Setsuna_LTX2 Beta1

This is a test LoRA model based on LTX-2, trained on 44 videos (most with audio) using LTX Trainer. The Lora still has some problems, but here I'd like to share some experiences from the training process. LTX-2 has significant limitations with Japanese language support, which appears to be a base model capability issue—most Japanese kanji characters cannot be pronounced correctly. However, there are some impressive aspects: lip-sync and audio cloning show great promise. The model fits quickly and shows results rapidly, but I believe more diverse training data (video dynamics, audio variations, etc.) is needed for better outcomes. Dataset Caption I highly recommend using Gemini for dataset captioning. Gemini can directly process both video and audio, delivering excellent results. While ComfyUI workflows with other caption nodes + Whisper audio recognition + refine are possible, they're not particularly recommended. Agent Prompt I used the following prompt for video analysis: You are a professional video analysis assistant responsible for generating descriptive text for video training datasets. Please analyze this video and output the description in the following format: 【Output Format】 Visual content: [Detailed description of visual content] Speech transcription: [Transcribe dialogue content, omit this line if none] Sounds: [Describe sounds and music, omit this line if none] On-screen text: [List on-screen text, omit this line if none] 【Rules】 1. Use plain text format, no Markdown bold or other formatting 2. "Visual content:" is required; omit entire lines (not "None") if other sections are empty 3. Descriptions should be natural and fluid, suitable for video generation model training 4. Focus on: actions, scenes, camera movement, lighting, atmosphere, emotions 5. Use present continuous tense for actions (e.g., is walking, is standing) 【Visual Content Description Points】 - Subject's actions and behaviors - Scene environment and background details - Camera movements (push, pull, pan, tilt, close-up, wide-angle, etc.) - Lighting and color tone (warm light, cool light, backlight, soft light, etc.) - Overall atmosphere and mood (serene, tense, warm, etc.) - Composition and visual elements Refine Prompt After generation, use this prompt to convert structured descriptions into natural paragraphs: You are a video description expert. Now you need to rewrite structured video descriptions into completely natural paragraph form. 【Task】 Rewrite the segmented video description into a smooth, natural paragraph, as if someone is describing what they see in the video. 【Rewriting Requirements】 1. Completely remove all labels (don't write "Visual content:", "Speech:", "Sounds:", etc.) 2. Naturally merge all information (visuals + dialogue + sounds into one fluid paragraph) 3. Like storytelling (use third-person perspective, present tense, vivid and specific with visual appeal) 4. Maintain technical details (camera movement, lighting, atmosphere descriptions should be preserved but naturally integrated) Preprocessing Configuration To prevent VRAM overflow, I used multiple bucket settings: - Maximum resolution: 768×?, 25 frames - Minimum resolution: 256×?, 121 frames - Unified frame rate: 24fps - Frame count requirement: Must satisfy frame_count % 8 == 1 (e.g., 1, 9, 17, 25...) Training Configuration Using RTX 4090 (48GB VRAM), no fp8/quanto quantization needed, training config was barely modified. Training Progress: - Sampling every 400 steps - First sample (400 steps) already captured basic character features - At 800 steps, ComfyUI testing showed incomplete character feature retention - Total training: 3400 steps (approximately 8 hours) - Final weight used: 2200 steps (1400 steps might also work well) Testing Findings Prompt Impact Prompts significantly affect results: - Without background description, likely generates simple solid color backgrounds - Need detailed descriptions of character movements (hair flowing, etc.) - Overall motion tends to be minimal Workflow Configuration Using official T2V workflow, found second-pass sampling has significant impact: - First pass (720p): More aligned with LoRA characteristics but slightly blurry - Second pass (spatial upscale): Clearer, but audio quality slightly degraded - Without LoRA in second pass, noticeable differences appear—more 2.5D style, most obvious in lip-sync details (presence/absence of teeth) Interesting Discovery Using cinematic-style descriptions produces interesting effects. Future Plans Planning to test training with larger-scale datasets. Reference Documentation: