<h1 id="galaxyace-lora">📱 GalaxyAce LoRA</h1><h2 id="minimax-h3-added">🎬 MiniMax H3 added</h2>Model: MiniMax H3Recommended LoRA Strength: 1.00Trigger Word: Not required<hr /><h2 id="krea-2-added">📷 Krea 2 added</h2>Model: Krea 2 (trained on Krea-2-Raw, works on Raw &amp; Turbo)Recommended LoRA Strength: 1.0 for in-dataset scenes — up to 2.0 (see Strength Behavior)<hr /><h3 id="ltx-2.3-added">🎬 LTX-2.3 added</h3>Model: LTX-2.3 Recommended LoRA Strength: 1.00 Trigger Word: Not requiredIn LTX 2.3 T2V&amp;I2V ComfyUI Workflow, LoRA is connected immediately after the checkpoint node inside the subgraph<hr /><h3 id="ltx-2-added">🎬 LTX-2 added</h3>Model: LTX-2 Recommended LoRA Strength: 1.00 Trigger Word: early 2010s low-end android phone footage This LoRA replicates the early 2010s video style of a Samsung Galaxy Ace smartphone. The training videos include streets, parks, everyday scenes, casual portraits, and real-life moments typical for low-end Android devices.The LoRA transfers:<ul><li>color reproduction of the Galaxy Ace camera</li><li>sensor noise &amp; compression artifacts</li><li>dynamic range and exposure behavior typical for small sensors</li><li>lens softness and subtle edge blur</li><li>subtle motion jitter and handheld feel</li><li>overall “early 2010s low-end Android phone video” aesthetics</li></ul>⭐ How to Use<ul><li>LoRA Strength: 1.00</li><li>Base Model: LTX-2</li><li>Best with natural scenes: street videos, parks, cafes, casual portraits, travel clips, everyday life</li><li>Adds subtle imperfections and low-fi realism for video</li><li>Recommended workflow: official ComfyUI LTX-2 Text-to-Video pipeline</li></ul>🧪 Notes<ul><li>Ideal for: street videos, diary-style clips, lifestyle captures, travel footage, nature</li><li>Preserves realism, avoids stylization, reproduces small-sensor behavior</li><li>Enhances authenticity while keeping motion, lighting, and colors faithful to early 2010s low-end Android look</li></ul><h3 id="z-image-turbo-(photos)">📷 Z-Image-Turbo (Photos)</h3>Model: Z-Image-Turbo Recommended LoRA Strength: 1.20⭐ How to Use<ul><li>LoRA Strength: 1.20</li><li>Base Model: Z-Image-Turbo</li><li>Works perfectly with natural, simple prompts</li><li>Adds subtle imperfections and real-photo feel of a low-megapixel phone camera</li></ul>🧪 Notes<ul><li>Best results: street photos, portraits, pets, travel shots, nature, everyday objects</li><li>Maintains realism, avoids stylization</li><li>Recreates small-sensor characteristics of the original device</li></ul>

GalaxyAce LoRA

🎬 MiniMax H3Model: MiniMax H3 (open weights, H3-Base)Recommended LoRA Strength: 1.00Trigger Word: Not requiredThis LoRA brings the Galaxy Ace look to MiniMax H3 — the color reproduction, exposurebehavior and lens softness of an early 2010s low-end Android phone, now on a model withnative audio.⭐ How to UseBase Model: MiniMax H3-Base — 768p, 24 fpsguidance_scale: 1 — H3 is guidance-distilled. Side effect: the negative prompt doesnothing, so anything unwanted has to be handled with positive wording insteadFrame count must be 17n+5 (5, 22, 39 ... 107, 149, 209, 345). Resolution in multiples of 32Best with natural scenes: street clips, car interiors at night, kitchens, courtyards,casual portraits, action from a phone in someone's handDrop to 0.7 only if speech goes out of sync — softness blurs the mouth. Don't rewrite theprompt to fix sync🧪 Prompting H3 (different from LTX and Krea)H3's text encoder is a language model (Qwen3-VL), not CLIP. That changes how you write:– (text:1.3) weighting does nothing — the encoder reads brackets as characters– tag tails (24yo, car selfie, night, candid) don't work either — write connected prose– structure that works: scene → beats with timecodes → how the camera behaves → audio,spelled out → no text or logos– write the audio explicitly or the model will put music under the clip– dialogue works, in Russian and the other languages H3 supports. Keep the prompt body inEnglish and swap only the spoken line– say nothing about image quality. No "low quality", no "grain", no "compression" — that'sthe LoRA's job. Describe light as a fact instead (one bare bulb overhead, a street lamppassing across a face, sun straight into the lens) and let the sensor react🧪 Notes– Trained with Ostris AI-Toolkit, on RTX Pro 6000 Blackwell (96GB)— about an hour of training– Trigger-less — the style applies on its own– Picture only. H3's audio branch is untouched by the LoRA, so voice artefacts belong to thebase model📝 Example promptClose-up selfie video of a beautiful blonde woman in her early twenties in the passengerseat of a car at night, her face filling most of the frame. Long light blonde hair, blueeyes, soft natural makeup, a thin gold chain at her throat. The only light on her face isthe phone screen from below and orange street lamps sliding past the window behind her asthe car moves.0:00-0:02 - She looks out through the side window, then turns her eyes into the phone andsays, "We're almost there."0:02-0:04 - A street lamp passes and washes orange across her face and the seat behind her.She says, "Twenty minutes, maybe."0:04-0:06 - She smiles slightly, looks down, then back up into the phone.She is holding the phone herself, close to her face, and her hand moves with the car.Audio: her voice close to the microphone, quiet and unperformed. Tyre noise on wet asphalt,the low hum of the engine, a turn signal ticking twice. No music.No text, no captions, no logos, no watermarks anywhere in the frame.149 frames, 768p, strength 1.00, guidance_scale 1, empty negative.

Files with SHA256: 98aa5e0b5ab3343191a6c5aa670e7fa73e2c8479b08a89955c2c8b555bc381d9