<h3 id="if-you-like-the-model-i-will-also-make-the-fp8-version">If you like the model, I will also make the FP8 version</h3>VelvetForge_ZIT is a hybrid model developed inside the Z‑Image Turbo pipeline, created through direct LoRA integration within a flow‑based architecture using QWEN 3.4B CLIP and BF16 precision. It is not an SDXL model, does not rely on diffusion sampling, and does not support traditional schedulers. VelvetForge_ZIT operates entirely through the AuraFlow system, using low‑step sampling with deterministic behavior.This model blends the structural consistency of Z‑Image Turbo with a velvet‑soft semi‑real aesthetic, characterized by:<ul><li>smooth transitions</li><li>clean, non‑plastic skin</li><li>soft, controlled lighting</li><li>natural editorial look</li><li>absence of diffusion‑style micro‑noise</li><li>stable color response even at low CFG</li></ul>The fusion was calibrated to preserve the physics of the flow model, avoiding drift, artifacts, and instability commonly seen in static merges. VelvetForge_ZIT uses dynamic LoRA injection during inference, with adaptive strength handling and internal normalization, ensuring consistent results even with minimal prompting.<h1 id="pipeline-and-architecture">⭐ Pipeline &amp; Architecture</h1><ul><li>Base Model: Z‑Image Turbo (flow‑based)</li><li>CLIP: QWEN 3.4B</li><li>Precision: BF16</li><li>Sampling: AuraFlow (res_multisteps / simple)</li><li>Optimal Steps: 8–9</li><li>Optimal CFG: 0–1</li><li>Compatibility: Z‑Image Turbo pipeline only</li><li>Not compatible with: Euler, DPM++, Heun, UniPC, Karras, DDIM, or any SDXL‑based scheduler</li></ul><h1 id="model-identity">⭐ Model Identity</h1>VelvetForge_ZIT produces a semi‑real aesthetic — a balanced midpoint between realism and soft editorial rendering. It is not photographic and not illustrative: it is a hybrid look with smooth light, natural textures, and clean detail.Key characteristics:<ul><li>velvet‑smooth skin</li><li>soft cinematic lighting</li><li>consistent results on male and female subjects</li><li>excellent handling of hair, clothing, and materials</li><li>clean, non‑noisy backgrounds</li><li>high coherence even with short prompts</li></ul><h1 id="recommended-usage">⭐ Recommended Usage</h1><ul><li>Sampler: res_multisteps (recommended), simple</li><li>Steps: 8–9</li><li>CFG: 0–1</li><li>Prompting: short, clean prompts without style stacking</li><li>Negative prompt: minimal or none</li><li>External LoRAs: not recommended (breaks flow‑based consistency)</li></ul><h1 id="intended-use">⭐ Intended Use</h1>VelvetForge_ZIT is ideal for:<ul><li>semi‑real portraits</li><li>half‑body editorial shots</li><li>fashion soft‑light compositions</li><li>clean, natural, velvet‑toned aesthetics</li><li>professional Z‑Image Turbo workflows</li></ul>

Velvet_Forge

This release uses FP8 quantization which significantly reduces the weight of the model and memory consumption, while maintaining style, visual rendering and perceptual quality unchanged compared to the FP16 version. The pipeline remains stable across complex portraits, high-frequency textures, natural lighting, and photographic fidelity.

Files with SHA256: afb14f70e33f4da43f8b883e6484662f46ba2e214a8e9156c491d66d9fd10233