ComfyUI: Krea 2 editing and local MiniMax H3 prompts

This tutorial covers Krea 2 image editing workflows in ComfyUI, then uses a local model to write prompts for MiniMax H3. The presenter says the Krea 2 edit model had not been released at recording time: the editing workflow uses an Edit LoRA and custom nodes instead.

Setup includes an INT8 or FP8 model, CLIP, a VAE, Pixaroma nodes, Krea 2 Enhancer and Krea 2 Identity. The examples start at 1 megapixel with dimensions snapped to multiples of 16. Reference Boost controls how closely edits retain the source identity. The presenter recommends trying values between 1 and 4; higher settings can prevent requested pose or outfit changes. Separate workflows cover portrait and landscape ratios, placing a character into a supplied background, and outfit transfer with a different LoRA. Small clothing details can disappear, and an outfit reference containing another person can confuse the result.

The Video Prompt Pixaroma node expands a short idea into a video prompt using a local LLM. It supports text-only input, a first frame, or first and last frames. Duration settings influence prompt length and action count rather than inserting a seconds value directly. The presenter reports better results with an 8B model than a 4B model and faster FP8 performance on his 40-series card. He recommends freeing VRAM before continuing into a video workflow, since keeping both models loaded can exceed available memory.