Guides / Getting started with local image generation · 中文版

Getting started with local image generation

For people who want to generate images on their own computer: first separate what runs locally from what is only an online demo, then check the license; VRAM data is almost entirely missing and this site will not guess for you.

Editorial picks

Separate local runs from online demos first: an "Official demo" link only means you can try it online, not that your machine can hold it; some configurations need an extra base model or companion tool — go by the matching run notes. Then check the "Commercial" line: the FLUX, Krea and Ideogram families are mostly non-commercial or community licenses, so read the terms where it says check yourself. Four picks below, each leaning a different way; open "Show how to run" to copy a command; every image model is at /guide/image/.

  • Most mature quality and ecosystem; non-commercial only FLUX.1-dev (BFL)

    The material states 12B and a non-commercial license; "Commercial" is marked check-yourself, and community feedback likewise confirms non-commercial research and development only. There is an official Space to try online; local runs go through diffusers or ComfyUI, and the card has a command. No site VRAM estimate; community feedback says FP16 nearly fills 24GB and a 12GB card needs the NF4 quant.

  • 4-step fast generation; MIT license, but the material also advises against commercial use Lens-Turbo (Microsoft)

    The material states 3.8B, 4-step distillation and an MIT license, and "Commercial" is marked allowed — but another line of the material advises against commercial products or services, so go by the license text in the repo before commercial use. No online demo; locally you clone the official repo and run the examples, and the card has a command. No site VRAM estimate; community feedback says the 20B text encoder makes loading a real hurdle.

  • Anime style; Apache-2.0, commercial use allowed Z-Anime (SeeSee21)

    The material states a full fine-tune of Z-Image under Apache-2.0 with commercial use allowed, shipped in GGUF / AIO / Diffusers formats with a ComfyUI workflow; no online demo. No site VRAM estimate: the material mentions both fp8 and bf16, and this site does not pick one. Community feedback is split, with more saying Anima is better overall; test in small batches before committing.

  • Apple Silicon ultra-light; Apache-2.0, commercial use allowed Bonsai-Image-4B (Prism ML)

    The material states 1.21GB of weights (the source's own figure, not a site estimate) and Apache-2.0 with commercial use allowed; MLX weights run only on Apple Silicon, there is a Space to try online, and locally you follow the repo instructions — the card has a command. No site VRAM estimate. The material says it is not for fine text, strict composition or detail-heavy scenes.

Full comparison

  • FLUX.1-dev (BFL)

    文本到图像生成，用于开发者与创作者的视觉内容创建

    Params
    12B
    Commercial
    需自查

    Run config black-forest-labs/FLUX.1-dev

    VRAM
    no estimate
    Details
    Show how to run 
    diffusers: FluxPipeline.from_pretrained('black-forest-labs/FLUX.1-dev')
    Why no estimate
    权重格式未知,无法估算显存
    License
    非商业许可
    Context
    不适用
    China access
    需代理

  • Krea-2-Turbo (Krea)

    12B 文本到图像扩散模型，GGUF 量化版，用于本地生成

    Params
    12B
    Commercial
    需自查

    Run config krea/Krea-2-Turbo

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    python -c "from diffusers import Krea2Pipeline; import torch; pipe = Krea2Pipeline.from_pretrained('krea/Krea-2-Turbo', torch_dtype=torch.bfloat16).to('cuda'); pipe('a fox in the snow', num_inference_steps=8, guidance_scale=0.0).images[0].save('out.png')"

    1 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    Krea-2-Community-License
    China access
    需代理
    Other forms
    原生 · GGUF

  • Juggernaut-Z-Image (RunDiffusion)

    文本到图像扩散模型，面向 AI 应用构建者

    Commercial
    不可商用

    Run config RunDiffusion/Juggernaut-Z-Image

    VRAM
    no estimate
    Details
    Show how to run 
    diffusers: from_pretrained('RunDiffusion/Juggernaut-Z-Image')
    Why no estimate
    No usable VRAM basis on record
    License
    cc-by-nc-4.0
    China access
    需代理

  • Z-Anime (SeeSee21)

    动漫风格图像生成，面向二次元AI应用开发者

    Params
    4B
    Commercial
    可商用

    Run config SeeSee21/Z-Anime

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    huggingface-cli download SeeSee21/Z-Anime

    1 more way(s) to run on the detail page.

    Why no estimate
    格式线索互相矛盾(含估算映射未支持的位宽),无法估算显存
    License
    apache-2.0
    China access
    需代理

  • Krea2-FP8 (KREA)

    Krea 2 Turbo 的 FP8 量化版，16GB 显存可用

    Commercial
    需自查

    Run config AlperKTS/Krea2_FP8

    VRAM
    no estimate
    Details
    Show how to run 
    huggingface-cli download AlperKTS/Krea2_FP8 --local-dir ./krea2_fp8
    Why no estimate
    No usable VRAM basis on record
    License
    krea-2-license
    China access
    需代理

  • Ideogram-4-NF4 (Ideogram)

    Ideogram 4 官方 NF4 量化，低显存本地出图

    Commercial
    需自查

    Run config ideogram-ai/ideogram-4-nf4

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    pip install diffusers bitsandbytes accelerate

    1 more way(s) to run on the detail page.

    Why no estimate
    No usable VRAM basis on record
    License
    other
    China access
    需代理

  • ideogram-4-fp8 (ideogram-ai)

    文本到图像生成模型，Ideogram 4 的 FP8 量化版本

    Commercial
    需自查

    Run config ideogram-ai/ideogram-4-fp8

    VRAM
    no estimate
    Details
    Show how to run 
    diffusers: from_pretrained('ideogram-ai/ideogram-4-fp8')
    Why no estimate
    No usable VRAM basis on record
    License
    other
    China access
    需代理

  • ideogram_4_turbotime_lora (ostris)

    Ideogram 4 加速 LoRA，2 步出图，无需 CFG

    Params
    未公开（LoRA）
    Commercial
    需自查

    Run config ostris/ideogram_4_turbotime_lora

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    pip install -U diffusers transformers accelerate

    1 more way(s) to run on the detail page.

    Why no estimate
    No usable VRAM basis on record
    License
    ideogram-4-non-commercial
    Context
    不适用
    China access
    需代理

  • AsymFLUX.2-klein-9B (Stanford)

    像素空间文生图模型，基于AsymFlow，适合文字渲染与细节生成

    Params
    9B

    Run config Lakonik/AsymFLUX.2-klein-9B

    VRAM
    no estimate
    Details
    Show how to run 
    Python: pip install lakonlab 后用 PixelFlux2KleinPipeline 加载适配器
    Why no estimate
    权重格式未知,无法估算显存
    License
    flux-non-commercial-license
    China access
    需代理

  • M87 (mgwr)

    Krea-2 Turbo 风格化 LoRA，电影感氛围增强

    Commercial
    可商用

    Run config mgwr/M87

    VRAM
    no estimate
    Details
    Show how to run 
    python -c "from diffusers import DiffusionPipeline; pipe = DiffusionPipeline.from_pretrained('krea/Krea-2-Turbo'); pipe.load_lora_weights('mgwr/M87'); pipe('a cat --preview').images[0].save('out.png')"
    Why no estimate
    No usable VRAM basis on record
    License
    apache-2.0
    China access
    需代理

  • Krea2 Turbo Style Reference LoRA (Krea)

    风格参考 LoRA，将参考图风格注入 Krea2 Turbo 生成

    Commercial
    需自查

    Run config ostris/krea2_turbo_style_reference

    VRAM
    no estimate
    Details
    Show how to run 
    pipe = DiffusionPipeline.from_pretrained("krea/Krea-2-Turbo", custom_pipeline="ostris/Krea2OstrisEdit", torch_dtype=torch.bfloat16); pipe.enable_model_cpu_offload(); pipe.load_lora_weights("ostris/krea2_turbo_style_reference")
    Why no estimate
    No usable VRAM basis on record
    License
    krea-2-community-license
    Context
    不适用
    China access
    需代理

  • Mage-Flow-Turbo (Microsoft)

    4步蒸馏的4B文本到图像模型，适用快速出图

    Params
    4B
    Commercial
    可商用

    Run config microsoft/Mage-Flow-Turbo

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    MIT
    Context
    不适用
    China access
    需代理

  • LLaDA-Image (inclusionAI)

    6B文生图与编辑模型，含4步Turbo版

    Params
    6B
    Commercial
    可商用

    Run config inclusionAI/LLaDA-Image

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    Apache-2.0
    China access
    魔搭可用

  • Qwen-Image-Flash (NVIDIA)

    四步DMD2蒸馏，快速文本生成图像，开发者评估

    Params
    20.43B
    Commercial
    需自查

    Run config nvidia/Qwen-Image-Flash

    VRAM
    no estimate
    Details
    Show how to run 
    sglang generate --model-path nvidia/Qwen-Image-Flash --prompt "A red fox in a snowy pine forest at golden hour, photorealistic, sharp focus, soft bokeh" --width 1024 --height 1024 --num-inference-steps 4 --guidance-scale 1.0 --true-cfg-scale 1.0 --seed 42 --save-output --output-file-path qwen-image-flash-sglang.png
    Why no estimate
    权重格式未知,无法估算显存
    License
    NVIDIA Open Model Agreement
    China access
    需代理

  • Mage-Flow-Edit-Turbo (Microsoft)

    4B 指令式图像编辑 Turbo 模型，4 步蒸馏推理

    Params
    4B
    Commercial
    可商用

    Run config microsoft/Mage-Flow-Edit-Turbo

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    MIT
    China access
    需代理

  • Lens (Microsoft)

    3.8B文本生成图像基础模型，高效训练与高分辨率生成

    Params
    3.8B
    Commercial
    可商用

    Run config microsoft/Lens

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    git clone https://github.com/microsoft/Lens && python inference.py

    1 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    MIT
    China access
    需代理

  • Krea-2-Raw (Krea)

    12B 图像生成基座，用于 LoRA 微调训练

    Params
    12B
    Commercial
    需自查

    Run config krea/Krea-2-Raw

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    pip install diffusers && python -c 'from diffusers import Krea2Pipeline; pipe = Krea2Pipeline.from_pretrained("krea/Krea-2-Raw", torch_dtype=torch.bfloat16).to("cuda"); pipe("a fox in the snow", num_inference_steps=52, guidance_scale=3.5).images[0].save("out.png")'

    1 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    krea-2-community-license
    China access
    需代理

  • Cosmos3-Super-Text2Image (nvidia)

    文本生成高保真图像，面向物理AI与创作

    Params
    64B
    Commercial
    需自查

    Run config nvidia/Cosmos3-Super-Text2Image

    VRAM
    no estimate
    Details
    Show how to run (+2 more) 
    vllm serve nvidia/Cosmos3-Super-Text2Image (vLLM-Omni)

    2 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    OpenMDW1.1
    Context
    4k tokens (文本输入)
    China access
    需代理

  • Lens-Turbo (Microsoft)

    3.8B 高效文生图，4 步蒸馏，为开发者与创作者快速生成

    Params
    3.8B
    Commercial
    可商用

    Run config microsoft/Lens-Turbo

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    HF Diffusers: LensPipeline.from_pretrained('microsoft/Lens-Turbo')

    1 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    MIT
    China access
    需代理

  • PiD (NVIDIA)

    像素扩散解码器，将潜在表征一步超分至2K/4K图像

    Commercial
    需自查

    Run config nvidia/PiD

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    hf download nvidia/PiD --local-dir . --include 'checkpoints/*'

    1 more way(s) to run on the detail page.

    Why no estimate
    No usable VRAM basis on record
    License
    NVIDIA Internal SRDML (仅限研究)
    China access
    需代理

  • CharacterSheet (Alissonerdx)

    将角色图片转为三视图、四视图及动态角色设定表的 LoRA 集合

    Commercial
    需自查

    Run config Alissonerdx/CharacterSheet

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    No usable VRAM basis on record
    License
    Civitai 模型许可
    Context
    不适用
    China access
    需代理

  • Anima-2.9B (Gazingstars)

    动漫/插画文本生图模型，面向二次元创作者

    Params
    2.9B
    Commercial
    需自查

    Run config Gazingstars123/Anima-2.9B

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    CircleStone Labs Non-Commercial License
    China access
    需代理

  • Krea-2 Depth ControlNet (Patil)

    深度条件图像生成，保持输入图像3D结构，用于风格迁移与内容变换

    Params
    862MB LoRA
    Commercial
    需自查

    Run config Patil/Krea-2-depth-controlnet

    VRAM
    no estimate
    Details
    Show how to run 
    git clone https://github.com/Tanmaypatil123/Krea-2-controlnet && cd Krea-2-controlnet && pip install -q -r requirements.txt && hf download Patil/Krea-2-depth-controlnet depth-control-lora.safetensors --local-dir . && python inference.py input.jpg -p "prompt" --lora depth-control-lora.safetensors
    Why no estimate
    No usable VRAM basis on record
    License
    krea-2-community-license
    China access
    需代理

  • H3 Character Sheet Generator (PoopMan333)

    ComfyUI 角色表生成，多参考图 360° 一致角色

    Commercial
    需自查

    Run config PoopMan333/H3_Character_Sheet_Generator

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    No usable VRAM basis on record
    License
    minimax-h3-community-license (EU/UK/韩国/美国除外)
    China access
    需代理

  • Sun-Direction LoRA (eric-venti-seeds)

    为Flux.2 Klein 9B图像指定阳光方向，球体参考工作流

    Params
    9B
    Commercial
    可商用

    Run config eric-venti-seeds/Sun-Direction-Lora-Flux2Klein9B

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    Apache-2.0
    Context
    不适用
    China access
    需代理

  • Kroma (lodestones)

    Krea 2 的 LoRA 适配器，ComfyUI 直接加载

    Commercial
    可商用

    Run config lodestones/Kroma

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    No usable VRAM basis on record
    License
    MIT
    China access
    需代理

  • Mage-Flow-4B (Microsoft)

    紧凑4B原生分辨率文本图像生成与编辑模型

    Params
    4B
    Commercial
    可商用

    Run config microsoft/Mage-Flow

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    MIT
    China access
    需代理

  • Bonsai-Image-Binary-4B (Prism ML)

    1-bit量化文生图模型，专为Apple Silicon本地运行

    Params
    4B
    Commercial
    可商用

    Run config prism-ml/bonsai-image-binary-4B-mlx-1bit

    VRAM
    ~0.6GB 1-bit(estimate)
    Details
    Show how to run 
    MLX Python: BONSAI_VARIANT=binary ./scripts/generate.sh --prompt "..."
    License
    apache-2.0
    China access
    需代理

  • Bonsai Image Ternary 4B (Prism ML)

    1.21GB三元文生图模型，4步采样，本地GPU运行

    Params
    4B
    Commercial
    可商用

    Run config prism-ml/bonsai-image-ternary-4B-gemlite-2bit

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    git clone https://github.com/PrismML-Eng/Bonsai-Image-Demo && cd Bonsai-Image-Demo && ./setup.sh && ./scripts/download_model.sh && ./scripts/serve.sh

    1 more way(s) to run on the detail page.

    Why no estimate
    格式线索互相矛盾(含估算映射未支持的位宽),无法估算显存
    License
    Apache-2.0
    Context
    不适用
    China access
    需代理

  • Bonsai-Image-4B (Prism ML)

    1.21GB 三元文生图扩散模型，苹果芯片本地运行

    Params
    4B
    Commercial
    可商用

    Run config prism-ml/bonsai-image-ternary-4B-mlx-2bit

    VRAM
    no estimate
    Details
    Show how to run 
    macOS: git clone https://github.com/PrismML-Eng/Bonsai-Image-Demo && cd Bonsai-Image-Demo && ./setup.sh && ./scripts/download_model.sh && ./scripts/generate.sh
    Why no estimate
    格式线索互相矛盾(含估算映射未支持的位宽),无法估算显存
    License
    apache-2.0
    Context
    不适用
    China access
    需代理

Common pitfalls

  • GGUF Q8 模型的效果优于 FP8 量化版 — AlperKTS/Krea2_FP8
  • FP8 是显存不足时的折中方案，非最优画质选择 — AlperKTS/Krea2_FP8
  • FP16下几乎占用RTX 3090/5090全部24GB显存 — black-forest-labs/FLUX.1-dev
  • 需NF4量化才能在12GB显存上舒适运行 — black-forest-labs/FLUX.1-dev
  • 持续生成时功耗达440W，散热不佳的机箱热点温度可超85°C — black-forest-labs/FLUX.1-dev
  • 安全过滤器比SDXL更严格，会误拦合法的医学或艺术史类prompt — black-forest-labs/FLUX.1-dev
  • 仅限非商业研究和开发用途 — black-forest-labs/FLUX.1-dev
  • 部分社区用户认为效果差，不如SD1.5配合loras/controlnet/adetailer — black-forest-labs/FLUX.1-dev
  • 相比SDXL缺少抽象关键词触发意外好结果的创造性和混沌感 — black-forest-labs/FLUX.1-dev
  • 生成速度极慢，单张图可能耗时13分钟 — ideogram-ai/ideogram-4-fp8
  • FP8量化在3090等显卡上文字质量下降明显 — ideogram-ai/ideogram-4-fp8
  • 学习曲线陡峭，几乎必须依赖第三方工具和区域提示才能出好图 — ideogram-ai/ideogram-4-fp8
  • 首发ComfyUI工作流有误，社区初体验差 — ideogram-ai/ideogram-4-fp8
  • 若不需文字渲染或区域提示，有更快的替代方案 — ideogram-ai/ideogram-4-fp8
  • NF4 量化版在低于 24GB 显存的 GPU 上仍可能 OOM — ideogram-ai/ideogram-4-nf4
  • JSON 结构化提示词学习曲线高，手写几乎不可行，依赖工具链 — ideogram-ai/ideogram-4-nf4
  • 发布时 ComfyUI 模板存在多处错误，上手体验不佳 — ideogram-ai/ideogram-4-nf4
  • 对简单肖像/风景图的生成效率不如 ZIT、Anima 等工具 — ideogram-ai/ideogram-4-nf4
  • 在小细节上容易出错，精细度不足 — krea/Krea-2-Raw
  • 缺乏后训练对齐和 RLHF，开箱即用的提示词生图效果不佳 — krea/Krea-2-Raw
  • bfloat16 推理需 16GB 以上显存，低显存需用量化版 — krea/Krea-2-Turbo
  • 许可证限制使用场景，不得用于移除安全措施、水印等 — krea/Krea-2-Turbo
  • 对提示词措辞和采样参数非常敏感，需要反复调试 — Lakonik/AsymFLUX.2-klein-9B
  • 不同设置下出图风格差异大，可复现性偏弱 — Lakonik/AsymFLUX.2-klein-9B
  • 20B 文本编码器体量偏大,部署门槛不低 — microsoft/Lens-Turbo
  • 不支持图像编辑能力 — microsoft/Lens-Turbo
  • 非 Turbo 版在 4090 上 50 步约需 40 秒,速度仍偏慢 — microsoft/Lens-Turbo
  • 细节偶尔与 prompt 偏离——食物摄影示例中部分细节未按提示生成 — microsoft/Lens-Turbo
  • 非 Turbo 版在 4090 上仅约 1.2it/s,对硬件要求高 — microsoft/Lens-Turbo
  • 已从应用商店下架,不再提供下载 — microsoft/Lens
  • 官方引导用户迁移至OneDrive/OneNote扫描,工作流变复杂 — microsoft/Lens
  • 本地直接保存功能随应用下架不可用 — microsoft/Lens
  • 结构化文本（信息图、图表）在所有变体上均失败 — microsoft/Mage-Flow-Turbo
  • 微软已从HuggingFace下架模型，社区获取困难 — microsoft/Mage-Flow-Turbo
  • 编辑推理比生成慢，Base/RL编辑在RTX 4090上约12秒 — microsoft/Mage-Flow-Turbo
  • 仍需约17-18.5GB VRAM，需中高端显卡才能运行 — microsoft/Mage-Flow-Turbo
  • 局部编辑细节（如衣物纹理）可能不够精细 — microsoft/Mage-Flow-Turbo
  • 本地部署需要 128GB+ 显存，单张 H100 80GB 都无法运行，本地部署基本不可行 — nvidia/Cosmos3-Super-Text2Image
  • NVIDIA 官方 API 仍在推进中，尚未正式上线 — nvidia/Cosmos3-Super-Text2Image
  • 容易破坏小细节，尤其在图像细节丰富的区域 — nvidia/PiD
  • 使用 Euler 采样器时会产生粗糙断边和噪点背景 — nvidia/PiD
  • 官方推荐使用 SDE 采样器加噪声注入和固定 sigma 曲线，对采样参数要求严格 — nvidia/PiD
  • 社区对 PiD 与已有方案的区别存在混淆 — nvidia/PiD
  • 降低步数存在画质取舍,证据中未完整描述具体损失 — ostris/ideogram_4_turbotime_lora
  • 1 兆像素图像在 8GB VRAM 上生成耗时约 185 秒，速度较慢 — ostris/krea2_turbo_style_reference
  • Identity Edit LoRA 是社区非官方微调，非 Krea 官方产品 — ostris/krea2_turbo_style_reference
  • 使用 Turbo 时需将 CFG 设为 1，与常规用法不同 — ostris/krea2_turbo_style_reference
  • 社区曾对其是否已发布/可用存在疑问 — Patil/Krea-2-depth-controlnet
  • 压缩技术/编码效率是核心创新，模型本身并非全新架构 [3] — prism-ml/bonsai-image-binary-4B-mlx-1bit
  • 被社区指出只是 FLUX.2 Klein 4B 的量压缩(经后训练恢复性能),而非全新模型。 — prism-ml/bonsai-image-ternary-4B-gemlite-2bit
  • 官方未在 Prism-ML 的 HF Web demo 页、HF 模型页与 GitHub 上提及 FLUX 团队或原始模型,招致批评。 — prism-ml/bonsai-image-ternary-4B-gemlite-2bit
  • 底层模型是 Qwen，模型本身并非自研架构 — prism-ml/bonsai-image-ternary-4B-mlx-2bit
  • Juggernaut X 在部分用户的批量测试中表现逊于 v9 — RunDiffusion/Juggernaut-Z-Image
  • 基于 Juggernaut Z 制作的 LoRA 与 Base Z Image 的兼容性不确定 — RunDiffusion/Juggernaut-Z-Image
  • 模型自 V6 起合并了闭源的 RunDiffusion Photo Model — RunDiffusion/Juggernaut-Z-Image
  • 标签式(tags)提示词效果不如 Anima — SeeSee21/Z-Anime
  • 有用户认为 Anima 综合表现远优于 Z-Anime — SeeSee21/Z-Anime
  • 已有更轻量的动漫专用模型抢先发布 — SeeSee21/Z-Anime
  • The VRAM line on this page is almost entirely "no estimate" because image models have several parts and several formats and this site does not pick one to guess; that does not mean they cannot run. Before downloading, read the community feedback on the detail page and the hardware notes in the official repo, and let an actual run decide.
  • An online demo does not mean it runs locally, and running locally does not mean your machine can hold it; some configurations (LoRAs, ControlNets, quantized repacks, for example) need an extra base model or companion tool — go by the matching run notes.
  • The "Commercial" line groups by license: the FLUX / Krea / Ideogram families are mostly non-commercial or community licenses, so read the terms where it says check yourself; this site's record may lag the repo's current license.

FAQ

Why do most models have no VRAM data?
An image model is usually several parts — text encoder, diffusion backbone, VAE — and its weights often come in fp8 / bf16 / GGUF variants; this site gives an estimate only when a single weight format can be determined, and does not pick one to guess. That is a gap in this site's material, not a sign the model cannot run locally. Before downloading, read the community feedback on the detail page and the hardware notes in the official repo.
Which one for anime style?
On this page the dedicated anime model under Apache-2.0 with commercial use allowed is Z-Anime; more community voices rate Anima better overall, and this site also indexes Anima-2.9B, but it carries a non-commercial license. Test both in small batches before deciding.
I just want to try the output without setting anything up?
Check the detail page for an "Official demo" link: among the picks, FLUX.1-dev and Bonsai-Image have official Spaces; Lens-Turbo and Z-Anime do not. Producing an image online does not mean it fits locally — judge those two things separately.