<h2 id="t800_robot-half-robot-and-full-endoskeleton-lora-(v2)-for-wan-2.1-t2v-lkvtne24r">T800_robot - Half-Robot &amp; Full Endoskeleton LoRA (V2) for Wan 2.1 T2V </h2>Fuel my GPU's caffeine addiction! (Ko-fi) <a target="_blank" rel="ugc" href="https://ko-fi.com/cyberaimania">https://ko-fi.com/cyberaimania</a>Trigger Word: T800_robotThe Upgraded Terminator Effect!Hello everyone! I'm excited to present the second version (V2) of my T800_robot LoRA, meticulously designed to generate a compelling "half-robot / damaged endoskeleton" face effect, as well as full endoskeleton bodies and robotic limbs, specifically for the Wan 2.1 T2V 14B model.Building upon the feedback and limitations of V1 (trmi800), this V2 has undergone significant improvements to offer a cleaner, more versatile, and powerful robotic transformation. The primary focus was to eliminate the "Arnold bleeding" issue and expand the creative possibilities.What's New in V2?Thanks to extensive work on the dataset and training process, V2 introduces the following key enhancements:<ul><li>Significantly Reduced "Arnold Bleeding": Through advanced dataset editing techniques, the influence of the actor's likeness has been substantially minimized, allowing for the robotic effect to be applied to a much wider range of faces.</li><li>Enhanced Versatility - Beyond the Half-Face: This LoRA now goes beyond just the half-robot face effect! V2 is capable of generating:<ul><li>Full Endoskeleton Bodies: Create complete robot characters.</li><li>Robotic Limbs: Apply robotic arms, legs, or other mechanical parts to human figures.</li><li>Mix and Match: Combine these effects! Generate a full human body with only a half-robot face, or a complete endoskeleton with a human hand – the possibilities are now truly expansive.</li></ul></li><li>Potentially Improved Results for Full Body Shots: With the inclusion of more full-figure reference materials in the dataset, V2 may offer better results when generating the robotic effect on entire bodies (though close-ups and medium shots are still recommended for optimal detail).</li></ul>I put a lot of effort into this version, as I had to practically create the dataset from scratch using other AI tools.Training Data:<ul><li>PNG Images: 257 (divided into categories)</li><li>Video Clips: 16</li><li></li></ul>Key Training Parameters:<ul><li>Base Model: Wan 2.1 T2V 14B</li><li>Epochs: 14</li><li>Micro Batch Size: 1</li><li>Gradient Accumulation: 4</li><li>Learning Rate: 5e-5</li><li>LoRA Rank: 32</li><li>Resolution: 1024 (images), 480 (videos)</li><li></li></ul>Work on V3 is underway... <a target="_blank" rel="ugc" href="https://www.google.co.uk/search?sca_esv=fa772f1b92e09aa5&amp;sxsrf=AHTn8zr2ccSl06tXFJvIIfAs8g4yXSdw3A:1746870593641&amp;q=Hasta+la+vista+baby&amp;spell=1&amp;sa=X&amp;ved=2ahUKEwiwnc-J0JiNAxWzQEEAHQ25KaYQkeECKAB6BAgREAE">Hasta la vista baby</a> ------------------------------------------------- OLDER_ V1 version: trmi800- Half-Robot Face Effect LoRA (V1) for Wan 2.1 T2V<h3 id="fuel-my-gpu's-caffeine-addiction!-(ko-fi)https:ko-fi.comcyberaimania-0i3l2ttk8"> Fuel my GPU's caffeine addiction! (Ko-fi) <a target="_blank" rel="ugc" href="https://ko-fi.com/cyberaimania">https://ko-fi.com/cyberaimania</a></h3>Trigger Word: trmi800<h3 id="introduction-and-challenges-dpjufn28w">Introduction &amp; Challenges</h3>Hello everyone! This is the first version (V1) of my <code>rbt800</code> LoRA, designed to create a specific "half-robot / damaged endoskeleton" face effect for the Wan 2.1 T2V 14B model.Creating this LoRA was a long and challenging journey. The main difficulty was gathering a suitable dataset, as high-quality source material for this specific concept is scarce. A significant portion of the available reference images and videos prominently features actor Arnold Schwarzenegger. As you'll see, this had an impact on the final result. <h3 id="training-details-806sq2gjv">Training Details</h3>This version was trained on an RTX 4090 (24GB VRAM) using <code>diffusion-pipe</code>. The process involved significant effort in data curation and parameter tuning.<ul><li>Base Model: Wan 2.1 T2V 14B</li><li>Dataset: Mixed<ul><li>114 PNG Images (Target resolution: 512p) - <code>num_repeats = 5</code></li><li>13 Video Clips (Various 480p resolutions: 480x480, 480x270, 480x320) - <code>num_repeats = 30</code></li></ul></li><li>Training Parameters (Approximation):<ul><li>Epochs 10</li><li>Learning Rate: <code>4e-5</code></li><li>LoRA Rank: 32</li><li>Optimizer: AdamW8bitKahan</li><li>Gradient Accumulation Steps: 2</li><li>Key Optimizations: <code>transformer_dtype = float8</code>, <code>activation_checkpointing = 'unsloth'</code>, <code>blocks_to_swap</code> 20</li></ul></li></ul><h3 id="results-limitations-and-known-issues-iio6gofoc">Results, Limitations &amp; Known Issues </h3>I'm not entirely satisfied with this V1, mainly due to the dataset limitations mentioned earlier.<ul><li>Arnold Bleeding: When using this LoRA alone, you will likely notice that the generated character often resembles Arnold Schwarzenegger. This is because, besides learning the half-robot effect, the LoRA also picked up the actor's distinct features present in much of the dataset.</li><li>Works Well When Mixed: The good news is that this "bleeding" effect is usually overridden when you combine this <code>rbt800</code> LoRA with another LoRA representing a different person's face or character. In such cases, the robotic effect applies correctly without imposing Arnold's likeness.</li><li>Best Use Cases: This LoRA performs best for medium shots (waist-up) and close-ups focusing on the face.</li><li>Limitations: Achieving satisfactory results for full-body shots (head-to-toe) can be challenging with this version.</li></ul><h3 id="recommended-generation-settings-thtv6u9bk">Recommended Generation Settings</h3>Based on my tests, I achieved the best results with:<ul><li>Steps: 25-30</li><li>CFG Scale: 6-7</li><li>Flow Shift (Wan Node): 6-9</li><li>Sampler: Euler (or Euler Ancestral)</li></ul>Feel free to experiment, but these are good starting points.<h3 id="future-work-and-concepts-(work-in-progress)-z4jukc0zk">Future Work &amp; Concepts (Work in Progress)</h3>I'm exploring ways to improve this LoRA and mitigate the "Arnold bleeding" issue. Here's what I've considered or am working on:<ol><li>Blurring (Failed Attempt): My initial idea was to blur the human part of the face in the dataset. Unfortunately, this backfired – the LoRA learned the blur itself, resulting in blurry human faces in generations. I don't recommend this approach.</li><li>Masked Training: The <code>diffusion-pipe</code> documentation suggests mask support (<code>mask_path</code>). This could be the ideal solution, allowing the training to focus only on the robotic parts while ignoring the human face. However, manually masking over 100 images and 13 videos is extremely time-consuming, and automated tools I've tried haven't yielded satisfactory results due to the variability between images.</li><li>Next Steps (Concept): My current plan involves a multi-pronged approach:<ul><li>Finding better source material showing the full figure for improved full-body training.</li><li>Using other tools (like Flux) to change the human face in the existing dataset from Arnold to more generic faces before retraining. This should hopefully prevent the LoRA from learning his specific features.</li></ul></li></ol>I don't have an ETA for a V2 yet. For now, I hope this V1 proves useful or fun for some of you, especially when combined with other character LoRAs. 1. LORA Trigger:<ul><li><code>trmi800</code> (Always include at the beginning of the prompt)</li></ul>2. Subject Description (The Man with the T-800 Face):<ul><li><code>a man with a battle-damaged face</code></li><li><code>half-human, half-robot face</code></li><li><code>severe damage on the [left/right] side</code> (Specify the side!)</li><li><code>exposing metallic endoskeleton</code></li><li><code>intricate metallic endoskeleton structure</code></li><li><code>torn skin and flesh</code></li><li><code>shredded, burnt flesh</code></li><li><code>glistening dark metallic endoskeleton</code></li><li><code>chrome components / chrome parts</code></li><li><code>a piercing glowing red robot eye</code></li><li><code>intact human eye and skin on the [opposite] side</code></li><li><code>expression is menacing / grim / determined</code></li><li><code>eyes are intense / scanning</code></li></ul>3. Scene/Environment Description:<ul><li><code>in a [type] environment</code></li><li><code>in a metal foundry</code></li><li><code>in a bleak, desolate, snowy landscape</code></li><li><code>on a New York City street in 1981 during the daytime</code></li><li><code>inside a classic 1980s USA telephone booth</code></li><li><code>in a dark, narrow alleyway at night</code></li><li><code>in a bustling cyberpunk city street at night</code></li><li><code>amidst the ruins of an ancient, crumbling structure</code></li><li><code>inside a dark, functional laboratory</code></li><li><code>in a scene of fire and destruction</code></li><li><code>with intense fire and smoke in the background</code></li><li><code>glowing molten metal tanks and heavy machinery are visible in the background</code></li><li><code>rain-soaked [location]</code></li><li><code>amidst vintage cars and pedestrians</code></li></ul>4. Action/Motion Description:<ul><li><code>standing still</code></li><li><code>walking [direction, e.g., to the right side of the frame]</code></li><li><code>walking sideways</code></li><li><code>walking with a determined stride</code></li><li><code>kneeling on one knee, head down</code></li><li><code>slowly stands up</code></li><li><code>rises slowly from the flames</code></li><li><code>looking directly at the camera</code></li><li><code>staring directly into the camera</code></li><li><code>turns its head [direction, e.g., towards the camera / to face slightly forward]</code></li><li><code>smoothly moves his head from left to right, scanning the area</code></li><li><code>energetically shakes his head from left to right</code> (Use for violent movement)</li><li><code>eyes darting and focusing</code></li><li><code>holding a telephone receiver to his damaged, robotic ear and talking</code></li><li><code>in slow motion</code> (Apply at the beginning of the motion description)</li></ul>5. Camera Language:<ul><li>Shots:<ul><li><code>extreme close-up (ECU) on the face / on the eye</code></li><li><code>close-up shot</code></li><li><code>medium close-up</code></li><li><code>medium shot</code></li><li><code>full shot</code></li><li><code>overhead shot</code></li><li><code>high angle perspective</code></li><li><code>low angle shot</code></li><li><code>Point-Of-View (POV) shot</code></li><li><code>establishing shot of the scene</code></li></ul></li><li>Movements:<ul><li><code>camera slowly pushes in</code> (Dolly in)</li><li><code>camera slowly zooms in</code></li><li><code>camera tracks the figure's face</code></li><li><code>camera performs a smooth tracking shot, moving alongside the man</code></li><li><code>camera pans left / right</code></li><li><code>camera tilts up / down</code></li><li><code>camera rolls</code></li></ul></li></ul>6. Atmosphere and Lighting:<ul><li><code>dramatic lighting</code></li><li><code>high-contrast lighting</code></li><li><code>dark atmosphere</code></li><li><code>intense atmosphere</code></li><li><code>gritty atmosphere</code></li><li><code>moody lighting</code></li><li><code>realistic lighting appropriate for the environment</code></li><li><code>fire provides the sole, harsh, flickering illumination</code></li><li><code>illuminated by [light source, e.g., the intense heat / flickering neon signs]</code></li><li><code>cold, harsh atmosphere</code></li><li><code>bleak and desolate atmosphere</code></li><li><code>eerie atmosphere</code></li><li><code>realistic rain effect</code></li><li><code>wet pavement reflections</code></li></ul>7. Style and Quality:<ul><li><code>cinematic video</code></li><li><code>gritty realism</code></li><li><code>high detail on textures ([type] textures, e.g., skin, metal, dust, snow])</code></li><li><code>sharp focus</code></li><li><code>cinematic quality</code></li><li><code>realistic rendering</code></li></ul>You can combine these phrases and elements to build your prompts, following the general structures we discussed earlier. Good luck with your video generation! I'll be back... 😉Let me know what you think!

TERMINATOR - Half-Robot Face Effect LoRA (V2) for Wan 2.1 T2V

V1 Notes:<ul><li>Initial release for Wan 2.1 T2V 14B. Trigger: <code>rbt800</code>.</li><li>Trained on a mixed dataset (114 images + 13 videos) featuring the half-robot effect.</li><li>Known Issue: Due to dataset sources, using this LoRA alone may result in the character resembling Arnold Schwarzenegger.</li><li>Recommendation: Works best when combined with another character/face LoRA to override the likeness bleeding.</li><li>Optimal for medium shots and close-ups. Full body shots may be inconsistent.</li><li>Exploring masking and dataset improvements for future versions to address the likeness issue.</li></ul>

Files with SHA256: d81a510ce8c611f4170ff17c4c7cce2eeaff4290fcd8874d502dd494515edaef