Hardware requirements, as published.
The VRAM and GPU figures each developer publishes for running its own weights, and the setup each figure assumes. They're the developers' figures, not Windpaint's: Windpaint runs z-image-turbo and flux-2-klein-4b on its own GPUs, so calling the API needs none of this.
- Models
- 19 in the catalogue
- VRAM figure
- Stated for 7 of 19
Image models: VRAM requirements.
The developer's figure, where it states one, and the setup it assumes. Each row links to the model's self-hosting section.
| Model | Published VRAM | GPUs named | Applies to | Source |
|---|---|---|---|---|
| Z-Image Turbo | Fits within 16 GB | None named | 16GB-VRAM consumer devices, bf16 weights. CPU offload shown for memory-constrained devices. | |
| FLUX.2 [klein] 4B | About 13 GB | RTX 3090, RTX 4070 and above | Model card figure. BFL's GitHub README says about 8 GB instead. | |
| Z-Image | Not stated | None named | No VRAM figure on the card. The example loads bf16 weights. | |
| Qwen-Image-2512 | Not stated | None named | No VRAM figure on the card. bf16 on CUDA; the repo ships a multi-GPU server. | |
| FLUX.1 [schnell] | Not stated | None named | No VRAM figure on the card. Its example uses CPU offload to save VRAM. | |
| HiDream-O1-Image | Not stated | A CUDA-capable GPU | No VRAM figure on the card. flash-attn recommended; avoid PyTorch 2.9.x. | |
| Stable Diffusion XL | Not stated | None named | No VRAM figure on the card. fp16 weights; CPU offload if VRAM-limited. | |
| Qwen-Image-Edit-2511 | Not stated | None named | No VRAM figure on the card. bf16 on CUDA. | |
| FLUX.2 [klein] 9B | About 29 GB | RTX 4090 and above | Model card figure for the distilled 9B. FP8 and NVFP4 versions are also published. | |
| Qwen-Image-2.1 | Not stated | None named | No VRAM figure on the card or README. bf16 weights; CPU offload shown for limited memory. | |
| MiniMax H3-Image | Not stated | H200, RTX PRO 6000 Blackwell, MI350X (supported list) | No VRAM figure on the card. BF16 DiT 66.28 GB plus text encoder 66.7 GB. |
Video models: VRAM requirements.
The developer's figure, where it states one, and the setup it assumes. Each row links to the model's self-hosting section.
| Model | Published VRAM | GPUs named | Applies to | Source |
|---|---|---|---|---|
| Wan 2.2 image-to-video | At least 80 GB | None named | Single-GPU 720p command with model offload and dtype conversion. Multi-GPU example uses 8 GPUs. | |
| Wan 2.2 text-to-video | At least 80 GB | None named | Single-GPU 720p command with model offload and dtype conversion. Multi-GPU example uses 8 GPUs. | |
| Wan 2.2 TI2V 5B | At least 24 GB | RTX 4090 | Single-GPU 720p command with model offload, dtype conversion and T5 on CPU. Drop those options at 80 GB or more. | |
| LongCat-Video | Not stated | None named | No VRAM figure in the README. Single- and multi-GPU commands; FlashAttention-2 by default. | |
| HunyuanVideo 1.5 | 14 GB minimum | NVIDIA GPU with CUDA | Measured with model offloading on. Linux, Python 3.10 or newer. | |
| LTX-2.5 | Not stated | None named | No inference VRAM figure. The README suggests FP8 casting and CPU or disk offload for limited GPU memory. | |
| LTX-2.3 | Not stated | None named | No inference VRAM figure. Tested with Python 3.12+, CUDA above 12.7; FP8 casting and offload suggested for limited memory. | |
| MiniMax H3 | Not stated | 4 GPUs in the SGLang example | No VRAM figure on the card. BF16 checkpoints with a Qwen3-VL-32B text encoder. |
How this is put together.
From each developer's own model card, README or docs.
A figure appears only where the developer states one. Each assumes the developer's own setup, such as model offloading or dtype conversion, which is listed beside it. Where sources from the same developer disagree, the row says so.
Some official READMEs also name community tools that report lower figures. Those aren't the developer's numbers, so they're left out of the table and listed on each model page, marked as community tools. For licences, see commercial use.
Questions.
Which models have a published VRAM figure?
Z-Image Turbo (fits within 16 GB), FLUX.2 [klein] 4B (about 13 GB), Wan 2.2 image-to-video (at least 80 GB), Wan 2.2 text-to-video (at least 80 GB), Wan 2.2 TI2V 5B (at least 24 GB), HunyuanVideo 1.5 (14 GB minimum) and FLUX.2 [klein] 9B (about 29 GB). The others don't state one on their card or in their README.
Do I need a GPU to use these models on Windpaint?
No. Windpaint runs z-image-turbo and flux-2-klein-4b on its own GPUs and you call them over the API. The figures here are for running the weights yourself. The other models aren't served on Windpaint yet.
Why do some READMEs mention 4GB of VRAM?
The Z-Image README says stable-diffusion.cpp, a third-party engine, can run Z-Image in as little as 4GB, and the Qwen-Image README says DiffSynth-Studio can run Qwen-Image within 4GB with layer-by-layer offload. Those are community tools, not the developer's own releases, so they aren't in the table. Each model page lists them, marked as community tools.