Model Directory
Open-source image and video generation models — what each one needs to run, and whether it's practical on a single consumer GPU. Includes models we run in FlixML and models we've evaluated and passed on, with the real reason why.
SDXL
Text → ImageThe general-purpose open image checkpoint most people start with. Fast, well-supported, runs comfortably on consumer VRAM.
~8 GB
FLUX.2
Text → ImageHigher-fidelity image generation than SDXL. The distilled Klein 4B runs on 8 GB in seconds; full dev needs a 24 GB card.
8 GB for Klein 4B; 24 GB+ for dev
Wan 2.2
Text/Image → VideoImage-to-video and text-to-video that fits a single consumer GPU once quantized.
~12 GB (fp8)
InfiniteTalk
Image/Video → Video (lip-sync)Audio-driven lip-sync, layered on a Wan 2.1 backbone. Makes a still or a clip talk.
~12 GB (GGUF-quantized)
Qwen-Image-Edit-2511
Image → Image (instruction edit)Instruction-driven image editing — re-angle a shot or change a pose from plain English, holding identity and lighting.
~12 GB
MiniMax-H3
Text/Image → Video + AudioOne of the most-pulled open video models on Hugging Face — natively generates synchronized audio and video. We evaluated it and it doesn't fit our 12 GB target.
~15.5 GB UNet + ~14.6 GB text encoder · evaluated, not in FlixML