Guides / Getting Started with Local Image Generation (2026-07) · 中文版
Getting Started with Local Image Generation (2026-07)
18 locally runnable open-source text-to-image models, VRAM data mostly missing; read the intro before the table.
Editorial picks
This page covers all 18 text-to-image model cards on the site. The image domain has the lowest VRAM field completeness site-wide (only 1 of 19 cards has data, ~5.3%), so the VRAM column is empty for most rows in this table — that doesn't mean the models can't run, but that data hasn't been collected. Before choosing, click into a model's detail page or official repository to check hardware recommendations; don't rely solely on the blank cells in this table.
- Flagship quality/most mature ecosystem FLUX.1-dev (BFL)
why_trust notes 708k downloads and 12.9k likes, the largest community validation on this page; community.consensus states it "is widely regarded as the most powerful open-weight image generation model," solves mentions it can render readable text and is compatible with ComfyUI/Diffusers/TensorRT; but the license is non-commercial, watch_outs explicitly says "cannot be used for commercial purposes or commercial services," and commercial_use indicates you must check for yourself.
- Fast generation / high training efficiency Lens-Turbo (Microsoft)
key_info.size 3.8B, license MIT for commercial use; community.solves mentions "extremely high training efficiency, requiring about one-fifth the compute of similar models" and "with only 3.8B parameters, it achieves image quality competitive with or even surpassing large models like FLUX and SD3"; watch_outs also warns "text encoder size reaches 20B, causing significant loading and inference overhead" and "it's questionable whether 16GB VRAM can run it smoothly."
- Anime style / commercially usable Z-Anime (SeeSee21)
key_info.license is apache-2.0 (commercial use allowed), key_info.image.base_model is a fine-tune of Tongyi-MAI/Z-Image; good_for explicitly says "generating anime-style illustrations / 2D character designs"; but community.consensus also states outright "the sample preview quality is questioned by the community as not good enough" and "there is very little real-world community feedback, lacking large-scale comparative validation," so it's advisable to test in small batches first.
- Mobile / Apple Silicon ultra-lightweight Bonsai-Image-4B (Prism ML)
why_trust says "6.4× compressed to 1.21GB, Apple Silicon MLX deployment," license apache-2.0 for commercial use; community.solves mentions fast inference (about 17k t/s) and a free iPhone app Bonsai Studio; but watch_outs also notes "generation accuracy/quality is not ideal," default 4-step sampling yields limited results, and 10–20 steps are needed.
Full comparison
| Model | Params | VRAM | Context | License | Commercial use | China access | Deploy |
|---|---|---|---|---|---|---|---|
| FLUX.1-dev (BFL) | 12B | ~7.2GB 4-bit | 不适用 | 非商业许可 | 需自查 | 需代理 | diffusers: FluxPipeline.from_pretrained('black-forest-labs/FLUX.1-dev') |
| Krea-2-Turbo (Krea) 原生 · GGUF | 12B | ~7.2GB 4-bit | — | Krea-2-Community-License | 需自查 | 需代理 | 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')" / sglang generate --model-path krea/Krea-2-Turbo --prompt "a red fox sitting in fresh snow, golden hour, photorealistic" --num-inference-steps 8 --height 1024 --width 1024 --save-output |
| Juggernaut-Z-Image (RunDiffusion) | — | — | — | cc-by-nc-4.0 | 不可商用 | 需代理 | diffusers: from_pretrained('RunDiffusion/Juggernaut-Z-Image') |
| Z-Anime (SeeSee21) | 4B | ~4.4GB FP8 | — | apache-2.0 | 可商用 | 需代理 | huggingface-cli download SeeSee21/Z-Anime / from diffusers import StableDiffusionPipeline; pipe = StableDiffusionPipeline.from_pretrained('SeeSee21/Z-Anime') |
| Krea2-FP8 (KREA) | — | — | — | krea-2-license | 需自查 | 需代理 | huggingface-cli download AlperKTS/Krea2_FP8 --local-dir ./krea2_fp8 |
| Ideogram-4-NF4 (Ideogram) | — | — | — | other | 需自查 | 需代理 | pip install diffusers bitsandbytes accelerate / python -c 'from diffusers import DiffusionPipeline; pipe = DiffusionPipeline.from_pretrained("ideogram-ai/ideogram-4-nf4", torch_dtype=torch.bfloat16); pipe.to("cuda")' |
| ideogram-4-fp8 (ideogram-ai) | — | — | — | other | 需自查 | 需代理 | diffusers: from_pretrained('ideogram-ai/ideogram-4-fp8') |
| ideogram_4_turbotime_lora (ostris) | 未公开（LoRA） | — | 不适用 | ideogram-4-non-commercial | 需自查 | 需代理 | pip install -U diffusers transformers accelerate / from diffusers import DiffusionPipeline; pipe = DiffusionPipeline.from_pretrained('ideogram-ai/ideogram-4-fp8', dtype=torch.bfloat16, device_map='cuda'); pipe.load_lora_weights('ostris/ideogram_4_turbotime_lora') |
| AsymFLUX.2-klein-9B (Stanford) | 9B | ~5.4GB 4-bit | — | flux-non-commercial-license | — | 需代理 | Python: pip install lakonlab 后用 PixelFlux2KleinPipeline 加载适配器 |
| M87 (mgwr) | — | — | — | apache-2.0 | 需自查 | 需代理 | 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')" |
| Krea2 Turbo Style Reference LoRA (Krea) | — | — | 不适用 | krea-2-community-license | 需自查 | 需代理 | 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") |
| Lens (Microsoft) | 3.8B | ~2.3GB 4-bit | — | MIT | 可商用 | 需代理 | git clone https://github.com/microsoft/Lens && python inference.py / LensPipeline.from_pretrained('microsoft/Lens') |
| Krea-2-Raw (Krea) | 12B | ~7.2GB 4-bit | — | krea-2-community-license | 需自查 | 需代理 | 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")' / pip install git+https://github.com/sgl-project/sglang.git && sglang generate --model-path krea/Krea-2-Raw --prompt "a fox in the snow" --num-inference-steps 52 --height 1024 --width 1024 --guidance-scale 3.5 --save-output |
| Cosmos3-Super-Text2Image (nvidia) | 64B | ~39GB 4-bit | 4k tokens (文本输入) | OpenMDW1.1 | 需自查 | 需代理 | vllm serve nvidia/Cosmos3-Super-Text2Image (vLLM-Omni) / diffusers.Cosmos3OmniPipeline.from_pretrained() / docker pull vllm/vllm-omni:cosmos3 |
| Lens-Turbo (Microsoft) | 3.8B | ~2.3GB 4-bit | — | MIT | 可商用 | 需代理 | HF Diffusers: LensPipeline.from_pretrained('microsoft/Lens-Turbo') / CLI: git clone https://github.com/microsoft/Lens && python inference.py --repo_id microsoft/Lens-Turbo |
| PiD (NVIDIA) | — | — | — | NVIDIA Internal SRDML (仅限研究) | 需自查 | 需代理 | hf download nvidia/PiD --local-dir . --include 'checkpoints/*' / git clone https://github.com/nv-tlabs/PiD |
| Krea-2 Depth ControlNet (Patil) | 862MB LoRA | — | — | krea-2-community-license | 需自查 | 需代理 | 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 |
| Sun-Direction LoRA (eric-venti-seeds) | 9B | ~5.4GB 4-bit | 不适用 | Apache-2.0 | 需自查 | 需代理 | — |
| Bonsai-Image-Binary-4B (Prism ML) | 4B | ~0.6GB 1-bit | — | apache-2.0 | 可商用 | 需代理 | MLX Python: BONSAI_VARIANT=binary ./scripts/generate.sh --prompt "..." |
| Bonsai Image Ternary 4B (Prism ML) | 4B | ~1.1GB ternary | 不适用 | Apache-2.0 | 可商用 | 需代理 | git clone https://github.com/PrismML-Eng/Bonsai-Image-Demo && cd Bonsai-Image-Demo && ./setup.sh && ./scripts/download_model.sh && ./scripts/serve.sh / Python API: from backend_gpu.server import build_pipeline; pipe = build_pipeline('prism-ml/bonsai-image-ternary-4B-gemlite-2bit') |
| Bonsai-Image-4B (Prism ML) | 4B | ~1.1GB ternary | 不适用 | apache-2.0 | 可商用 | 需代理 | macOS: git clone https://github.com/PrismML-Eng/Bonsai-Image-Demo && cd Bonsai-Image-Demo && ./setup.sh && ./scripts/download_model.sh && ./scripts/generate.sh |
Common pitfalls
- FLUX.1-dev 使用非商业许可，不能用于商业用途或商业服务 — black-forest-labs/FLUX.1-dev
- 部分社区模型将 dev 与 Schnell 合并后以 Apache 等商业友好许可分发，存在法律风险 — black-forest-labs/FLUX.1-dev
- 生成速度比同系列的 Schnell 慢约 4 倍 — black-forest-labs/FLUX.1-dev
- 约需 20 步推理才能产出高质量图像，对低配硬件仍有算力压力 — black-forest-labs/FLUX.1-dev
- 社区 LoRA 与衍生模型(如 Krea-dev)的兼容性不保证，迁移需实测验证 — black-forest-labs/FLUX.1-dev
- 仅针对室外场景训练，室内或非外景图像效果无法保证 — eric-venti-seeds/Sun-Direction-Lora-Flux2Klein9B
- FP8 量化版本质量下降明显，尤其是文字渲染 — ideogram-ai/ideogram-4-fp8
- 自然语言提示效果不如 JSON 结构化提示 — ideogram-ai/ideogram-4-fp8
- NF4 量化版画质明显下降，尤其文字渲染变差 — ideogram-ai/ideogram-4-nf4
- 生成速度较慢 — ideogram-ai/ideogram-4-nf4
- 低 CFG 值（如 1）跑图效果不稳定 — ideogram-ai/ideogram-4-nf4
- JSON 格式提示词上手门槛高，自然语言效果不如结构化输入 — ideogram-ai/ideogram-4-nf4
- 下载模型需同意共享联系方式 — ideogram-ai/ideogram-4-nf4
- 超过 50 人团队需购买企业许可，非完全免费商用 — krea/Krea-2-Raw
- 所有用户必须实施技术安全措施以防止生成非法内容 — krea/Krea-2-Raw
- 并非传统意义上的开源，而是自定义许可证下的开放权重 — krea/Krea-2-Raw
- 真实感/照片级渲染不如 Z-Image Turbo — krea/Krea-2-Turbo
- 复杂提示词下可靠性偏低 — krea/Krea-2-Turbo
- 团队承认模型未针对真实感调优,后续版本拟改善 — krea/Krea-2-Turbo
- 人物生成误差大，经常无法正确绘制人物 — Lakonik/AsymFLUX.2-klein-9B
- 提示跟随能力弱，生成错误多 — Lakonik/AsymFLUX.2-klein-9B
- 编辑时无法保持人物面部一致性，且会轻微改变整体光照和色彩 — Lakonik/AsymFLUX.2-klein-9B
- 每条帖子垂直间距过大，在类似 Reddit 的站点上浏览效率低 — mgwr/M87
- 文本编码器体积达 20B,加载和推理开销较大 — microsoft/Lens-Turbo
- 不支持图像编辑/inpainting 等能力,功能较单一 — microsoft/Lens-Turbo
- 16GB 显存能否流畅运行存在疑问 — microsoft/Lens-Turbo
- 人体解剖结构渲染不够稳定，手部/姿态容易出错 — microsoft/Lens
- 冷门/小众题材生成质量下降明显 — microsoft/Lens
- 需使用ComfyUI nightly版本，稳定版(≤0.22.0)不兼容 — microsoft/Lens
- 模型定位为机器人与自动驾驶AI训练的世界模型，非消费级图像生成产品，与 Seedance 等消费级产品不构成直接竞争 — nvidia/Cosmos3-Super-Text2Image
- 目前社区仅有 Python 脚本形式的简易封装，尚无 ComfyUI 工作流节点 — 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
- Krea 2 Turbo 原生缺乏参考模式，保持角色或风格一致性必须依赖 LoRA 训练 — ostris/krea2_turbo_style_reference
- 原生模型有安全限制，需通过 prompt rebalancing 或 LoRA 才能解除 — 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
- 生成准确性不佳，输出质量不够理想 — prism-ml/bonsai-image-binary-4B-mlx-1bit
- 生成图像的准确度/质量明显不如全精度模型 — prism-ml/bonsai-image-ternary-4B-gemlite-2bit
- 文字生成结果差 — prism-ml/bonsai-image-ternary-4B-mlx-2bit
- 整体准确度不佳 — prism-ml/bonsai-image-ternary-4B-mlx-2bit
- 采用 CC BY-NC 4.0 许可，不可商用 — RunDiffusion/Juggernaut-Z-Image
- 基于 Z-Image Base 微调，对上游基座模型有依赖 — RunDiffusion/Juggernaut-Z-Image
- 与 Anima 对比中被多位社区用户认为动漫风格差距明显 — SeeSee21/Z-Anime
- 使用 Stable Diffusion 风格提示词时可能出现效果不佳的问题 — SeeSee21/Z-Anime
- VRAM field completeness site-wide is only 5.3% (only 1 of 19 cards in the image domain has data); blank VRAM cells in most rows of this table do not mean the model can't run. Refer to the hardware recommendations on the model detail page or official repository. For data methodology, see docs/research/2026-07-guide-inventory.md §1.3.
FAQ
- Why do most models lack VRAM data?
- This is a real gap in the current data: VRAM field completeness in the image domain is only 5.3% across 19 cards (only 1 has data). The VRAM column is empty for most rows on this page, not because models can't run locally. Before downloading, check the hardware recommendations on the corresponding model detail page or official repository.
- Which one should I use for anime style?
- SeeSee21/Z-Anime is the option on this page that specializes in anime style and is commercially usable (apache-2.0), based on a fine-tune of Tongyi-MAI/Z-Image; but community feedback on the card questions sample quality and independent validation is scarce. It's recommended to try a demo or small-scale test before deciding.