<h1 id="yozora-xl:-a-rectified-flow-sdxl-model">Yozora-XL: A Rectified Flow SDXL Model</h1>Yozora-XL is a rectified-flow model based on <a target="_blank" rel="ugc" href="https://civarchive.com/models/2363696/chenkinnoob-xl-v02-rectified-flow">Chenkin 0.2 RF</a>, fine-tuned using the <a target="_blank" rel="ugc" href="https://github.com/Hysocs/Aozora_SDXL_Training/tree/main">Aozora</a> Training Script now supporting Flow-based SDXL architectures. Aozora It enables full/partial fine-tuning on 12GB consumer GPUs such as the RTX 3060. The training script is available on GitHub at [<a target="_blank" rel="ugc" href="https://github.com/Hysocs/Aozora_SDXL_Training/tree/main">Aozora</a>] for community use but requires general understanding to setup and use. <ul><li>Never merged</li><li>No internally merged loras </li></ul><h3 id="version-0.1">Version 0.1</h3>The initial release (v0.1 alpha) is a proof-of-concept demonstrating the trainer's Rectified Flow support. It was trained to validate Flow-based fine-tuning on consumer hardware rather than achieve final quality. Future versions will utilize 50k+ images and extended training schedules. Even at this stage, the model improved and provides decent colors and improved lighting in some scenes, with stable performance across wide CFG ranges without offset noise. <h2 id="training-settings">Training Settings</h2><ul><li>Base Model: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2363696/chenkinnoob-xl-v02-rectified-flow">ChenkinNoob-XL-V0.2 RF</a></li><li>Max Train Steps: 91,567</li><li>Batch Size: 1 with 16 Gradient Accumulation Steps</li><li>Learning Rate: 2e-5 (Graph shown below, The lr was spiked half way into the run due to unforeseen issue)</li><li>Shift: 2.0</li><li>Optimizer: <a target="_blank" rel="ugc" href="https://github.com/Hysocs/Aozora_SDXL_Training/blob/main/optimizer/raven.py">Raven</a></li><li>Mixed Precision: bfloat16</li><li>VRAM Usage: ~11.8GB</li><li>Timestep Mode: Uniform</li><li>UNET Training: ~92% of parameters</li><li>Dropout: 15% unconditional</li><li>Loss: Semantic Loss (used 0.2x) </li></ul><h2 id="training-graphs">Training Graphs <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e0d7bdfd-667b-41a6-84b2-5df9626b783a/original=true/e0d7bdfd-667b-41a6-84b2-5df9626b783a.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/dc8ce86f-a1ca-4784-8f73-d227f483d590/original=true/dc8ce86f-a1ca-4784-8f73-d227f483d590.jpeg" /></h2><ul><li>This model was trained with a semantic-aware loss that approximates expensive perceptual metrics (e.g., LPIPS) with analytical importance maps. Rather than running auxiliary networks per step, it combines color saliency (bilateral-filtered LAB deviation) with structural edges (Sobel filtering) to weight the diffusion training loss spatially. This prioritizes semantically important regions—subjects and fine details—without the computational overhead of network-based evaluation." <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6d220975-b0ab-4e2b-902f-9591022985fa/original=true/6d220975-b0ab-4e2b-902f-9591022985fa.jpeg" /> </li></ul><h3 id="quick-start">Quick Start</h3><ul><li>Sampler: Euler | CFG: 6 | Steps: 25 | Shift: 3</li><li>Positive: masterpiece, best quality, aesthetic</li><li>Negative: worst quality, low quality, bad anatomy</li><li>ModelSamplingSD3 node in comfyui or Advanced Model Sampling wth sd3 in reforge</li></ul><h3 id="recommended-settings">Recommended Settings</h3><ul><li>Positive Prompt: masterpiece, best quality, aesthetic</li><li>Negative Prompt: worst quality, low quality, bad anatomy, low resolution</li><li>Sampler: Euler (Quality may vary with others)</li><li>Scheduler: Normal/Simple/SGM Uniform</li><li>Steps: 20-50</li><li>CFG Scale: 4-8</li><li>Shift: 3-8</li><li>Resolution: 1024x1024 (up to 1152x1152) or any aspect ratio variations of this range </li></ul>This is the workflow i use, in json format if needed [<a target="_blank" rel="ugc" href="https://github.com/Hysocs/Aozora_SDXL_Training/blob/main/misc/YozoraComfyuiWorkflow.json">YozoraComfyuiWorkflow</a>]Note: This model requires SD3 Flow loading to work, You will need the ModelSamplingSD3 node. You can copy the workflow form any of the preview images as a example For A1111/ReForge, enable the Advanced Model Sampling extension and use RF-specific samplers (Euler Comfy, etc.). For ADETAILER compatibility, add <code>advanced_model_sampling_script</code> to your builtin scripts list. <h2 id="license">License</h2>This model follows the license of its base, ChenkinNoob-XL RF. Review and comply with those terms.

Yozora-XL Rectified-Flow

Initial proof of concept release - The model was trained on 13k images, this is not enough to full converge and stabilize the model and may cause degradation on some seeds but quality improvements on others. future releases will be trained on 5x more data

Files with SHA256: 00bffffdd9a6056edeb6daf8d9273f6d9cfc247aebf8541ed0a7dec4ff702b56