FramePack Timestamped Prompts
Efforts are underway to integrate timestamped prompts into FramePack, allowing specific actions or style changes at defined times within longer video generations. Early tests show promise for controlling narrative flow in extended clips. Initial setup can be tricky, requiring adjustments for timing precision.
Links:
- https://github.com/colinurbs/FramePack/
- https://github.com/colinurbs/FramePack/blob/main/multi_prompt.py
HiDream Model Enhancements
Combine HiDream with Flux.Dev as a refiner for improved photorealism, using low denoise (around 0.15). Experiment with Detail Daemon or specific Sigmas node setups to enhance skin texture and reduce the "plastic look" often seen in base HiDream outputs, particularly for portraits.
Links:
VRAM Optimization and Storage
For GPUs with limited VRAM (like 8GB or 16GB), use launch arguments like --medvram or --medvram-sdxl and --xformers. Install Tiled VAE extensions for upscaling tasks to prevent Out-of-Memory errors. Use symbolic links (symlinks) to store models on different drives while keeping your UI pointed correctly.
Links:
FLUX ControlNet FP8 Quantization
An FP8-quantized version of FLUX.1-dev-ControlNet-Union-Pro-2.0 is available. This significantly reduces memory requirements while maintaining quality, making complex ControlNet workflows (pose, depth, canny) feasible on consumer GPUs previously prone to OOM errors. Requires specific setup and models.
Links:
- https://civitai.com/models/1488208
- https://huggingface.co/ABDALLALSWAITI/FLUX.1-dev-ControlNet-Union-Pro-2.0-fp8
- https://github.com/al-swaiti/ComfyUI-OllamaGemini
Advanced Prompting and LoRA Training
Structure complex prompts using detailed layouts, negative tags, and token weighting like in some Sora examples. For LoRA training, tailor your dataset specifically (e.g., only portraits or a mix of portraits and full body) based on the desired output capabilities of the final LoRA.
