How to get started adapting LightX2V techniques to Text to Image Models

[MS00-GitIt]

Hi team—thanks for LightX2V, it’s great work. I’ve read the docs and README and see official support for HunyuanVideo, Wan 2.1/2.2, SkyReels-V2-DF, and CogVideoX1.5.

My goal: explore whether LightX2V methods (e.g., step distillation, quantization, attention kernels, offloading) can be adapted to image models like Skyworks UniPic (MIT), SDXL, or Chroma (Flux Schnell variant). I realize these are image T2I models (not video), so I’m asking:

1. Which components are model-agnostic vs. tightly coupled to current video backbones?
2. Is there a minimal path or example showing how to plug a new (non-video) UNet/DiT + VAE stack into LightX2V’s runtime?
3. If adaptation is feasible, what would you recommend as the first milestone (e.g., quantized inference only, then step-distill)?
4. Any constraints I should expect around schedulers, CFG removal, or resolution strategies?

My environment (for context): Linux, CUDA 12.x, Python 3.10; GPU: RTX 3090 24GB VRAM.
Happy to prototype and contribute a PR or notes if you can point me to the right extension points.

Links I consulted:

• README + Supported Models
• Method tutorials (Step Distillation, Quantization, Attention, Offloading)
• Deployment guides