FramePack Video Generation Updates
FramePack offers an easy install process for video animation. Recent pull requests and forks add Start+End Frame conditioning alongside the original image input. Higher definition outputs (e.g., 1080p) are achievable, though prompting can need refinement to ensure consistent activity throughout the video.
Links:
- https://github.com/lllyasviel/FramePack/releases/tag/windows
- https://github.com/lllyasviel/FramePack/pull/167
- https://www.patreon.com/posts/126855226
HiDream Settings and Training
For HiDream inference, use the "Dev" model with Euler sampler, ddim_uniform scheduler, 20 steps, and SD3 Sampling at 1.72 for optimal balance. Use the "Full" model for training LoRAs, which preserves base model knowledge well but requires significant VRAM (24-48GB currently).
Advanced LoRA Training Techniques
When training LoRAs for models like HiDream or Flux, consider using adaptive optimizers like Prodigy ('d_coef=2') or DAdaptAdam. For HiDream LoRAs, train on the "Full" model but use the LoRA with the "Dev" model during inference for potentially better results.
Links:
Performance Acceleration and AMD Support
Easier methods for installing Triton and SageAttention on WIndows using pre-built wheels or .bat scripts are available, offering significant speed boosts. AMD GPU support continues to improve, with ZLUDA enabling tools like ComfyUI and Forge forks to run on Windows/AMD setups.
Links:
- https://github.com/woct0rdho/SageAttention/releases
- https://civitai.com/articles/12851/easy-installation-triton-and-sageattention
- https://github.com/patientx/ComfyUI-Zluda
Workflow and Tooling Enhancements
Diffusion Toolkit v1.9 adds features like file renaming, compact view, external app integration, and archived folders. ComfyUI sees new nodes like NoteManager and Prompt Composer. Skip Layer Guidance for Wan video works by making the unconditional estimate worse, improving the final result.
Links:
