Hugging Face's logo

waveshaker
/
loras

Text-to-Image
lora
character
girl
flux
pony
sd1.5
krea2
moss
Model card Files
xet
 Community

Viewer: 🖼️LoRAs Gallery

MainPage

Dataset: 📊Dataset Collections

Tools for this repo: 🔨Loras Tools s

Training Configuration & Training Tools

  • Moss Image (Click to expand)
    • Training Tools: Onetrainer (Koratahiu‘s Fork) / Fizgig / Fizgig-Advanced(Murf-Code's Fork)
    • Dataset: Mostly 50+ images (averaging 100+)
    • Tagging Style: Mixed format (Booru tags as prefixes combined with natural language)
    • Repeats per Image: 50+
    • Training Parameters: Default settings from Onetrainer / Fizgig
      • Onetrainer tweak:
        • Attention:torch CuDNN (overriding the default SDPA)
        • Training Res: 512
        • LR:1.5e-04 - 2.5e-04
        • Batch Size:：3 - 4
      • Fizgig tweak:
        • Presets: Moss-Image Fast(rank 16, 0.25MP)
        • Pictrue Size: 0.26 MP ≈ 512 res
        • LR:1e-04
        • Gradient Accumulation: 3
      • Fizgig-Advanced tweak:
      • Presets: Moss-Image Fast(rank 16, 0.25MP)
      • Pictrue Size: 0.26 MP ≈ 512 res
      • LR:1.5e-04 - 2.5e-04
      • Batch Size: 3 - 4
    • Training Speed (Benchmark: 39 imgs, 0.26MP, r16):
      • OneTrainer: 910 steps, 1.12 s/it (Local Batch Size: 3, LR: 0.0003) → ~0.37s/img (17m 43s)
      • Fizgig(Batch-Size: Not Supported): 2730 steps, 1.25 it/s (Gradient Accumulation: 3, LR: 0.0001) → ~0.83s/img (36m 17s)
      • Fizgig-Advanced(Batch-Size: Supported): 910 steps, 2.07 s/it (Batch Size: 3, LR: 0.0003) → ~0.69s/img (31m 20s)
      • (Note on Speed & Formats: OneTrainer used the official sharded Diffusers format (bfloat16), while Fizgig used the ComfyUI single-file int8 format.)
    • Showcase & Card Image Parameters:

    LR / Batch Size / Gradient Accumulation

    • Effective Batch Size = Batch Size × Gradient Accumulation. Gradient accumulation saves VRAM but is slower; a larger Batch Size is faster but uses more VRAM.
    • LR Scaling: Don't blindly apply the linear rule to LoRA. Validate with loss curves. (My baseline: Batch 3 + LR 0.0003). Note: The vanilla Fizgig benchmark used the default LR 0.0001, which is under-scaled for an effective batch of 3. As widely discussed on GitHub and Reddit, increasing the batch size generally requires a higher LR to compensate for fewer update steps.

    Paper TL;DR:

    • Accurate, Large Minibatch SGD (2017): The "Linear Scaling Rule" — if you multiply batch size by k, multiply LR by k. This was designed for full fine-tuning on ImageNet, not LoRA (e.g., BS 1, LR 1e-4 → BS 4, LR 4e-4).
    • One weird trick for parallelizing CNNs (2014): The "Square Root Scaling Rule" — if you multiply batch size by k, to keep gradient variance stable, scale LR by k^0.5 instead of k. More conservative when batch gets very large (e.g., BS 1, LR 1e-4 → BS 4, LR 2e-4).
    • Beware of the Batch Size (2026): For LoRA specifically, batch size is not just a multiplier for LR — it's a first-order design parameter. Optimal batch depends on dataset size (most sensitive), rank, and model capacity. Vanilla LoRA matches fancy variants when batch is tuned properly.
  • Krea2 (Click to expand)
  • Anima (Click to expand)
    • Training Tools: Onetrainer
    • Dataset: 20 ~ 100 images
    • Tagging Style: Mixed format (Booru tags as prefixes combined with natural language)
    • Training Res: 512
    • Repeats per Image: 50+
    • Training Parameters: Mostly default settings from Onetrainer
    • Showcase & Card Image Parameters:
  • Flux.1-dev (Click to expand)
  • Stable Diffusion v1.5 & Pony Diffusion V6 XL (Click to expand)

Usage Disclaimer

  1. Educational & Research Purpose Only: This repository hosts LoRA models trained on various base models, including Pony, Flux 1 Dev, Anima, Krea2, Moss Image, and SD1.5. All models are intended solely for academic exchange, technical testing, and personal research into image generation algorithms.
  2. Prohibition of Illegal Use: Users must strictly adhere to local laws and regulations when downloading and using these models. It is strictly forbidden to use these models to infringe upon others' privacy, reputation, portrait rights, or other legitimate interests. Do not use them for creating or spreading misinformation, fraudulent activities, or any unethical purposes.
  3. Takedown Notice: These models are shared for non-commercial research purposes. If you are a copyright owner or an authorized agent and believe that any content in this repository infringes upon your rights, please contact me directly. I will promptly review and remove the infringing material upon verification, without any delay.
  4. Content Compliance Notice: Please be aware that certain base models (e.g., Pony) may exhibit a high degree of openness regarding specific prompts. Users are entirely responsible for filtering and auditing the compliance of the generated content.
  5. Limitation of Liability: As the creator/tester of these experimental models, I take no responsibility for any direct or indirect legal liabilities, losses, or disputes arising from the use of this repository. Any image generated and its subsequent consequences are the sole responsibility of the user.

使用免责声明

  1. 仅限教育与研究目的：本仓库托管了基于多种基础模型（包括 Pony、Flux 1 Dev、Krea2、Moss Image、Anima 和 SD1.5）训练的 LoRA 模型。所有模型仅供学术交流、技术测试以及个人对图像生成算法的研究使用。
  2. 禁止非法使用：用户在下载和使用这些模型时，必须严格遵守当地的法律法规。严禁使用这些模型侵犯他人的隐私、名誉、肖像权或其他合法权益。不得将其用于制造或传播虚假信息、欺诈活动或任何不道德的目的。
  3. 侵权通知与删除：本仓库模型仅限非商业研究分享。如果您是版权所有者或授权代理人，并认为本仓库中的任何内容侵犯了您的权利，请直接与我联系。核实后，我将立即审查并毫不拖延地删除相关的侵权内容。
  4. 内容合规提示：请注意，某些基础模型（如 Pony）在特定提示词方面可能具有较高的开放性。用户需全权负责对生成内容的合规性进行过滤和审查。
  5. 责任限制：作为这些实验性模型的创建者/测试者，我对因使用本仓库而产生的任何直接或间接的法律责任、损失或纠纷概不负责。任何生成的图像及其后续产生的一切后果，均由用户自行承担全部责任。
Downloads last month
1,184
Inference Providers NEW
Text-to-Image
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for waveshaker/loras

Base model

krea/Krea-2-Raw
Adapter
(1,166)
this model

Spaces using waveshaker/loras 3

Papers for waveshaker/loras