New Models & 3D Generation
HiDream-I1, a new 17B parameter open-source base model, is available (MIT license). It boasts strong prompt adherence and image quality but requires high VRAM. TripoSF, a high-resolution (1024³) 3D VAE using SparseFlex, improves 3D reconstruction detail, aiding future image-to-3D pipelines. Inference code is released.
Links:
- https://huggingface.co/HiDream-ai/HiDream-I1-Full
- https://github.com/VAST-AI-Research/TripoSF
- https://arxiv.org/abs/2503.21732
Advanced Faceswapping & Consistency
VACE combined with Wan2.1 shows promise for high-quality video faceswapping, with workflows emerging for expression tracking. For consistent characters (games/comics), generate expression sheets, use IP-Adapters (Face ID Plus v2 recommended), or inpaint faces using CLIPSeg masks for specific expressions.
Links:
- https://www.patreon.com/posts/best-faceswap-126056154
- https://stable-diffusion-art.com/ip-adapter/#IP-Adapter_Face_ID_Plus_v2_SDXL
- https://liveportrait.github.io/
Performance Tuning & Troubleshooting
Severe slowdowns reported in Automatic1111 v1.10.1; switching to ComfyUI or Forge is advised for better memory management. Common issues include Python version/path conflicts (use 3.10.x) and ZLUDA setup errors for AMD GPUs (check paths/DLLs, delete venv folder after changes).
Links:
Video Generation Techniques
Extend video length using frame interpolation like RIFLEx, or use the last frame as input for the next segment (common with Hunyuan). In Sora, use the custom blend feature for fine control over transitions via time/weight points. Smooth looping requires careful frame matching.
Links:
Model Control & Prompting
Use Regional Prompter in Stable Diffusion to apply different prompts/styles to specific image areas, useful for multi-character control. Pony models generally require specific scoring tags (e.g., score_9_up) and Clip Skip set to 2 for optimal output quality.
Links: