<h1></h1>The Best Result: V1 Adapter Training Style: Anime / Illustration Trigger Word: <code>aimaginedworlds</code> Base Model: <code>Tongyi-MAI/Z-Image-Turbo</code><hr /><h2 id="my-story:-the-road-to-the-&quot;perfect&quot;-lora">📖 My Story: The Road to the "Perfect" LoRA</h2>I want to share my experience training this LoRA not just the final product, but the entire journey, because I believe transparency helps the community learn.<h3 id="attempt-1:-the-1000-image-dataset-+-v2-adapter">🚫 Attempt 1: The 1000-Image Dataset + V2 Adapter</h3>I started big. I thought more data = better results, so I gathered 1000 images and used the newest adapter:<ul><li>Adapter: <a target="_blank" rel="ugc" href="https://huggingface.co/ostris/zimage_turbo_training_adapter/tree/main">ostris/zimage_turbo_training_adapter V2</a></li><li>Result: Complete failure. The LoRA didn't capture the anime style at all. The outputs looked generic and lacked any personality from the training data.</li></ul><h3 id="attempt-2:-curated-100+-image-dataset-+-v2-adapter">⚠️ Attempt 2: Curated 100+ Image Dataset + V2 Adapter</h3>I realized quality beats quantity. I carefully curated a smaller dataset of ~118 high-quality anime images with detailed captions.<ul><li>Result: Better, but still not amazing. The V2 adapter seemed to struggle with strong style transfer. The outputs were "okay," but not the striking anime aesthetic I was aiming for.</li></ul><h3 id="attempt-3:-trying-z-image-de-turbo">🔄 Attempt 3: Trying Z-Image-De-Turbo</h3>I switched gears entirely. I thought maybe training on the non-turbo base model would give me more control:<ul><li>Model: <a target="_blank" rel="ugc" href="https://huggingface.co/ostris/Z-Image-De-Turbo">ostris/Z-Image-De-Turbo</a></li><li>Result: Nothing amazing. While technically capable, it didn't produce the vibrant, stylized anime look I wanted. It felt "flat."</li></ul><h3 id="attempt-4:-the-v1-adapter-the-winner!">✅ Attempt 4: The V1 Adapter THE WINNER!</h3>Out of frustration, I went back to the original V1 adapter. And guess what?<ul><li>Adapter: <code>ostris/zimage_turbo_training_adapter_v1.safetensors</code></li><li>Dataset: My curated 118 anime images</li><li>Result: AMAZING! This was the breakthrough. The V1 adapter, combined with the right settings, finally captured the anime style beautifully. Fast inference, strong style, and consistent quality.</li></ul>Sometimes, the "old" version just works better.<hr /><h2 id="the-real-cost:-what-this-training-cost-me">💸 The Real Cost: What This Training Cost Me</h2>Training LoRAs isn't free. Here's the honest breakdown of what I spent on Modal cloud compute to reach this result:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/48ecf48e-20a5-4f5f-8532-a9a8b4d838f0/width=525/48ecf48e-20a5-4f5f-8532-a9a8b4d838f0.jpeg" />GPU Used NVIDIA H200Total Training Runs 10+Total Cost~$60Time Invested Multipledays of experimentation...That's $60 and countless hours of debugging, testing different adapters, adjusting hyperparameters, and waiting for training jobs to complete, all to find the perfect combination.<hr /><h2 id="the-winning-configuration">⚙️ The Winning Configuration</h2>Here is the exact configuration that produced the best results. Feel free to use it as a starting point for your own training!<pre><code>job: "extension"
config:
 name: "aimaginedworlds_turbo"
 process:
 - type: "diffusion_trainer"
 training_folder: "/root/ai-toolkit/modal_output"
 device: "cuda"
 trigger_word: "aimaginedworlds"
 network:
 type: "lora"
 linear: 32
 linear_alpha: 32
 conv: 16
 conv_alpha: 16
 save:
 dtype: "bf16"
 save_every: 250
 max_step_saves_to_keep: 4
 datasets:
 - folder_path: "/root/ai-toolkit/training_data/aimaginedworlds"
 caption_ext: "txt"
 caption_dropout_rate: 0.05
 resolution:
 - 512
 - 768
 - 1024
 train:
 batch_size: 1
 steps: 5000
 gradient_checkpointing: true
 noise_scheduler: "flowmatch"
 optimizer: "adamw8bit"
 lr: 0.0001
 dtype: "bf16"
 model:
 name_or_path: "Tongyi-MAI/Z-Image-Turbo"
 arch: "zimage:turbo"
 assistant_lora_path: "ostris/zimage_turbo_training_adapter/zimage_turbo_training_adapter_v1.safetensors"
 sample:
 sampler: "flowmatch"
 sample_every: 250
 guidance_scale: 1
 sample_steps: 8
</code></pre>Key Settings:<ul><li>Rank 32/Alpha 32: The sweet spot for style without overfitting.</li><li>V1 Adapter: The secret sauce!</li><li>5000 Steps: Enough for full convergence.</li><li>FlowMatch Scheduler: Native to Z-Image Turbo.</li></ul><hr /><h2 id="how-to-use-this-lora">🚀 How to Use This LoRA</h2>This LoRA was trained specifically for anime/illustration style. It works best when you keep prompts simple and let the trigger word do the heavy lifting.<h3 id="the-trigger-word">✨ The Trigger Word</h3>Just add <code>aimaginedworlds</code> at the start of your prompt:<pre><code>aimaginedworlds, a girl with blue hair sitting in a cafe
</code></pre>That's it! You don't need complex prompting, the style is baked in.<h3 id="recommended:-z-image-turbo-prompt-template-node">🔌 Recommended: Z-Image-Turbo Prompt Template Node</h3>For optimal results with this LoRA, use my ComfyUI-OllamaGemini node with the new Z-Image-Turbo prompt template:🔗 <a target="_blank" rel="ugc" href="https://github.com/abdallalswaiti/ComfyUI-OllamaGemini">ComfyUI-OllamaGemini</a>It does magic prompting using Flux, Veo3.1, Qwen, Gemini, Banana Pro, Imagen4, and more!<hr /><h2 id="support-my-work">❤️ Support My Work</h2>Creating high-quality LoRAs takes real time, effort, and money. As you saw above, this project alone cost me ~$60 in cloud compute and days of experimentation.If this LoRA helps you create beautiful images, please consider supporting my work. Even a small contribution helps me:<ul><li>🖥️ Cover cloud compute costs for future models</li><li>🎨 Train more high-quality anime LoRAs</li><li>📚 Share my findings with the community</li></ul>Every bit of support means the world to me and keeps me continue !<hr /><h2 id="tools-and-credits">🛠️ Tools &amp; Credits</h2>This LoRA was trained using the amazing AI-Toolkit by Ostris:🔗 <a target="_blank" rel="ugc" href="https://github.com/ostris/ai-toolkit">https://github.com/ostris/ai-toolkit</a>If you're interested in training your own LoRAs, I highly recommend checking it out. It's powerful, well-documented, and actively maintained!<hr /><h2 id="how-you-can-help">🙏 How You Can Help</h2>If you found this useful, here are some ways to support:<ol><li>💸 Support via PayPal — Help cover those GPU costs!</li><li>📢 Share your creations — Tag me so I can see what you make!</li></ol><a target="_blank" rel="ugc" href="https://paypal.me/abdallalswaiti"><img src="https://img.shields.io/badge/PayPal-Support_My_Work-00457C?style=for-the-badge&amp;logo=paypal&amp;logoColor=white" alt="PayPal" /></a><hr />Made with ❤️, frustration, and a lot of GPU hours.

Files with SHA256: 34d27f3c3ef36791284726d684877f445453b57089bb05d47c4b6aaf4a40b7ae