Z-IMAGE IMG2IMG for Characters V5: Best of Both Worlds (workflow included)
New ComfyUI workflow combines Z-Image Base's prompt adherence with Turbo's quality for 1536px character images.
A new, optimized workflow for generating high-quality AI character images has gone viral in the Stable Diffusion community. Created by developer RetroGazzaSpurs and dubbed 'Z-Image IMG2IMG for Characters V5,' it solves a common problem by strategically chaining two distinct Z-Image models. The process first uses a Z-Image Base model combined with a character-specific LoRA (like those from Malcolm Rey's collection) to establish strong prompt adherence and variety. The output is then passed through a Z-Image Turbo model for a final quality refinement, delivering what the creator calls 'the best of both worlds.'
The workflow is shared as a ComfyUI JSON file on Hugging Face, making it easily importable for users. A key improvement in V5 is optimization for lower VRAM systems, broadening its accessibility. The creator notes that while 1280px resolution yields great results, a well-trained LoRA can perform even better at 1536px. An important caveat is that the workflow currently relies on SAM3 nodes in ComfyUI, which have a known bug; however, a quick two-minute manual fix is provided via a GitHub issue link until the developer releases an official patch.
- Workflow chains Z-Image Base (for prompt adherence) and Z-Image Turbo (for quality refinement) in a single ComfyUI pipeline.
- Optimized for Malcolm Rey's LoRA collection and lower VRAM systems, supporting outputs up to 1536px resolution.
- Available as a JSON file on Hugging Face, but requires a manual fix for currently broken SAM3 nodes in ComfyUI.
Why It Matters
Provides a reproducible, high-quality pipeline for AI artists and developers creating consistent character assets, streamlining a complex technical process.