Gemma4 Prompt Engineer - Early access -

A developer known as 'Gemma4Prompt' has released early access to Gemma4 Prompt Engineer, a custom node for the popular ComfyUI workflow manager. The tool leverages the Gemma 4 31B 'abliterated' model—a community-modified version that removes restrictive content filters—running entirely locally via llama-server. This approach eliminates cloud API keys and costs, keeping all prompt generation private on the user's machine. The node specifically tailors outputs for six distinct AI media models: LTX 2.3 (cinematic paragraphs), Wan 2.2 (motion-first), Flux.1 (natural language), SDXL 1.0 (booru tag style), Pony XL (e621 format), and SD 1.5 (weighted classic), ensuring each prompt format matches the target model's optimal input structure.

Key features include 48 detailed environment presets covering styles from Wes Anderson to K-drama, a PREVIEW/SEND mode for workflow control, character locking for LoRA consistency, screenplay formatting for LTX 2.3, and dialogue injection for video prompts. The setup process is automated for Windows users via a batch file that handles llama-server installation and model downloading to C:\models\. Current limitations restrict it to Windows/CUDA systems with substantial VRAM (~20GB for the 31B Q4_K_M quantized model). The developer notes this is early-access software with ongoing tuning, particularly around interrupt behavior and server management improvements.