<h1 id="wan2.2-ultima-14b-t2v-and-i2v-low-vram-fp8-edition">WAN2.2 ULTIMA 14B T2V &amp; I2V — Low VRAM FP8 Edition</h1><h3 id="overview">📝 Overview</h3>WAN2.2 ULTIMA is a low-VRAM optimized hybrid of Wan2.2 T2V High &amp; Low Noise with carefully tuned LoRAs, built to run efficiently in FP8 while preserving cinematic visuals, motion stability, and Lightning-fast 4-step convergence.This model has been tested across six different builds before release, resulting in a stable, refined version that delivers smooth motion without sacrificing quality.<hr /><h3 id="core-setup">⚡ Core Setup</h3><ul><li>Base: Wan2.2 T2V High &amp; Low Noise (14B, FP8/FP16 friendly)</li><li>LoRAs Applied:<ul><li>Wan2.2-Lightning_T2V-v1.1-A14B-4steps-lora_HIGH_fp16 → (Strength: 0.45)</li><li>Wan2.2-Lightning_T2V-v1.1-A14B-4steps-lora_LOW_fp16 → (Strength: 0.45)<ul><li>Enables true 4-step speed with crisp detail.</li></ul></li><li>Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V1.1-HIGH_NOISE → (Strength: 0.20)</li><li>Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V1.1-LOW_NOISE → (Strength: 0.20)<ul><li>Adds motion energy and prevents the “slow-motion look.”</li></ul></li><li>Wan2.2-Fun-A14B-InP-high-noise-MPS → (Strength: 0.15)</li><li>Wan2.2-Fun-A14B-InP-low-noise-MPS → (Strength: 0.15)<ul><li>Enhances flow &amp; expressiveness in frame-to-frame motion.</li></ul></li></ul></li></ul><ul><li>Base: Wan2.2 I2V High &amp; Low Noise (14B, FP8/FP16 friendly)</li><li>LoRAs Applied:<ul><li>Wan2.2-Lightning_I2V-A14B-4steps-lora_HIGH_fp16 → (Strength: 0.45)</li><li>Wan2.2-Lightning_I2V-A14B-4steps-lora_LOW_fp16 → (Strength: 0.45)<ul><li>Enables true 4-step speed with crisp detail.</li></ul></li><li>Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1_HIGH → (Strength: 0.35)</li><li>Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1_LOW → (Strength: 0.35)</li><li>Adds motion energy and prevents the “slow-motion look.”</li></ul></li></ul><hr /><h3 id="key-improvements">🔧 Key Improvements</h3>✅ Low VRAM optimized — works under 16 GB VRAM with fp8_e4m3fn precision. ✅ Fast rendering — 4-step Lightning efficiency maintained. ✅ Fluid motion — balanced LoRA blend avoids the usual Lightning “pose lock” and slow-motion artifacts. ✅ Cinematic visuals — detailed, consistent frames without over-cooked sharpness or contrast.<h3 id="workflow-notes">🛠 Workflow Notes</h3><ul><li>WAN2.2 Native workflow → Runs well, fast, and efficient.</li><li>WAN2.2 Video Wrapper workflow → Produces better prompt alignment and more realistic visuals, especially when combined with motion-oriented models aka High and Low noise Vace (without using the wanvace2video)</li><li>Recommended if you’re pushing for cinematic realism and stronger adherence to storytelling prompts</li></ul><h3 id="recommended-ksampler-setup-for-t2v">🎚 Recommended KSampler Setup for T2V</h3>Dual KSampler chain tested for WAN2.2 ULTIMA:KSampler 1 (Noise Init)<ul><li><code>steps</code>: 8 ( step: 10 is better ofcouse but takes lil bit longer)</li><li><code>cfg</code>: 1.5</li><li><code>sampler_name</code>: euler</li><li><code>scheduler</code>: simple</li><li><code>add_noise</code>: enable</li><li><code>control before generate</code>: randomize</li><li><code>start_at_step</code>: 0</li><li><code>end_at_step</code>: 4</li><li><code>return_with_leftover_noise</code>: enable</li></ul>KSampler 2 (Refinement)<ul><li><code>steps</code>: 8 ( step: 10 is better ofcouse but takes lil bit longer)</li><li><code>cfg</code>: 1.5</li><li><code>sampler_name</code>: uni_pc_bh2</li><li><code>scheduler</code>: normal</li><li><code>add_noise</code>: disable</li><li><code>control before generate</code>: fixed</li><li><code>start_at_step</code>: 4</li><li><code>end_at_step</code>: 10000</li><li><code>return_with_leftover_noise</code>: enable (You can disable this but i prefer keeping it enabled)</li></ul><h3 id="recommended-ksampler-setup-for-i2v">🎚 Recommended KSampler Setup for I2V</h3>See Image which contains the I2V workflow.<h1 id="wan2.2-ultima-experiments-for-wanvideo2image">🚀 Wan2.2 ULTIMA Experiments for Wanvideo2Image</h1>I’ve been testing the Wan2.2 ULTIMA 14B F8 Low &amp; High Noise models (T2V + I2V) inside Wanvideo2Image, and the results are 🔥.Important: 👉 Ksampler 1 should always be Wan2.2 ULTIMA I2V 14B F8 CLEAN HIGH (uploaded here).<ul><li>This model has no baked-in LoRAs, making it perfect for the “foundation pass.”</li><li>Keep CFG ~3–4 here for smooth motion + strong prompt response.</li></ul>After that, you can stack/mix I2V High/Low and T2V High/Low models. Each combo changes how motion, texture, and coherence play out.<hr /><h3 id="example-1-stable-cinematic">⚡ Example 1 — Stable Cinematic</h3><ul><li>Ksampler 1: ULTIMA I2V CLEAN HIGH</li><li>Ksampler 2: ULTIMA I2V HIGH</li><li>Ksampler 3: ULTIMA I2V HIGH</li><li>Ksampler 4: ULTIMA I2V LOW</li></ul>💡 Result: Strong structure + cinematic motion. “High + High” builds detail and energy, while the final Low smooths transitions so it feels stable and film-like.<hr /><h3 id="example-2-dreamlike-flow">⚡ Example 2 — Dreamlike Flow</h3><ul><li>Ksampler 1: ULTIMA I2V CLEAN HIGH</li><li>Ksampler 2: ULTIMA I2V HIGH</li><li>Ksampler 3: ULTIMA T2V LOW</li><li>Ksampler 4: ULTIMA T2V LOW</li></ul>💡 Result: Adds a flowing, dreamlike quality. The T2V LOW samplers inject organic fluidity, making results feel painterly and surreal. Perfect for fantasy/artistic prompts.<hr /><h3 id="example-3-balanced-hybrid">⚡ Example 3 — Balanced Hybrid</h3><ul><li>Ksampler 1: ULTIMA I2V CLEAN HIGH</li><li>Ksampler 2: ULTIMA I2V HIGH</li><li>Ksampler 3: ULTIMA I2V LOW</li><li>Ksampler 4: ULTIMA T2V LOW</li></ul>💡 Result: Structure up front, looseness later. Great if you want characters to stay recognizable while backgrounds or effects evolve with more freedom.<hr /><h3 id="example-4-controlled-chaos">⚡ Example 4 — Controlled Chaos</h3><ul><li>Ksampler 1: ULTIMA I2V CLEAN HIGH</li><li>Ksampler 2: ULTIMA T2V HIGH</li><li>Ksampler 3: ULTIMA T2V LOW</li><li>Ksampler 4: ULTIMA I2V LOW</li></ul>💡 Result: Adds punchy, chaotic motion (T2V HIGH), softens it with T2V LOW, then stabilizes everything with I2V LOW. Feels like a dynamic camera shot that still keeps coherence.<hr /><h3 id="final-notes">🧪 Final Notes</h3><ul><li>Always anchor with Ksampler 1 = CLEAN HIGH.</li><li>Mix/match after that depending on whether you want stability, dreaminess, balance, or chaos.</li><li>Tweak CFG between 3–6 on later samplers for subtle motion/style shifts.</li></ul><hr /><h3 id="disclaimer">⚠️ Disclaimer</h3><ul><li>I do not claim authorship of Wan2.2 or the LoRAs used in this merge.</li><li>Full credit to the original model and LoRA creators — please support them on their pages.</li><li>This release is a merge + tuning effort to help the community run Wan2.2 with lower VRAM requirements.</li></ul><hr /><h3 id="notes">💡 Notes</h3><ul><li>Results may vary depending on workflow, sampler, fps settings, and hardware.</li><li>This version is stable and production-ready — tested across multiple setups and iterations.</li><li>Feedback and comparisons are welcome to further improve future merges.</li></ul>

WAN2.2 ULTIMA 14B FP8

wan22ULTIMA14BFP8_t2vHighNoise.safetensors

Files with SHA256: d93870c80c379475a9917e3b0805b26f3247d4d5b425b8be9b77995f4ca39802