<h1 id="moss-pixelart-v.0.3">moss-pixelart-v.0.3</h1><h2 id="model-description">Model Description</h2>moss-pixelart-v.0.3 is a Low-Rank Adaptation (LoRA) model fine-tuned on OpenMOSS-Team/MOSS-Image-SFT specifically for generating clean, grid-aligned pixel art and 2D game assets.The model is trained to render at a strict 8x pixel scale, where every 8×8 block of output pixels represents a single logical pixel. This enforces strong grid adherence during generation and makes downsampling to native pixel resolutions straightforward.Baked with &lt;3 at <a rel="ugc" href="https://fingergun.dev/">fingergun.dev</a><h3 id="resolution-mapping-(8:1-scale)">Resolution Mapping (8:1 Scale)</h3><pre><code>Generation Resolution Native Pixel Art Grid Downscale Factor
512 × 512 (Training res) 64 × 64 px 8x (12.5%)
768 × 768 96 × 96 px 8x (12.5%)
1024 × 1024 128 × 128 px 8x (12.5%)</code></pre><h2 id="usage-and-inference">Usage &amp; Inference</h2><h3 id="trigger-word">Trigger Word</h3>Prepend pxl8x (or [pxl8x]) to your prompt to activate the 8x pixel art style.<h3 id="recommended-inference-settings-(moss-image-defaults)">Recommended Inference Settings (MOSS-Image Defaults)</h3>This LoRA is designed to run using the standard MOSS-Image-SFT default parameters:<ul><li>Sampler / Scheduler: Euler (flowmatch / sgm_uniform)</li><li>Inference Steps: 50 (or 30–50 steps for standard generation)</li><li>Guidance Scale (CFG): 3.0 (recommended range 3.0–6.0)</li><li>Flow Shift: 5.0</li><li>Skip Layer Guidance (SLG): Unconditional skip on layer 22 (uncond_skip_layers: [22], slg_start: 0.01, slg_stop: 1.0)</li><li>LoRA Weight: 1.0</li></ul><h3 id="turbo-step-distilled-lora-compatibility">Turbo / Step-Distilled LoRA Compatibility</h3>moss-pixelart-v.0.3 stacks cleanly with MOSS Turbo / DMD (8-step) distillation LoRAs. When pairing with an 8-step Turbo LoRA:<ul><li>Set Inference Steps to 8.</li><li>Keep Sampler on Euler and Flow Shift at 5.0.</li><li>Keep both LoRA weights around 1.0.</li></ul><h3 id="example-prompts">Example Prompts</h3><ul><li>pxl8x an isometric cyberpunk ramen stall at night, glowing neon sign reading "NOODLES", steam rising from bowls, rain puddles on asphalt, crisp pixel art</li><li>pxl8x a retro sci-fi space marine in bulky industrial armor holding a motion tracker, dark derelict spacecraft corridor, green CRT monitor glow</li><li>pxl8x a 3/4 top-down RPG potion shop interior, wooden shelves packed with glowing glass vials, a sleepy alchemist cat on the counter, warm lantern light</li><li>pxl8x a classic 16-bit boss sprite of a mechanical dragon breathing plasma fire, dark volcanic cavern background, high contrast arcade palette</li></ul><h2 id="post-processing-(recommended)">Post-Processing (Recommended)</h2>While the model maintains an 8×8 pixel grid natively, diffusion decoding can occasionally introduce subtle sub-pixel edge bleeding or slight grid drift ("mixels").For game-ready 1:1 assets, post-process your raw outputs with <a target="_blank" rel="ugc" href="https://www.spritefusion.com/pixel-snapper">Sprite Fusion Pixel Snapper</a> (<a target="_blank" rel="ugc" href="https://github.com/Hugo-Dz/spritefusion-pixel-snapper">GitHub</a>). It automatically detects the pixel grid, cleans up blurry or inconsistent blocks, and snaps the image down to a crisp, true-resolution sprite sheet or tile.<h2 id="training-details-and-hyperparameters">Training Details &amp; Hyperparameters</h2>Trained via Ostris's ai-toolkit (diffusion_trainer) using the following configuration:<h3 id="architecture-and-network">Architecture &amp; Network</h3><ul><li>Base Model: OpenMOSS-Team/MOSS-Image-SFT (moss_image arch)</li><li>Network Type: LoRA (diffusers format, bf16)</li><li>Linear Rank / Alpha: 32 / 32</li><li>Conv Rank / Alpha: 16 / 16</li><li>Excluded Layers: layers.22. (paired with uncond_skip_layers: [22] on the base model)</li><li>Base Quantization: convrot8 (applied to both Transformer and Text Encoder)</li><li>Trained Components: Transformer / UNet only (train_text_encoder: false)</li></ul><h3 id="training-configuration">Training Configuration</h3><ul><li>Dataset: pixel snapped 64x64 images (512x512 resolution, flip_x: false, flip_y: false)</li><li>Training Steps: 3000</li><li>Batch Size: 2 (Gradient Accumulation 2 = Effective Batch Size 4)</li><li>Optimizer: adamw8bit (lr: 0.0001, weight_decay: 0.001)</li><li>Noise Scheduler: flowmatch (timestep_type: shift)</li><li>Loss Type: mse</li><li>Differential Guidance: Enabled (scale: 3.0)</li><li>Caption Dropout Rate: 0.05 (5%)</li><li>Precision: bf16</li></ul><h2 id="limitations">Limitations</h2><ul><li>Resolution Multiples: Always generate at dimensions divisible by 16 (and ideally multiples of 64 like 512x512, 768x768, or 1024x1024) to keep the 8x pixel grid evenly aligned with the VAE latent space.</li><li>Fine Text At Low Resolutions: Generating at 512x512 yields a 64x64 logical canvas—keep in-image text short (1–2 words in block font) or generate at 1024x1024 (128x128 logical grid) if you need readable signage.</li></ul>

Files with SHA256: 77aed56b4e016101c6c58ebeafd9ff30f27bc3b8591157a96acc35948cd643e7