Wan2GP vs ComfyUI: Which to Use for Local AI Video on a Small GPU

Quick answer

Wan2GP vs ComfyUI comes down to fixed tabs versus a graph you build yourself. Wan2GP, which calls itself WanGP, is one developer's Gradio app, made by DeepBeepMeep "for the GPU poor". You pick a model from a dropdown, fill in a form and press generate. It downloads the model files for you and offloads to system RAM by default. ComfyUI is a node editor: every model, loader, sampler and decoder is a node you can rewire, and custom nodes extend it.

Both run Wan 2.1 and 2.2, LTX, HunyuanVideo 1.5 and MiniMax H3. Neither project has published a like-for-like speed comparison with the other, and we have not run Wan2GP. Every speed or VRAM figure below is labelled with whose it is.

We read both projects' READMEs, documentation, licence files and source on 2026-10-10. File sizes are from the Hugging Face API, in decimal gigabytes.

What Wan2GP is

Wan2GP started on 2 March 2025 as "Wan2.1GP", a low-VRAM front end for Wan 2.1, per its changelog. The repository is still deepbeepmeep/Wan2GP; the app now calls itself WanGP and covers far more than Wan. Its README lists, today:

The interface is a Gradio web page, served at http://localhost:7860 by default. It adds a generation queue, galleries, a prompt enhancer, an offline assistant called Deepy, a headless mode (python wgp.py --process my_queue.zip) and a plugin manager.

ComfyUI's README lists Wan 2.1 and 2.2, LTX-Video 2 and 2.3, HunyuanVideo 1.5, CogVideoX, Mochi and others for video, and MiniMax H3 and LTX-AV for audio with video. It calls that list representative, not complete.

Does Wan2GP support MiniMax H3?

Yes. The changelog entry dated 5 August 2026, WanGP v12.42, added MiniMax H3 FL2VA and Ref2VA, each in the full 33B and the pruned 20B version. Later entries added Turbo and PDD accelerator LoRAs, sliding windows for longer clips, ControlNet-Union 2.0, outpainting and an H3 voice-cloning preset.

Wan2GP's own VRAM claims for H3, none of them with stated test hardware:

• 5 August 2026: 5-6 GB of VRAM for 5 seconds (124 frames), and 8-9 GB for 15 seconds, at 832×480.
• 7 October 2026, v17.17: 15 seconds at 1080p down from 25 GB of VRAM to 11 GB, after a rewrite of its memory library.

These say nothing about system RAM, which is where H3 bites. Our own measured run was in ComfyUI, not Wan2GP: on an RTX 3060 the card peaked at 11,649 of 12,288 MiB and system memory at 43,587 MiB, for a 1344×768, 124-frame clip in about 43 minutes. The full test card is on our MiniMax H3 RTX 3060 page. That page also lists one owner's Wan2GP report; it used a different task, resolution and length, so the two cannot be compared.

How Wan2GP saves memory: the profiles

Wan2GP's offloading comes from its own library, mmgp, "Memory Management for the GPU Poor". Since v17.17 it lives in the Wan2GP repository rather than a separate one. You pick a profile with --profile or in the settings. Profile 4 is the default, both in the CLI documentation and in wgp.py.

Two more memory controls sit in Config → RAM/VRAM Management: a VRAM allocator choice (the MMGP Optimized VRAM Allocator is the default) and a VRAM Preload setting. The v17.17 notes say several of its savings depend on SageAttention 2; our SageAttention page covers that library on the ComfyUI side.

ComfyUI offloads by default too. Its README says it can run large models on as little as 4 GB of VRAM and 8 GB of RAM through weight streaming; that is Comfy Org's claim, with no hardware named. In comfy/cli_args.py, "dynamic VRAM" is on unless you pass --disable-dynamic-vram, --highvram, --gpu-only, --novram or --cpu, and the help text for --lowvram says it does nothing while dynamic VRAM is on. When ComfyUI runs out anyway, see our out-of-memory page.

Installing Wan2GP

The README offers four routes:

1. Its own scripts. scripts\install.bat on Windows or scripts/install.sh on Linux and macOS, with an auto or manual mode. run, update and manage scripts sit beside it; the README says the installer also sets up acceleration kernels such as Triton and SageAttention.
2. Pinokio. A one-click install through the Pinokio app. The README recommends the community scripts wan2gp or wan2gp-amd by Morpheus over the official Pinokio install, and says Pinokio users should update through Pinokio.
3. Wan2GP Desktop. A third-party launcher by GKArtist that installs, updates and runs it.
4. By hand. The README's commands for RTX 20 to RTX 50 cards:

git clone https://github.com/deepbeepmeep/Wan2GP.git
cd Wan2GP
conda create -n wan2gp python=3.11.14
conda activate wan2gp
pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt
python wgp.py

GTX 10-series cards get a separate recipe with Python 3.10.9 and PyTorch 2.7.1. A Docker script, run-docker-cuda-deb.sh, covers Debian and Ubuntu. To update a manual install:

git pull
conda activate wan2gp
pip install -r requirements.txt

The README also warns that only the official GitHub repository and the wangp.ai and wan2gp.ai sites are the project's; other services using the name are not. For ComfyUI's three install routes, portable, Desktop and git, see our ComfyUI download page.

Where each keeps its models

Wan2GP downloads for you. Pick a model and it fetches the files it needs into its ckpts/ folder; the README says it picks the files suited to your hardware. For H3, its default FL2VA Pruned 20B definition points at the DeepBeepMeep/MiniMax-H3 repository on Hugging Face:

ComfyUI downloads nothing on its own. Its README says the core downloads only what you ask for; you place files in ComfyUI/models/ yourself, or follow a template's download links. Comfy-Org's matching file, minimax_h3_fl2va_pruned_int8_convrot.safetensors, is 20,970,379,616 bytes (20.97 GB).

The two H3 diffusion files are not the same file: different names, sizes and SHA-256 hashes. Wan2GP's changelog calls its pruned H3 checkpoints ComfyUI-compatible so that H3 LoRAs work in both apps, which is not the same as byte-identical. Two files are identical in both repositories, by SHA-256: the NVFP4 text encoder qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (15.69 GB) and the INT8 video VAE minimax_h3_video_vae_int8_convrot.safetensors (2.81 GB).

Sharing a folder. Wan2GP v8.994, in its 6 October 2025 changelog entry, added extra checkpoint folders set in the Config tab, with reusing files already in a ComfyUI folder named as the reason. In the code, the setting is checkpoints_paths in wgp_config.json, defaulting to ckpts and .. Whether Wan2GP would use ComfyUI's differently named H3 files in place of its own is not documented, and we have not tested it. For LoRAs, --loras PATH moves the whole LoRA root and --lora-config lora_paths.json points single model folders elsewhere, per its LoRA guide. ComfyUI's own route is extra_model_paths.yaml.

Side by side

Update pace is high on both sides. Wan2GP's README carries update notes dated 6, 16, 24 and 29 September and 7 October 2026, and its repository took at least 100 commits between 10 September and 10 October.

What nobody has published

• The same video model, settings and clip length run in Wan2GP and in ComfyUI on one machine, with peak VRAM, peak system RAM and time for each.
• A measured Wan2GP profile 4 versus profile 5 run on a 12 GB card for MiniMax H3, with the hardware named.

If you run either, record what our test card records. For the broader picture on a 12 GB card, see our RTX 3060 12GB page.

To check a GPU and system RAM size against MiniMax H3, use the system requirements checker. Its only preset is MiniMax H3; Wan, LTX and HunyuanVideo are not presets yet.

Licence and downloads

Wan2GP is not under a standard open-source licence. Its LICENSE.txt is the WanGP Community License 2.0, and GitHub shows it as "Other". In summary: free use is allowed, including inside a company and for client work. You may sell what you make, crediting WanGP only when you sell or license an output directly. Selling WanGP itself, white-labelling it, or offering paid hosted, API or SaaS access needs a separate written licence. Third-party code and model weights keep their own licences. The bundled mmgp library is under the same licence.

ComfyUI is under GPL-3.0.

The model weights are a third layer. MiniMax H3 is under the MiniMax H3 Community License, and Comfy-Org's repository marks it license: other. The DeepBeepMeep/MiniMax-H3 repository sets no licence field at all. Our licence map covers which H3 file falls under which terms.

Territory. The MiniMax licence grants use, modification, distribution and display only outside the EU, the UK, the Republic of Korea and the United States, and applies the same limit to outputs. Read the licence map before downloading any H3 file named here or reusing what it generates.

We do not host any of these files. Download them from the repositories named below.

GenVidKit is an independent guide. It is not affiliated with DeepBeepMeep or the WanGP project, Comfy Org, Pinokio, Alibaba, the Wan team, MiniMax or Hugging Face.

Sources

All read on 2026-10-10.

• Wan2GP README — supported models, the September and October 2026 update notes including v17.17, install routes, Pinokio recommendation, update commands.
• Wan2GP changelog — the 5 August 2026 H3 entry and VRAM claims, extra checkpoint folders (v8.994), the first release.
• Wan2GP CLI documentation — the profiles, the default profile, LoRA path flags, the server port.
• Wan2GP LoRA guide, finetunes guide and getting started guide — LoRA folders, ckpts/, first launch.
• Wan2GP source: wgp.py, shared/utils/files_locator.py, mmgp/ and defaults/minimax_h3_fl2va_pruned.json — default profile, checkpoint paths, mmgp, H3 file URLs.
• Wan2GP LICENSE.txt — the WanGP Community License 2.0.
• ComfyUI README — features, supported models, offline downloads, release cycle, ComfyUI-Manager.
• ComfyUI comfy/cli_args.py — dynamic VRAM and the VRAM flags.
• ComfyUI releases and its LICENSE — release dates, GPL-3.0.
• DeepBeepMeep/MiniMax-H3 and Comfy-Org/MiniMax-H3 — file sizes, SHA-256 hashes, licence fields.