The latest articles on DEV Community by Mild King (@mnblabs). https://dev.to/mnblabs https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3028911%2Ffedd0dec-a784-4844-8ccc-e4c791f03a28.png https://dev.to/mnblabs en Mild King Tue, 08 Apr 2025 07:10:55 +0000 https://dev.to/mnblabs/how-to-train-flux1-for-custom-emoji-generation-dataset-size-script-and-deployment-4dn7 https://dev.to/mnblabs/how-to-train-flux1-for-custom-emoji-generation-dataset-size-script-and-deployment-4dn7 <p>I'm working on a personal project where I want to generate custom emoji-style images from text prompts — like turning this:</p> <blockquote> <p><code>Flying pig</code> → 🐖 with wings<br> <strong><em>(see cover image!)</em></strong></p> </blockquote> <p>I'm using <a href="https://huggingface.co/black-forest-labs/FLUX.1-dev" rel="noopener noreferrer">black-forest-labs/FLUX.1-dev</a> as the base model. It’s a diffusion model similar to Stable Diffusion, but optimized for low-VRAM generation.</p> <h2> What I have: </h2> <ul> <li>~25k 512x512 emoji-style images</li> <li>Captions for each (in .txt files)</li> <li>A train.json mapping image to caption </li> </ul> <div class="highlight js-code-highlight"> <pre class="highlight css"><code><span class="nt">dataset</span><span class="o">/</span> <span class="err">├──</span> <span class="nt">images</span><span class="o">/</span><span class="nt">image_001</span><span class="nc">.png</span><span class="o">,...</span> <span class="err">├──</span> <span class="nt">captions</span><span class="o">/</span><span class="nt">caption_001</span><span class="nc">.txt</span><span class="o">,...</span> <span class="err">└──</span> <span class="nt">train</span><span class="nc">.json</span> <span class="err">#</span> <span class="o">[</span><span class="p">{</span> <span class="err">"image":</span> <span class="err">"images/image_001.png",</span> <span class="err">"caption":</span> <span class="err">"captions/caption_001.txt"</span> <span class="p">}</span><span class="o">,</span> <span class="o">...]</span> </code></pre> </div> <h2> What I need help with: </h2> <ol> <li>How many images is “enough”? Is 25k too much or just fine?</li> <li>Any working training script for FLUX.1? <ul> <li>I tried one (PyTorch + diffusers), but outputs look like noise.</li> </ul> </li> <li>Best training config? <ul> <li>Should I freeze VAE/text encoder?</li> <li>Recommended batch size, LR, etc?</li> </ul> </li> <li>How do I export the model to ONNX or TFLite?</li> </ol> <p>Planning to use it in a Flutter app later.<br> <em>A sample setup or any advice would be helpful for beginners to get started.</em></p> machinelearning flutter diffusionmodels flux1dev