ARIA ZIBAria ZIB is a downloadable Z-Image Base derivative built for cinematic photography and stylized film stills: rich but controlled color, organic surface texture, soft halation, volumetric light, and deliberate photographic composition.This release is presented and validated as an SFW cinematic model. It was not developed or validated as an explicit-anatomy or NSFW-specialist checkpoint.THREE DOWNLOADS, ONE ARIA LOOKBF16 Quality — reference-quality rendering and the full Z-Image Base CFG workflow. Start with 28 steps and CFG 4.0.Turbo INT8 Metal — a faster, lower-memory option for a compatible Metal/INT8 stack. Start with 8 steps, CFG 1.0, Euler, simple scheduler, and model sampling shift 3.0.No additional Aria LoRA is required. The Aria cinematic adaptation is already baked into all three files. Only the Turbo INT8 version contains the 2603 8-step distillation at strength 0.8; do not load that distillation LoRA again.RAW INT8 ConvRot — direct quantization of the undistilled Aria ZIB BF16 master, without the Turbo distillation adapter. 5.87 GiB; row-wise INT8, ConvRot group 64, sensitive matrices retained in BF16.RAW starting preset: 16 steps, CFG 4.0, res_multistep sampler, bong_tangent scheduler, model sampling shift 3.0. For faster exploration, 12 steps produced usable architectural and portrait images in our local checks. If your ComfyUI installation does not offer that sampler/scheduler combination, start with Euler/simple at 28 steps instead. Euler/simple at 12 and 16 steps failed in our architectural test; those low-step Euler settings are not recommended.Quantization reduces the memory footprint; it is not step distillation. The reduced-step presets are empirical results from two prompts at 512 x 768, not a universal optimum or a claim of BF16 parity. Subject, seed, resolution, sampler and backend can change the outcome.REQUIRED SEPARATE COMPONENTSThese downloads contain the Z-Image diffusion transformer only. Install a compatible Qwen3 4B text encoder and the Z-Image ae.safetensors VAE separately, then load Aria through ComfyUI's Load Diffusion Model / UNETLoader in a normal split-component Z-Image Base workflow.OPTIONAL STYLE TOKENaria_cinemaThe token can reinforce the intended house style, but it is optional.HONEST LINEAGEAria ZIB was not trained from scratch and is not a full 6B-parameter fine-tune. It starts from Tongyi-MAI Z-Image Base. The owner trained a rank-32, alpha-16 cinematic LoRA for 600 optimizer updates on 299 owner-generated reference images, then fused the final adapter into Z-Image Base at strength 1.0.BF16 Quality is the fused Aria model in native ComfyUI Z-Image format.Turbo INT8 Metal additionally fuses Alibaba-PAI Z-Image-Fun 2603 8-step distillation at strength 0.8, then applies INT8 ConvRot quantization while retaining sensitive matrices in BF16. It is not the official Z-Image-Turbo model, and no BF16-parity claim is made.The training reference set consisted of images generated locally in ComfyUI by the model owner using Krea 2 / Krea 2 Turbo-family community derivative checkpoints, with optional third-party LoRAs in parts of the source archive. Those tools rendered reference images only: no Krea-family checkpoint or LoRA weights were copied or merged into these downloads.TESTED SETUPApple M4 Pro with 24 GB unified memory, ComfyUI 0.31.0, comfy-kitchen 0.2.28, PyTorch 2.13.0, and locally modified Apple-Silicon INT8/FP8 support based on ComfyUI-AppleSilicon-FP8 1.3.1.Other Macs, NVIDIA/AMD systems, and unmodified plugin builds have not been validated. Loading behavior, speed, memory use, and output can vary by backend. This release does not configure or promise Civitai on-site generation.KNOWN LIMITATIONS• Small or lengthy text inside images can be malformed.• Complex human anatomy and close-contact poses can fail.• The cinematic adaptation can be visually strong and varies with subject, seed, aspect ratio, and settings.• The distilled INT8 variant can change composition or reduce fine detail relative to BF16.• Third-party LoRA compatibility has not been comprehensively tested.RELEASE TERMSAria ZIB v1.0 is a free, non-commercial community release. There is no paid access, per-generation fee, donation goal, or tip request. Credit is required. Commercial use, resale, paid generation-service use, and distribution of merges or derivatives are not granted by this release page. Upstream components remain subject to their own licenses and notices.CREDITSTongyi-MAI — Z-Image BaseAlibaba-PAI — Z-Image-Fun-Lora-DistillHugging Face DiffusersComfyUIKrea 2 — referenced only for the lineage of locally generated training outputs; no Krea weights were merged

Aria ZIB

Initial Aria ZIB release with three downloadable diffusion-transformer files.BF16 Quality: start at 28 steps, CFG 4.0, standard Z-Image Base FlowMatch Euler.Turbo INT8 Metal: start at 8 steps, CFG 1.0, Euler, simple scheduler, model sampling shift 3.0.The Aria cinematic LoRA is already fused into all three files. Only the Turbo INT8 version has Alibaba-PAI 2603 8-step distillation fused at 0.8. Do not load either LoRA again.Requires a separate Z-Image-compatible Qwen3 4B text encoder and ae.safetensors VAE. Tested locally on Apple M4 Pro / 24 GB. No on-site generation configuration or BF16-parity claim.Added September 22, 2026: Aria ZIB RAW INT8 ConvRot, directly quantized from the original undistilled BF16 master. Existing BF16 and Turbo downloads are unchanged.RAW INT8 ConvRot — direct quantization of the undistilled Aria ZIB BF16 master, without the Turbo distillation adapter. 5.87 GiB; row-wise INT8, ConvRot group 64, sensitive matrices retained in BF16.RAW starting preset: 16 steps, CFG 4.0, res_multistep sampler, bong_tangent scheduler, model sampling shift 3.0. For faster exploration, 12 steps produced usable architectural and portrait images in our local checks. If your ComfyUI installation does not offer that sampler/scheduler combination, start with Euler/simple at 28 steps instead. Euler/simple at 12 and 16 steps failed in our architectural test; those low-step Euler settings are not recommended.Quantization reduces the memory footprint; it is not step distillation. The reduced-step presets are empirical results from two prompts at 512 x 768, not a universal optimum or a claim of BF16 parity. Subject, seed, resolution, sampler and backend can change the outcome.

Files with SHA256: a082821a046a15a1f865f1694552f543ae0a49f747262aa8e83445cc9d26add3