Most anime-focused workflows rely on LoRAs or tag-heavy prompt tricks; Z-Anime takes a different route by providing a full fine‑tune family built on the Z-Image (S3‑DiT) base so you get consistent anime aesthetics, full negative-prompt support, and variants tuned for speed or final-render quality. The repo bundles model variants, quantized GGUF builds, AIO single-file checkpoints, and the VAE/text-encoder artifacts used for training, making it easier to run or adapt the model without stitching pieces together.
Key Capabilities
- Full fine-tune (not a LoRA): trained end-to-end on Z-Image Base (S3‑DiT 6B), so negative prompts and guidance behave predictably and the base variant preserves the widest creative range. This matters when you need reliable negative prompt conditioning or intend to further fine-tune/LoRA on top of a consistent checkpoint.
- Multiple quality/speed trade-offs: Base (highest quality), Distill-8-Step (balanced — 8 steps, CFG ~1.0), Distill-4-Step (ultra-fast — 4 steps). Use distills for rapid iteration and Base for final renders or LoRA training.
- Practical deployment formats: BF16/FP8 for CUDA/GPU workflows, GGUF quantized variants (~4–7GB) for low-VRAM or CPU/AMD inference, and AIO single-file checkpoints for simpler ComfyUI setups.
- Tooling & compatibility: includes a diffusers-format folder (ZImagePipeline.from_pretrained() compatible), recommended sampler/scheduler guidance, and an optional ComfyUI workflow. Recommended default settings and resolution presets are provided to reduce trial-and-error.
Who it's for & Trade-offs
Great fit if you: want an off-the-shelf anime-focused full fine-tune for high-quality image generation or further fine-tuning; need low-VRAM GGUF/AIO options to run on 8GB-class hardware; iterate quickly with 4/8-step distills. Look elsewhere if you: require non-anime photographic realism (this family is tuned for anime aesthetics), need production‑grade safety/filters out of the box (the model is noted as partially NSFW capable), or require tooling tied specifically to a different text‑encoder/architecture.
If you plan to train or chain further LoRAs, prefer the Base BF16 checkpoint; for fast batch generation or CPU-friendly runs, consider the GGUF or distill AIO variants. License: Apache‑2.0 (check the model page for attribution and commercial-use details).
