<h1 id="z-image-turbo-photorealistic-lighting-lora-(flow-dpo)">Z-Image-Turbo Photorealistic Lighting LoRA (Flow-DPO)</h1>This is a specialized LoRA adapter for <a target="_blank" rel="ugc" href="https://huggingface.co/Alibaba-Tongyi/Z-Image-Turbo">Alibaba-Tongyi/Z-Image-Turbo</a>, finetuned using Flow-DPO (Direct Preference Optimization for Flow Matching) to significantly enhance photorealistic lighting, cinematic shadows, and overall image quality.By utilizing Flow-DPO on perfectly spatially-aligned image pairs, this LoRA fixes the common "flat," "washed-out," or "plastic" artifacts often found in ultra-fast distilled models, delivering stunning, physically accurate lighting in just 8 inference steps.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/347a99d1-ce41-4f89-b56a-6fc5fed41092/original=true/347a99d1-ce41-4f89-b56a-6fc5fed41092.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2c8e6cf7-cb06-42ee-aa3a-61426deb6512/original=true/2c8e6cf7-cb06-42ee-aa3a-61426deb6512.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/4167ebe5-911e-479a-a454-284f1ca251c1/original=true/4167ebe5-911e-479a-a454-284f1ca251c1.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/20182d1a-2f1b-4e3d-bbfa-3a09dda997a7/original=true/20182d1a-2f1b-4e3d-bbfa-3a09dda997a7.jpeg" /><h2 id="training-details-and-methodology">🧠 Training Details &amp; Methodology</h2>This model was trained using a custom implementation of Flow-DPO (<a target="_blank" rel="ugc" href="https://arxiv.org/abs/2501.13918">Improving Video Generation with Human Feedback, arXiv:2501.13918</a>).<h3 id="1.-the-dataset-(strict-spatial-alignment)">1. The Dataset (Strict Spatial Alignment)</h3>To prevent the model from hallucinating or altering image structures (Catastrophic Forgetting), the preference dataset was constructed using strict spatial alignment:<ul><li>Win (Chosen): High-quality, professional photographs with perfect lighting and textures.</li><li>Lose (Rejected): The exact same images degraded programmatically (Gaussian blur, lowered contrast, extreme exposure shifts, gaussian noise, and heavy JPEG compression artifacts).</li><li>Alignment: No cropping or warping was applied, ensuring the Flow Matching trajectory learned to solely correct lighting and texture.</li></ul><h3 id="2.-discrete-timestep-distillation-preservation">2. Discrete Timestep Distillation Preservation</h3>Unlike standard diffusion models where $t$ is sampled continuously $t \in [0, 1]$, Z-Image-Turbo is a distilled model specifically optimized for 8 fixed timesteps. During the Flow-DPO training, we dynamically extracted the exact discrete $t$-distribution from the <code>FlowMatchEulerDiscreteScheduler</code> and restricted the random sampling to these exact 8 nodes. This ensures the LoRA retains the turbo model's extreme speed without causing output blurriness.<h3 id="3.-hyperparameters">3. Hyperparameters</h3><ul><li>Base Model: Alibaba-Tongyi/Z-Image-Turbo (6B Single-Stream DiT)</li><li>Learning Rate: <code>1e-4</code></li><li>KL Penalty ($\beta$): <code>1.0</code></li><li>Effective Batch Size: <code>1</code></li><li>Mixed Precision: <code>bfloat16</code></li></ul><h2 id="limitations">⚠️ Limitations</h2><ul><li>Not an Image-to-Image Restorer: This LoRA changes the prior distribution of the Text-to-Image generation. It is designed to generate better original images from text prompts, not to be used as an img2img filter to fix user-uploaded bad photos (unless combined with RF-Inversion techniques, which are highly unstable for 8-step models).</li><li>Color Saturation: Pushing the LoRA scale too high (e.g., &gt; 1.5) might result in over-sharpened or overly saturated images due to the nature of DPO margin maximization. Keep the scale around <code>0.6 - 1.0</code> for the most photorealistic results.</li></ul>

Files with SHA256: 1fd3c728ade7e1dec52f26ab5f860e7efd57abb06f740788f13b5b68bd94edc2