPU for AI film and video production

Frequently asked questions

What GPU do I need to make an AI short film?

Most stages run on a 24–32 GB card: storyboard images, Wan 2.2 TI2V-5B, voice, lip sync, music, upscaling and interpolation. The 14B and larger video models, such as Wan 2.2 A14B, MiniMax H3 and LTX-2.5, need 48 GB, 80 GB or more, or heavy offloading to host RAM.

Which open video models can I use commercially?

Wan 2.1 and Wan 2.2 are Apache 2.0. MiniMax H3 and HunyuanVideo 1.5 use their own community licences, so read them before commercial use; LTX-2.5 needs a paid licence for entities with at least $10,000,000 in annual revenue (read 2026-11-04).

How do I keep a character consistent across shots?

Train a LoRA on the character with musubi-tuner or ai-toolkit, or use a reference-image model such as Wan-Animate-2 or MiniMax H3 Ref2VA. musubi-tuner recommends 24 GB or more of VRAM for video training and 64 GB or more of main memory (read 2026-11-04).

Which tools add voice, lip sync and music?

Fun-CosyVoice 3.0 and IndexTTS-2.5 for voice; LatentSync, MuseTalk and InfiniteTalk for lip sync; ACE-Step 1.5 and MiniMax Music 3 for music. YuE2 weights are CC BY-NC 4.0, so they are not for commercial use without permission (read 2026-11-04).

How do I upscale and smooth AI video?

SeedVR2 or Real-ESRGAN for upscaling, and RIFE for frame interpolation. SeedVR2’s ComfyUI node suggests FP16 models on 24 GB or more and FP8 models with BlockSwap or VAE tiling on 12–16 GB (read 2026-11-04).

Sources (read 2026-11-04, Asia/Shanghai)

The Deploy page lists the cards you can start. Prices are on https://cloudgpu.app/pricing.

Last updated: 2026-11-04 (Asia/Shanghai).