fullgreed (Z-Image base) — ComfyUI format

Diffusion model (DiT, denoiser only) on the Z-Image architecture (Alibaba Tongyi, S3-DiT), converted from a Draw Things .ckpt export (SQLite tensor store) to ComfyUI-compatible safetensors.

• fullgreed_f16.safetensors — FP16, ~12.3 GB, 453 tensors.

Usage (ComfyUI)

Place in ComfyUI/models/diffusion_models/ and load with Load Diffusion Model (UNETLoader), weight dtype fp16. This is the diffusion model only; you also need:

• Text encoder: Qwen3-4B (qwen_3_4b.safetensors) in models/text_encoders/ — its hidden size (2560) matches this model's caption embedder.
• VAE: Flux VAE (ae.safetensors, 16-channel) in models/vae/

Use a standard Z-Image workflow (same setup as the greed model).

Conversion notes

Key-remapped from the Draw Things layout to the Z-Image/NextDiT naming used by ComfyUI: fused q/k/v into attention.qkv, concatenated the 4-way ada_ln into adaLN_modulation.0, mapped the SwiGLU FFN to feed_forward.w1/w2/w3, and flattened the norm dims. context_refiner blocks have no adaLN (modulation off).

Verified: all 453 tensor keys and shapes match John2386/greed (the official Comfy-Org Z-Image diffusion model layout) exactly.

Note: this is a single-file diffusion checkpoint, not a diffusers pipeline repo (no model_index.json). Load it as a single file (ComfyUI UNETLoader, or from_single_file / manual single-file loading) — not DiffusionPipeline.from_pretrained("John2386/fullgreed").