Advanced Parameters & Tuning

Kernel-level tuning, buffer pool sizing, and thread configuration

🚀 Advanced Parameter Control & Hybrid Scheduling

Termux-Diffusion exposes granular hardware control across CPU worker threads, Khronos Vulkan compute shaders, and memory-mapped buffers.

1. Production Parameter Taxonomy

Flag Type Default Engineering Function
--diffusion-model <path> Path Built-in Explicit path to 6.0B DiT or UNet weights (GGUF format)
--llm <path> Path Built-in Path to large language model text encoder (e.g. Qwen3-4B-Instruct)
--taesd <path> Path Built-in Path to ultra-lightweight Tiny AutoEncoder (taef1.safetensors) for ultra-fast VAE decode
--stream-layers Boolean Off Dynamically streams DiT layers across AXI bus to observe strict 1GB VRAM caps
--max-vram <target=size> String None Hard allocation ceiling on GPU heap (e.g. vulkan0=1) to prevent Android LMK kills
--params-backend <module=target> String Auto Binds model parameter storage (e.g. diffusion=cpu) for LPDDR5 residency
--backend <spec> String Auto Asymmetric engine assignment (e.g. clip=cpu,diffusion=vulkan0,vae=cpu)
--diffusion-fa Boolean Off Tiled Flash Attention kernel reducing attention memory complexity from O(N²) to O(N)
--vae-on-cpu Boolean Off Offloads VAE decoding to CPU cores, freeing GPU VRAM during final image restoration
--vae-tiling Boolean Off Tiled spatial latent reconstruction preventing 1.2GB decoding memory spikes

2. Documentation Overhaul Roadmap (차기 파라미터 전수 개편 계획)

Phase 1: Dynamic VRAM-Budget Autotuning Matrix (v1.8.1)

Interactive parameter calculator querying device SoC and physical RAM to automatically synthesize optimal CLI flags.

Phase 2: Mobile GPU Architecture Presets (v1.8.2)

Targeted configurations tuned for ARM Mali (G68/G78), Qualcomm Adreno (650/730/830), and Samsung Xclipse (RDNA).

Phase 3: Multi-Model Schema Unification (v1.9.0)

Unified JSON schema validation across Python SDK, Node.js SDK, and REST test harnesses with fail-fast assertions.