I2V on 2x T4 GPUs: Model Selection and Preventing Melting Artifacts
No_Cow3163 · reddit · 2026-08-11
A developer discusses building an automated image-to-video (I2V) workflow on constrained hardware (2x Tesla T4 16GB, no bf16 support). The goal is generating 3-4 second clips with subtle motion (like drifting smoke or flowing water) while maintaining a painted illustration look and avoiding the "everything melts" failure.
Key technical questions include:
- Model Selection: Weighing LTX-Video 0.9.8 2B distilled vs. Wan 2.2 TI2V-5B, excluding the larger LTX-2.3.
- Performance & Settings: Seeking real-world generation times on T4 and exact parameters (steps, cfg) to control minimal motion.
- Style Drift: Concerns that models trained on photorealistic footage will drag painted inputs toward realism.
- Alternatives: Exploring non-diffusion methods like displacement maps or particle overlays for cheaper, subtle effects.
More from Multimodal
- ComfyUI Adds Native Support for Wan Animate 2 with Real-Time Latency — petewoodbridge · 2026-08-11
- Minimax H3 vs. Gemini: Real-World Video Generation Test with Detailed Prompt — xdcfret1 · 2026-08-11
- Exploring Best ComfyUI Workflows for Realistic Photo and Video Sharpening — Mean-Crab1827 · 2026-08-11
- Turn Code Repos into Promo Videos in 10 Mins with Grok CLI and hyperframes — sujingshen · 2026-08-11
- MiniMax H3 Workflows: Turbo LoRA, Auto Prompting, and Video Previews — Hearmeman98 · 2026-08-11
- Vidu Introduces Native End-to-End Lip-Sync for Video Generation — Aiden_Tech_Ai · 2026-08-11