Description
The wall in front of local AI video generation was never the interface — it was VRAM. The official repos ask for 24GB up front, so a 3060 12G or an older 2080 cannot even load the weights, and you end up renting cloud compute by the minute, where iterating on one prompt ten times costs more than the result is worth. Wan2GP (WanGP) drops that floor to 6GB of VRAM, putting GTX 10-series and RTX 20-series cards back in the game.
It is not a launcher for one model either. Video, image and audio models all live behind the same browser UI: Wan 2.1/2.2, LTX-2, Hunyuan Video, Qwen Image, Flux, Z-Image, plus a long bench of TTS and music models. Switching models is a dropdown, one environment covers all of them, and you generate from a browser tab instead of memorizing command-line flags.
Runs on modest GPUs: selected models start at 6GB of VRAM, GTX 10XX and RTX 20XX cards are supported, and AMD RDNA 2 through RDNA 4 hardware works too — no GPU upgrade required just to press generate.
One app, every model: Wan 2.1/2.2, LTX-2, Hunyuan Video, MiniMax and Kandinsky for video; Qwen Image, Flux, Z-Image and HiDream for stills; Qwen3 TTS, Ace Step, Index TTS2 and voice cloning for audio.
Hardware-aware downloads: int8, fp8, gguf, NV FP4 and Nunchaku checkpoints are all understood, and the app picks the build that matches your card instead of making you cross-reference a compatibility table.
One-click install scripts: double-click install.bat on Windows or run install.sh on Linux, and the script also sets up acceleration kernels like Triton, Sage Attention and Flash Attention. Multiple sandboxed environments mean a broken experiment is one switch away from the setup that worked.
Generation queue: stack up a batch of video, image and audio jobs, walk away, and collect everything later. A headless mode runs the same batches straight from the command line.
LoRAs and finetunes: every model accepts LoRAs, existing LoRA folders from another app can be reused in place, and finetunes pulled from Hugging Face or CivitAI drop straight in.
Input preparation tools: a mask editor, background remover, pose/depth/flow extractors, speaker diarization and vocal/background separation are built in, so control-video and image-to-video inputs never need a second program.
Output postprocessing: RIFE interpolation and FlashVSR upscaling on the video side, MMAudio soundtracks, SeedVC voice replacement, and remuxing any soundtrack into a finished clip.
Workspaces: generations are grouped per project, collections and selections survive a restart, media moves between workspaces without duplicating files on disk, and a selection can be zipped for download.
Follow a job from anywhere: start a generation at your desk and check it from a phone or another machine on the same server — galleries, progress and the queue stay in sync, and you can add jobs from either end.
Per-model prompt enhancer: prompts are rewritten in the syntax each model actually responds to, and with a Qwen3-based enhancer the start frame, end frame and reference images feed into the rewrite as well.
It is not a launcher for one model either. Video, image and audio models all live behind the same browser UI: Wan 2.1/2.2, LTX-2, Hunyuan Video, Qwen Image, Flux, Z-Image, plus a long bench of TTS and music models. Switching models is a dropdown, one environment covers all of them, and you generate from a browser tab instead of memorizing command-line flags.
Features
Runs on modest GPUs: selected models start at 6GB of VRAM, GTX 10XX and RTX 20XX cards are supported, and AMD RDNA 2 through RDNA 4 hardware works too — no GPU upgrade required just to press generate.
One app, every model: Wan 2.1/2.2, LTX-2, Hunyuan Video, MiniMax and Kandinsky for video; Qwen Image, Flux, Z-Image and HiDream for stills; Qwen3 TTS, Ace Step, Index TTS2 and voice cloning for audio.
Hardware-aware downloads: int8, fp8, gguf, NV FP4 and Nunchaku checkpoints are all understood, and the app picks the build that matches your card instead of making you cross-reference a compatibility table.
One-click install scripts: double-click install.bat on Windows or run install.sh on Linux, and the script also sets up acceleration kernels like Triton, Sage Attention and Flash Attention. Multiple sandboxed environments mean a broken experiment is one switch away from the setup that worked.
Generation queue: stack up a batch of video, image and audio jobs, walk away, and collect everything later. A headless mode runs the same batches straight from the command line.
LoRAs and finetunes: every model accepts LoRAs, existing LoRA folders from another app can be reused in place, and finetunes pulled from Hugging Face or CivitAI drop straight in.
Input preparation tools: a mask editor, background remover, pose/depth/flow extractors, speaker diarization and vocal/background separation are built in, so control-video and image-to-video inputs never need a second program.
Output postprocessing: RIFE interpolation and FlashVSR upscaling on the video side, MMAudio soundtracks, SeedVC voice replacement, and remuxing any soundtrack into a finished clip.
Workspaces: generations are grouped per project, collections and selections survive a restart, media moves between workspaces without duplicating files on disk, and a selection can be zipped for download.
Follow a job from anywhere: start a generation at your desk and check it from a phone or another machine on the same server — galleries, progress and the queue stay in sync, and you can add jobs from either end.
Per-model prompt enhancer: prompts are rewritten in the syntax each model actually responds to, and with a Qwen3-based enhancer the start frame, end frame and reference images feed into the rewrite as well.
