Best open-source image and video generation models

Most open image and video models now run on a single consumer GPU, though the largest need workstation-class memory, and the license matters as much as quality: several of the best are free only for personal use, for companies below a revenue limit, or outside certain regions. Memory figures include the text encoder and decoder that these pipelines also need.

  • List updated November 11, 2026
  1. FLUX 3 Image Klein

    Black Forest Labs' 5.6B model generates and edits images and combines up to 10 reference images. An INT8 build runs in 8 GB of VRAM. Commercial use is free only below US$5 million of annual revenue, and the download is gated.

    • Black Forest Labs
    • 5.6B
    • FLUX Kommunity (Conditional)
    • from 8 GB
  2. Qwen-Image-2.1

    Alibaba's 7B model generates and edits images and can output real transparent PNGs from up to 10 references. The license is research-only, so commercial use needs a separate license.

    • Alibaba Qwen
    • 7.1B
    • Qwen Research License (Non-commercial)
    • from 20 GB
  3. Z-Image

    Alibaba Tongyi's 6B Apache 2.0 image model. Turbo draws photoreal images with English and Chinese text in 8 steps on a 16 GB GPU, and the base model is the better starting point for fine-tuning. Turbo had about 743,000 downloads in the 30 days to 11 November 2026.

    • Tongyi MAI
    • 6.15B
    • Apache 2.0 (Permissive)
    • from 16 GB
  4. Krea 2

    Krea's 12.8B model trained from scratch, as a Raw base for fine-tuning and an 8-step Turbo. The pipeline is about 36 GB, the download is gated, and commercial use is free only below US$1 million in annual revenue. Raw had about 108,000 downloads in the 30 days to 11 November 2026.

    • Krea
    • 12.8B
    • Krea 2 Community (Conditional)
    • from 41 GB
  5. MOSS-Image

    OpenMOSS's 5B Apache 2.0 image model draws long English and Chinese text and follows JSON layouts. Its low-memory profile peaks near 4 GB of GPU memory but needs about 14 GB of system RAM, and the repository is 27.2 GB with its text encoder.

    • OpenMOSS
    • 5.08B
    • Apache 2.0 (Permissive)
    • from 4.2 GB
  6. FIBO 2.0

    Bria's 8B model generates and edits images from structured JSON prompts, and Bria says it trained only on licensed data. The weights are CC BY-NC 4.0, so commercial use needs an agreement with Bria, and the pipeline is about 25.5 GB.

    • Bria AI
    • 8.13B
    • CC BY-NC (Non-commercial)
    • from 29 GB
  7. Ideogram 4

    Ideogram's first open-weight model: 9.3B parameters trained on structured JSON captions, with text rendering that Ideogram's own comparisons put ahead of Qwen-Image and FLUX.2 [dev]. The license is non-commercial and the download is gated. The fp8 repository is about 27.5 GB, and the 16.1 GB nf4 build needs an NVIDIA GPU.

    • Ideogram
    • 9.3B
    • Ideogram Non-Commercial (Non-commercial)
    • from 19 GB
  8. HiDream-O1-Image

    HiDream's 8B model under the MIT license drops the usual autoencoder and text encoder, so one network does text-to-image, instruction editing and multi-reference scenes at up to 2,048 by 2,048. Each repository is about 35.2 GB because the weights are stored in FP32, and a CUDA GPU is required.

    • HiDream.ai
    • 8.8B
    • MIT (Permissive)
    • from 20 GB
  9. Ming-Image 0.1 Design

    inclusionAI's two 6B models under the MIT license: one draws text-heavy designs, interfaces and infographics, with transparent RGBA output, and the other splits a flat design into editable layers. The README calls a single GPU with 80 GiB the minimum validated setup.

    • inclusionAI (Ant Group)
    • 6.2B
    • MIT (Permissive)
    • from 80 GB
  10. Anima

    A 2B anime and illustration model from CircleStone Labs and Comfy Org, with a Turbo version that needs only 8 to 12 steps. The weights are non-commercial, though the images you make can be sold.

    • CircleStone Labs
    • 2.09B
    • CircleStone Non-Commercial (Non-commercial)
    • from 6.44 GB
  11. MiniMax H3-Image

    MiniMax's 33B image model makes 16-megapixel pictures in six steps and edits with up to 16 references. By our estimate it needs about 155 GB of memory, and the license excludes the EU, the UK, South Korea and the United States.

    • MiniMax
    • 33.1B
    • MiniMax H3-Image Community (Conditional)
    • from 155 GB
  12. P-H3-Image

    Pruna's 8-step image and editing model, distilled from MiniMax-H3. The ComfyUI 4-bit set is a 20 GB download that peaks at about 21 GB of VRAM at 8 MP, and it edits from up to 8 references. It keeps H3's license, which excludes the EU, the UK, South Korea and the United States, and Pruna gives speed figures but no quality scores.

    • Pruna AI
    • 33.1B
    • MiniMax H3 Community (Conditional)
    • from 21 GB
  13. SenseNova-U1.5

    SenseTime's Apache 2.0 native unified model that generates images up to 4K, edits them with several references and also answers questions about pictures, with no separate image encoder or VAE. In SenseNova's own tables it scores 0.92 on GenEval and 0.988 and 0.989 on English and Chinese long-text rendering. A 24 GB card runs it with layer offloading.

    • SenseTime
    • 17.5B
    • Apache 2.0 (Permissive)
    • from 24 GB
  14. LTX-2.5

    LTX's 22B model makes video and synchronized audio together, with native multi-shot scenes. A distilled INT8 build runs in about 16 GB. The license is free only below US$10 million of annual revenue.

    • LTX (Lightricks)
    • 22B
    • LTX-2.x Community (Conditional)
    • from 16 GB
  15. Wan 2.2

    Alibaba's Apache 2.0 video family: two-expert 27B models for text-to-video and image-to-video that need about 80 GB, plus a 5B model that runs on a 24 GB GPU.

    • Alibaba Wan
    • 27B (14B active)
    • Apache 2.0 (Permissive)
    • from 24 GB
  16. Kandinsky 6.0 Video

    Kandinsky Lab's MIT-licensed video family: a 3B Lite and a 29B Pro that make 5-second clips with synchronized 44 kHz audio and lip-sync. The README times a distilled clip at 68 seconds for Lite and 235 seconds for Pro on an RTX 4090, against 955 seconds for the 50-step Pro.

    • Kandinsky Lab
    • 29B
    • MIT (Permissive)
    • from 16 GB
  17. Cosmos 3

    NVIDIA's open world model for robots and driving, in 4B, 16B and 64B sizes: it generates video, images and actions and also answers questions about video. OpenMDW-1.1 allows commercial use, and by our estimate the 4B Edge needs about 11 GB while the 64B Super needs about 153 GB.

    • NVIDIA
    • 64.6B
    • OpenMDW-1.1 (Permissive)
    • from 11 GB
  18. Hy-Nomas

    Tencent's Apache 2.0 interactive world model: it continues a video world segment by segment from text instructions and camera paths, with a cache that keeps earlier segments in memory. It outputs 720 by 1280 at 24 FPS in 8 steps. The default setup needs eight GPUs, and a separate offloading setup targets one RTX 4090.

    • Tencent
    • 8.55B
    • Apache 2.0 (Permissive)
    • from 24 GB
  19. Pika H3 Flex

    Pika's six-step post-train of MiniMax H3 makes video with stereo sound from text, frames or up to 12 references. It needs the 147 GB of H3 base files and keeps H3's territory limits, and Pika's 4K and audio-stem features are hosted only.

    • Pika
    • 33.1B
    • MiniMax H3 Community (Conditional)
    • from 169 GB
  20. HyperFlow H3 Flash

    Video Rebirth's compressed MiniMax H3: 13.7B parameters, video with stereo sound in 4 to 8 steps, an INT8 file of about 16 GB and a 9.1 GB GGUF that runs on a 24 GB GPU. It still needs H3's companion files and keeps H3's territory limits, and Video Rebirth gives speed figures but no quality scores.

    • Video Rebirth
    • 13.9B
    • MiniMax H3 Community (Conditional)
    • from 18 GB
  21. MiniMax H3

    MiniMax's 33B video model makes video and stereo sound in one pass, up to 15 seconds. By our estimate it needs about 166 GB of memory, and the license excludes the EU, the UK, South Korea and the United States.

    • MiniMax
    • 33.1B
    • MiniMax H3 Community (Conditional)
    • from 166 GB

How we chose

  • Open weights for image or video generation, with a license we have read in full.
  • Memory figures are the developer's stated minimum or a published measurement where one exists; otherwise download size plus about 15%.
  • Quality claims come from the developers and have not been independently verified.