Full canvas workflow combining GGUF-quantized Qwen-Image-2.1 diffusion weights with the refusal-ablated Heretic text encoder. Eliminates morality guardrails and runs smoothly on 12GB VRAM.
Why CFG 2.2: Modern DiT architectures (like Qwen and Flux) use flow-matching dynamics instead of classic epsilon-prediction. Pushing CFG above 3.0 creates severe oversaturation, plastic skin, and harsh edges. Keeping CFG between 2.0 and 2.5 delivers cinematic skin textures without highlight burnout. Memory offloading: Host the quantized Q4_K_M DiT in GPU VRAM (4.6GB) and offload the Heretic text encoder into System RAM to stay safely within 8GB to 12GB graphics card limits.
RTX 4070 (12GB VRAM): ~11.1 GB active VRAM footprint. Generation time: 24 seconds per 1280x800 image. RTX 3060 (12GB VRAM): 38 seconds. Apple Silicon M3 (16GB): ~2.5 minutes using MPS backend.