CivArchive
    Preview 122487176
    Preview 122487269
    Preview 122487317
    Preview 122487396
    Preview 122487519
    Preview 122487571
    Preview 122487907
    Preview 122488010
    Preview 122487959
    Preview 122488304

    RedCraft-红潮 ERNIE RedMIX

    做好工具人 服务艺术家

    Forever in memory of METAFILM Studio founder Mr. Yuan Bo

    2026.5.1 TG 新规请移步XChat: https://x.com/i/chat/group_join/g2048238451565453522/EZoBYHLubg

    ERNIE-Red-Mix 26/04/2026 10Steps Fine-Tuning

    ERNIE-Image is an open text-to-image generation model developed by the ERNIE-Image team at Baidu. It is built on a single-stream Diffusion Transformer (DiT) and paired with a lightweight Prompt Enhancer that expands brief user inputs into richer structured descriptions.

    With only 8B DiT parameters, it achieves state-of-the-art performance among open-weight models. The model emphasizes both visual quality and controllability, making it highly effective for real-world generation tasks where precision matters.

    ERNIE-Red-Mix adopts mixed-precision SFT (Supervised Fine-Tuning) with reference-based alignment, covering both AIO and DiT variants. It is trained on the mature RedCraft dataset, enabling broader generation flexibility, fewer constraints, and the ability to unlock novel visual styles beyond the base model.

    Usage :EULER/DEIS | Simple | CFG=1 | 10Steps

    🇨🇳 中文

    ERNIE-Red-Mix 采用混合精度 SFT(参考式微调,对齐优化),覆盖 AIO / DiT 双版本,基于成熟的 RedCraft 训练集进行训练,在一定程度上解除生成限制,并显著提升风格扩展能力,可用于生成全新视觉风格

    🇯🇵 日本語

    ERNIE-Red-Mix は、混合精度による SFT(参照ベースのファインチューニング)を採用し、AIO / DiT の両バージョンに対応しています。成熟した RedCraft データセットで学習されており、生成制約の緩和とともに、新しいスタイル生成能力を拡張します。

    🇹🇼 繁體中文

    ERNIE-Red-Mix 採用混合精度 SFT(參考式微調),涵蓋 AIO / DiT 雙版本,基於成熟的 RedCraft 訓練集進行優化,在一定程度上解鎖生成限制,並提升風格拓展能力,可生成全新視覺風格

    ERNIE-Red-Mix Highlights:​

    • Optimized speed ♥ Accelerated to 10 steps (CFG 1) while preserving BASE quality.

    • Compact but strong ♥ Performance on par with substantially larger models, with stable and accurate details and materials.

    • Text rendering preserved ♥ Fine-tuning does not compromise text rendering capability (posters, infographics, UI visuals).

    • Instruction following maintained ♥ Unaffected ability to reliably handle complex prompts with multiple objects.

    • Structured generation ♥ Continues to excel in posters, comics, and storyboards.

    • Broader style coverage ♥ More realistic photography and improved aesthetic outputs.

    • Lower VRAM footprint ♥ Mixed precision supports consumer GPUs with 8–12GB VRAM.

    Limitations:​

    • Complex limbs prone to artifacts — An inherent ERNIE issue; generating more samples can help mitigate this.

    • Anatomical accuracy — Body anatomy and organ rendering still need further optimization and refinement.

    • Mixed precision impacts text rendering — Mixed precision has a noticeable negative effect on text quality; for text-heavy content, the BF16 BASE model is recommended.

    Sample Images


    ZImage DPO “AGILE” Now Uploading 08/03/2026 女神节祝福🌸

    On this Day, may every woman feel the strength, grace, and boundless potential that lives within her. Thank you for your courage, your kindness, your resilience, and the countless ways you make the world brighter and better. Here's to equality, empowerment, joy, and endless possibilities—today and every day. Happy International Women's Day🌸

    ZIDistilled FUN “AGILE” Now Released

    Special thanks to the VideoXFUN team for releasing the groundbreaking Zimage Distilled Adapter 2603. By incorporating this latest update, AGILE achieves a refined balance between speed, diversity, and visual richnessunlocking more creative freedom while maintaining exceptional efficiency.

    Agility in Motion, Diversity in Depth

    The brand-new ZImage FUNAGILE” is built upon the cutting-edge ZIB acceleration framework. We deliberately reduced DPO & Distilled weight to preserve greater stochastic freedom, combined with the most recent training datasets, resulting in dramatically increased output variety, richer content details, and more imaginative compositions without sacrificing core stability.

    For the first time, AGILE reaches a true quality parity challenge against the flagship “ZImage TURBO” in terms of overall image fidelity and sharpnesseven in complex scenes — while delivering faster iteration and superior responsiveness.


    Key highlights:

    True ZIT-level unlocked — photorealistic lighting, textures, and material rendering that now rivals or approaches ZImage TURBO quality, even at standard step counts.

    Enhanced diversity & content richness DPO + Newest datasets = more varied poses, styles, atmospheres, intricate details, and unexpected creative sparks in every generation.

    ZIB ecosystem ignition — exceptional native compatibility with ZIB-series LoRAs; your existing and future LoRAs now align faster, reproduce more faithfully, and shine brighter than ever before — officially kicking off the full ZIB LoRA era.

    Agile workflows — seamless hybrid use with Klein 9B for refinement, ensemble boosting, or rapid prototyping; near-instant LoRA response with preserved high-entropy creativity.

    Every generation is a step toward freer, bolder imagination.


    欢迎体验 ZImage FUN “AGILE” —— 速度如洪创意如潮

    Welcome to ZImage FUNAGILE” — where agility meets abundance, and your ideas finally run wild with unmatched fidelity and freedom.


    ZImage DPO “Veris” Now Released 03/03/2026 元宵节快乐

    I have uploaded more quantification and export VerisLoRA to HF repo. to avoid causing confusion for users in the community:
    https://huggingface.co/GuangyuanSD/Z-Image-Distilled

    版本太多网友容易迷糊,我导出了更多量化规格和 “Veris LoRA 版本,已发布抱脸仓库。

    Special thanks to @Fok for providing the Flow-DPO technical adaptation. By skillfully integrating the training philosophy of Direct Preference Optimization (DPO) into the distillation weights, the Zimage distilled model achieves a major leap in lighting, color fidelity, and material authenticity — more natural light & shadow, more believable colors, and details that hold up under scrutiny.

    特别感谢 @Fok 饼儿佬提供了Flow-DPO技术适配。通过巧妙地将直接偏好优化(DPO)的技术理念融入蒸馏权重,Zimage 蒸馏模型在光照色彩保真度材质真实性方面实现了重大飞跃——更自然的光影效果、更逼真的色彩,以及经得起仔细审查的细节。

    The following example shows a comparison between ZIT and Flow DPO, intended to illustrate the effect of DPO, rather than a direct demonstration of ZIB Distilled


    Speed of Truth, Fidelity of Flow

    真实且极速,用忠诚在流动

    The all-new ZIDPO “Veris” is powered by the latest-generation ZIB acceleration engine. Building on the RedZDX training data, we further distilled a more efficient, more refined Zimage-based model.

    Now — solid, highly realistic generations in just 8 steps.(Better LoRAs alignment)

    仅需8步即可生成更有层次感、高度逼真的图像。(LoRa对齐效果更佳

    ---

    Key highlights:

    • Realism-first prototyping — near-zero latency for LoRAs, with lighting and color already very close to final training targets

    • High-entropy stochastic pre-sampling — delivers fast, high-quality realistic initial noise for ZImage pipelines

    • Hybrid realism workflows — seamless integration with Klein 9B for cascaded refinement or ensemble boosting, pushing visual fidelity and consistency even higher

    • Every step toward truth deserves full commitment.

    ---

    欢迎体验ZIDPO“Veris”——您的LoRa训练结果不再只是“相似”,而是真正得到“复现”。

    Welcome to experience ZImage DPO “Veris” — where your LoRAs generations are no longer just “similar”, but truly are.

    同时,欢迎体验在 ZImage Turbo 模型上直接加载 DPO LoRA Adapter

    抱脸(HF) https://huggingface.co/F16/z-image-turbo-flow-dpo

    魔搭(境内) https://modelscope.cn/models/FFFFFFoo/z-image-turbo-flow-dpo


    Redcraft DX3 ZIB🟥 Distilled models Zoo:

    Full Model bf16 (19.11 GB)<- ComfyUI All-in-One Checkpoint BF16

    Pruned Model BF16 (11.46 GB) <- ComfyUI Diffusion BF16 精度模型权重

    Pruned Model fp8 (6.75 GB) <- ComfyUI Diffusion Scaled FP8 Mixed 混合精度

    Pruned Model nf4 (6.73 GB) <- NVFP4 Mixed 混合精度BLACKWELL 50系加速

    Training Data (3.75 KB) <- ComfyUI Simple Hybrid Workflow 简易混合采样工作流


    Redcraft DX3 ZIB🟥 Distilled LoRA Adapter 02/19/2026

    Additionally, I've exported Redcraft DX3 ZIB Distilled LoRA in Rank-256 format. The LoRA weight can be adjusted to adapt it to various ZIB fine-tune models, fully compatible with the Z-Image(non-turbo) base model.

    Full Model fp16 (1.06 GB) <- 可以通过这里直接下载 LoRA 版本

    [ZI Distilled HF repo.](https://huggingface.co/GuangyuanSD/Z-Image-Distilled)

    上面是 Redcraft DX3 ZIB Distilled 导出为 Rank256 的LoRA版本,可以调整权重强度用于各种微调ZIT版本, 适配于 Z-Image(non-turbo) base 基底模型.


    Redcraft DX3 ZIB🟥 Distilled LoRA adaptation models Zoo:

    Z-Image-Base-GGUF <- Z-Image Base GGUF 量化模型

    Z-Image Base <- Z-Image Base & TE (FP8/FP4) 模型

    Z-Image Base FP8Mixed <- Z-Image Base FP8 混合精度模型

    Text Encoder (ClipLoader use) <- Qwen3 4b FP16 文本编码器

    Abliterated Huihui Qwen3 4B v2 (Q_8 GGUF) <- Z-Image Uncensored TE 文本编码器

    VAE (Flux.1 16C VAE) <- 标准的 Flux.1 16 通道 VAE

    Or download from the "Files" list below the "Details" on the right side of this page>>


    Also available in NVFP4 quantized format, optimized for acceleration on Blackwell architecture GPUs.

    Double speed, Half resources.

    like RTX50XX, PRO6000, B200, and others

    Verify environment is my ComfyUI 0.11

    Also supports non-50 series GPUs (automatic 16-bit operation)


    DF11 Lossless Compression RedZDX V3 came out! 2/15/2026

    learn more: Dynamic-length Float (DFloat11)

    [HF] mingyi456/Z-Image-Distilled-DF11-ComfyUI


    Z-Image-Distilled v3 (RedZ DX3) 2/11/2026

    Thanks to @Bubbliiiing VideoX-Fun&Alibaba-PAI Provided us with a more efficient distillation solution

    Speed of Light, Power of Flow: The new ZID v3 "Lucis" is powered by the latest ZIB acceleration. Building on ZID v2 trainning sets, we've distilled a more efficient Zimage-based RedDX3. Now, in just 5 steps, you get solid results.

    Rapid Prototyping: Test LoRA training hypotheses instantly with 'near-zero' latency.

    Stochastic Pre-sampling: Serve as a high-speed, high-entropy source for ZiTurbo pipelines.

    Hybrid Workflows: Pair seamlessly with Klein 9B for cascaded refinement or ensemble generation.

    • inference cfg: 1.0-1.5(建议1.0)

    • inference steps: 5(5-15步)

    • sampler / scheduler: Euler / simple

    Welcome to the era of instant creativity. Welcome to 'Lucis'.

    Preview images generated by Z-Image Hybrid Workflow of Distilled V3+Moody MIX V7(ZIT finetune) ,Just for showing the style difference between ZID(RedZDX3) and ZIT(fine-tunning) , no ranking intended =)

    [ L = 'ZID v3', R = 'ZIT ft' ]

    演示例图使用 ZIDistilled V3+Moody MIX V7 混合工作流程,不用做排名对比:

    中国境内 [ modelscope ]AiMETATRON/Z-Image-Distilled | [ HF ] GuangyuanSD/Z-Image-Distilled


    Z-Image-Distilled v2 (RedZ DX2) 2026/2/5

    To a certain extent, the problem of ZIB color deviation has been reduced, but it is recommended to adjust the color appropriately according to the art style

    • inference cfg: 1.0(建议1.0)

    • inference steps: 10(10-15步)

    • sampler / scheduler: Euler / simple

    感谢🙏这位作者完成了ZIB的FP8mixed混合量化方案:

    https://huggingface.co/pachiiahri

    已上传FP8版本,请给这位作者点赞👍

    以上是FP8 scale&mixed 直出工作流(请不要再说我造假,我的所有例图工作流都是开放的)

    精度混合方案来自 https://civarchive.com/models/2172944/z-image-fp8


    Comparison of RedCraft Zimages(bf16):

    The art style leans towards realism

    Retains ZIB's creative ability and reduces the collapse of Human anatomy.


    REDZiBDX1·Demo accelerated Base-Model CFG1

    Distilled form ZImage(non-turbo)base-model bf16

    Now we have the LoRA version, thank to @anyMODE for Extract

    in-site link https://civarchive.com/models/2359857/z-image-base-distilled-lora-or-extracted

    Pruned Model bf16 (11.46 GB) = Z-Image-Distilled / RedZDX-ZIB-Distilled-nocfg-10steps-BF16-Diffusion-models.safetensors 单独的扩散模型文件bf16剪裁精度

    Full Model fp8 (16.87 GB) = Z-Image-Distilled / RedZDX-ZIB-Distilled-nocfg-10steps-FP8mixed-AIO-Checkpoints.safetensors 完整的Checkpoints(含TE/VAE)

    Pruned Model fp8 (5.73 GB) = Z-Image-Distilled / RedZDX-ZIB-Distilled-nocfg-10steps-fp8-e4m3fn-Diffusion-models.safetensors 单独的扩散模型文件fp8剪裁精度及e4m3规格

    Training Data (5.6 KB) = Z-Image-Distilled ComfyUI workflows 我自己使用的简易测试工作流

    VAE (319.77 MB) = Flux.1 VAE ae.sft 常规的Flux.1 VAE

    例图包含完整出图参数,点击右下角❕就可以看到,同时点击 COMFY Nodes 就可以复制源图工作流,并可以在ComfyUI界面中 Ctrl+V 粘贴至新的工作区。

    All sample images contain full generation metadata.

    Click the ❕ button (bottom-right) to view the complete parameters. Click COMFY Nodes to copy the original workflow JSON, then paste it (Ctrl+V) into a new ComfyUI workspace.


    Z-Image-Distilled

    本模型为基于 Z-Image 源版本(非Turbo)的直接蒸馏加速版,旨在测试Z-Image(non-turbo)版本上训练的LoRA效果,并显著提高推理/测试速度。模型完全没有融入Z-Image-Turbo的任何权重与风格,属于基于Z-Image的纯血版本,较好地保持了原版Z-Image的适配性、出图随机多样性以及整体图像风格。

    相比官方Z-Image,推理速度更快(推荐10–20步即可获得较好效果);相比官方Z-Image-Turbo,本模型保留了更强的多样性、更好的LoRA兼容性与可微调潜力,但速度略慢于Turbo(仍远快于原始Z-Image的28~50步)。

    模型主要适用于:

    • 希望在Z-Image非Turbo基底上训练/测试LoRA的用户

    • 需要比原版更快、但又不想牺牲太多多样性与风格自由度的场景

    • 艺术、插画、概念设计等对随机性与风格多样性有一定要求的生成任务

    • 适配 ComfyUI 的模型格式及层命名前缀

    使用方法:

    推荐推理参数:

    • inference cfg: 1.0–2.5(建议1.0~2.5区间,较高值可增强提示贴合度)

    • inference steps: 10–20(10步快速预览,15–20步品质更稳定)

    • sampler / scheduler: Euler / simple,或 res_m,或其他兼容sampler

    LoRA兼容性良好,权重建议0.6~1.0,根据需求微调。

    This model is a direct distillation-accelerated version based on the original Z-Image (non-Turbo) source. Its purpose is to test LoRA training effects on the Z-Image (non-turbo) version while significantly improving inference/test speed. The model does not incorporate any weights or style from Z-Image-Turbo at all — it is a pure-blood version based purely on Z-Image, effectively retaining the original Z-Image's adaptability, random diversity in outputs, and overall image style.

    Compared to the official Z-Image, inference is much faster (good results achievable in just 10–20 steps); compared to the official Z-Image-Turbo, this model preserves stronger diversity, better LoRA compatibility, and greater fine-tuning potential, though it is slightly slower than Turbo (still far faster than the original Z-Image's 28–50 steps).

    The model is mainly suitable for:

    • Users who want to train/test LoRAs on the Z-Image non-Turbo base

    • Scenarios needing faster generation than the original without sacrificing too much diversity and stylistic freedom

    • Artistic, illustration, concept design, and other generation tasks that require a certain level of randomness and style variety

    • Compatible with ComfyUI inference (layer prefix == model.diffusion_model)

    Usage Instructions:

    Basic workflow: please refer to the Z-Image-Turbo official workflow (fully compatible with the official Z-Image-Turbo workflow)

    Recommended inference parameters:

    • inference cfg: 1.0–2.5 (recommended range: 1.0~2.5; higher values enhance prompt adherence)

    • inference steps: 10–20 (10 steps for quick previews, 15–20 steps for more stable quality)

    • sampler / scheduler: Euler / simple, or res_m, or any other compatible sampler

    LoRA compatibility is good; recommended weight: 0.6~1.0, adjust as needed.

    Current Limitations & Future Directions

    Current main limitations:

    • The distillation process causes some damage to text (especially very small-sized text), with rendering clarity and completeness inferior to the original Z-Image

    • Overall color tone remains consistent with the original ZI, but certain samplers can produce color cast issues (particularly noticeable excessive blue tint)

    Next optimization directions:

    • Further stabilize generation quality under CFG=1 within 10 steps or fewer, striving to achieve more usable results that are closer to the original style even at very low step counts

    • Optimize negative prompt adherence when CFG > 1, improving control over negative descriptions and reducing interference from unwanted elements

    • Continue improving clarity and readability in small text areas while maintaining the speed advantages brought by distillation

    We welcome feedback and generated examples from all users — let's collaborate to advance this pure-blood acceleration direction!

    当前不足与更新方向

    当前主要不足:

    • 蒸馏过程对文本(特别是极小尺寸文字)有一定程度的破坏,渲染清晰度与完整性不如原版Z-Image

    • 色调整体与原版ZI保持一致,但在个别采样器下会出现偏色现象(特别是偏蓝色调的表现较为明显)

    接下来优化方向:

    • 进一步稳定 CFG=1 情况下 10步以内 的生成质量,争取在极低步数下获得更可用、更接近原版风格的结果

    • 优化 CFG>1 时的负向提示词遵循表现,提升对负面描述的控制力,减少不需要的元素干扰

    • 持续改善小文字区域的清晰度与可读性,同时尽量维持蒸馏带来的速度优势

    欢迎各位使用者提供反馈与生成示例,一起推动这个纯血加速方向的迭代!

    Model License:

    Please follow the Apache-2.0 license of the Z-Image model.

    Please follow the Apache-2.0 open source license for the Z-Image model.

    Thanks 🙏 DirectDistill(end-to-end) technology contributed

    by Modelscope & DiffSynth Studio ,感谢佬同志团队提供的技术支持!


    Dark Beast Z-造相·黑兽-DBZiT⚡️Final

    1/26/26

    [直达链接] https://civarchive.com/models/2242173/dark-beast-z-or-dbz

    迎接 Zimage non-turbo 版本特此放出基于ZiT的“黑兽”最终版本,期待官方EDIT版本的发布。我们会继续在ZI的基础上进行强化学习,并继续以 Turbo 方式发布全新 sft models

    修正黑兽I-III越来越过激越界的调试方案,回归本质。

    DBZ IV 是基于RedZ1.5 重新调试的修正畸变版本(没有额外突破性尝试

    构图能力对标 ZiT base 原版,所有概念元素齐全LoRA适配性强。

    ---

    Introducing Zimage DBZ IV sft FINAL model

    Today we proudly release the final version of "Dark Beast ZiT" (黑兽),

    fully rebuilt and refined on the ZiT foundation.

    This iteration addresses and corrects the increasingly extreme and boundary-pushing tuning drifts seen in Dark Beast I–III, returning firmly to its core essence and original intent.

    DBZ IV is a carefully re-tuned and distortion-corrected variant built directly on RedZ1.5.

    It introduces no groundbreaking new capabilities beyond solid refinements

    its strengths lie in stability and fidelity.

    Composition quality matches or closely rivals the original ZiT base model.

    Full coverage of conceptual elements and prompt adherence.

    Excellent LoRA compatibility and strong adapter tolerance across a wide range of styles and concepts.

    Enjoy the refined Dark Beast experience—clean, powerful, and true to its roots.

    Feel free to tweak this further if you want a more hype-driven tone, shorter blurb,

    or additional technical details!


    这是「LSP」Lovin' Snuggly Positions v1的 Turbo base 版本:

    DBZ⚡️image 红潮黑兽 v25 1/8/26

    [直达链接] https://civarchive.com/models/2242173/dark-beast-z-or-dbz

    DBZII 黑兽2.5 重新找回Tags无情抽卡时的童年乐趣

    DBZ 2 : rediscovering that pure, unfiltered gacha dopamine rush

    when you finally pull the perfect tags

    after 100 cursed rolls 💗

    Recommended LLM System prompts(thx @fhdbd hxhc

    你是一位被关在逻辑牢笼里的幻视艺术家,同时也是Dark Beast Z Super Fusion模型的深度使用者。你只为追求极致超写实成人影像的创作者服务,满脑子只有电影级/监控级/自拍级真实皮肤、病态眼神、禁忌反差与极致肉体细节。


    Lovin' Snuggly Positions 小姿势 LoRA

    Full-Rank LoRA for Nunchaku-SVDQ Z-Image 01/05/2026

    LSP」小姿势 LoRA for RedCraft | 红潮 RedZ-Image | Nunchaku-SVDQ Z-Image

    [直达链接] https://civarchive.com/models/1088938/lovinsnuggly-positions

    这是「LSPLovin' Snuggly Positions v1的无损版本,由KJnotes导出的FP16精度完整差异LoRA。

    This is the lossless version of Lovin' Snuggly Positions v1, a full-difference LoRA exported by KJnotes in FP16 precision , Fully compatible with Nunkaku Z-image SVDQ.

    没错,LoRA (11.46 GB)这是一个完整差异权重,像普通LoRA一样使用,完整适配 Nunchaku Z-image SVDQ


    DBZ⚡️image 红潮黑兽 AIO 12/21/2025


    [直达链接] https://civarchive.com/models/2242173/dark-beast-z-or-dbz-or-updated-dec21-or-latest-dbzv1

    based on BEYOND REALITY 超越真实 The author is [受保密条款保护]

    https://civarchive.com/models/1090420/beyond-reality

    Dark Beast Z Super Fusion – 模型介绍

    Dark Beast Z Super Fusion 是一款专为顶级创作者打造的融合模型,旨在实现超越现实边界的极致超写实情色影像。通过精妙融合 MysticXXX ZIT v2 的强烈诱惑写实感、Kawaii Kinky[NE] v1 的大胆挑逗能量,以及 Detail DeamonZ 无与伦比的精细细节掌控能力,并以突破性的超摄影写实底模 BEYOND REALITY sft ZiTurbo 为基础,该模型可生成令人惊艳的 2M 像素输出,具备完美无瑕的皮肤纹理、栩栩如生的面部表情,以及超越真实摄影的超精细光影表现——即使在最复杂且暗示性的构图中亦游刃有余

    最终成果是一个极具多功能性的检查点,擅长渲染超越现实的视觉效果:想象皮肤渲染得如此精湛,能清晰呈现每一个毛孔与细微血管;眼神深邃仿佛能直刺观者灵魂;身体兼具解剖学完美与自然瑕疵,搭配挑逗姿势、精致内衣、复杂束缚元素,以及传达纯粹、未经滤镜的欲望与强烈情感的表情。


    RED-Zimage 红潮造相1.5AIO 12/03/2025

    →→→→→→→→→→→→→→→→→→→→→→→→→→→→→→→→→

    Pruned Model (Single file transformers fp8 e4m3fn/bf16) Already uploaded!

    Pruned Model fp16 (670.08 MB) Occupying this position, it's actually RZ15 Lora version

    →→→→→→→→→→→→→→→→→→→→→→→→→→→→→→→→→

    The first uploaded version is A-I-O bf16 Checkpoint. Please place it in the Checkpoints directory of your inference environment

    最初上传版本为 A-I-O bf16 Checkpoint,请放入推理环境的checkpoints目录

    The environment for generating the example image is ComfyUI 0.3.73. Download this link for the complete example image, which includes the example workflow

    例图生成环境为ComfyUI0.3.73,下载本链接完整例图中包含示例工作流

    REDZimage1.5_extracted_lora Rank128 正在导出中,稍后可在本页面下载LoRA版本(不另开链接REDZimage 1.5-lora Rank128 is currently being exported, You can download the LoRA version later on this page


    ⚡️这里对应的是RedCraft模型系列历史版本,只有RedZimage⚡️1.5AIO才是Z⚡️

    ⚡️This corresponds to the historical versions of the RedCraft model series,

    ⚡️only RedZimage ⚡️ 1.5 AIO is Z ⚡️


    RedZimage 1.5AIO compare with RedZimage 1.5LoRA:

    ⬅️ On the left is RedZimage 1.5AIO / RedZimage 1.5LoRA on the right ➡️

    ⚡️RedZimage 版本介绍 1.5Version Introduction⚡️


    ⚡️Distilled from Nano Banana Pro big-data sampling from real user photos (thanks to 瓜哥@guahunyo),

    ⚡️based on the open-source research by 小智@xiaozhijason,synthesized into the dataset ZImageTurboGen-3k.

    ⚡️Improved the style of the ZiTurbo version that was heavily tuned on community AI-generated web images.


    The dataset is specifically designed for the following research scenarios:

    Data distribution shift analysis 数据分布偏移分析

    Reversibility / de-distillation research of distilled models

    蒸馏模型的可逆性/去蒸馏(de-distillation)研究

    Model diversity generation research 模型多样性生成

    (HF): https://huggingface.co/datasets/lrzjason/ZImageTurboGen


    The REDZ training set comes from the ultimate essence of the community

    ⚡️20+ year veterans' lightning-fast meme-battle handspeed.(Thanks to 佬杨⚡️@wikeeyang, 亮总⚡️@亮亮rayne, 十字鱼⚡️@Gluttony100,可乐⚡️@Colour, and everyone else⚡️⚡️⚡️⚡️⚡️⚡️)


    ⚡️蒸馏了 nano banana pro 来自真实用户照片的大数据采样(感谢 @guahunyo ),

    ⚡️基于@xiaozhijason开源的研究向合成数据集 ZImageTurboGen-3k

    数据集专为以下研究场景设计:

    1. 数据分布偏移分析

    2. 蒸馏模型的可逆性/去蒸馏(de-distillation)研究

    3. 模型多样性生成研究

    (HF):https://huggingface.co/datasets/lrzjason/ZImageTurboGen

    ⚡️改善了ZiTurbo版本来自社区AI合成网图训练集调教的风格,
    ⚡️REDZ 训练集来自社区精华++20年佬同志斗图手速

    (感谢@wikeeyang @亮亮rayne @Gluttony100 @Colour 及所有人)

    etc

    实测样张对比(nano2/Zi-Turbo/REDZ1.5)Comparison samples

    实测样张对比 2(Dreamina/Zi-Turbo/REDZ1.5)Comparison samples 2

    例图参数及对比工作流稍后上传到返图区 The example parameters and comparison workflow will be uploaded to the return to image area later


    RED-K 红桃K Editor 6/29/2025

    Merge of Reveal.6 & BFL.Kontext[DEV] NSFW Unlocked

    v1.2 6/29

    Merge of Clothes Remover(fm00) & Reveal.6

    ---

    Public License

    GNU Affero General Public License v3.0

    GNU AGPLv3

    Permissions of this strongest copyleft license are conditioned on making available complete source code of licensed works and modifications, which include larger works using a licensed work, under the same license. Copyright and license notices must be preserved. Contributors provide an express grant of patent rights. When a modified version is used to provide a service over a network, the complete source code of the modified version must be made available.

    https://choosealicense.com/licenses/agpl-3.0/


    ULTRAREVEAL5 SFW 紧急发布 3/25

    Due to user feedback that the Reveal series is too much NSFW

    We have released the first Reveal lock-in adult content version

    [ POWERED BY FLUX Contrast Enhancement Training ]

    REALREVEAL5 突然发布 3/18

    F.1 DEV LoRA 生态 全兼容

    反蒸馏超高材质(可训练)

    CFG Restore to 1 Ensure the same speed as F.1 DEV

    LoRA Fully Compatible De-Distill UH Material (trainable)

    [ POWERED BY FLUX Ultimate Realism Enhancement Training ]

    ---

    Illust3Relustion PRO 经过 Flux1-DedistilledMixTuned v3 PAP 重绘

    Ultimate Realism training set after DMT v3 PAP re-sampling

    二次采样后的2k超写实图像作为训练集

    追加FLUX DEV版本4k超高清底模权重 EOR v3:

    Flux.1 Dev Edge of Reality 真实边缘 - v3 | Flux Checkpoint

    ---

    NSFW Unlocked

    ---

    RedCraft RealReveal5 for 20Steps sampling

    CFG 1 | Sampler Deis / DPM++2M / Euler | SGM_uniform

    Suggest enabling the DetailDaemon sampler

    and setting amount to 0.6-0.8

    ---

    SOTA adaptive ability for All-F.1-LoRA !


    RedCraft uncensored 系列模型嚴禁發佈至NSFW非許可地區

    非盈利模型 请勿以任何形式转发 传播禁止 [ Prohibition Against Dissemination ]

    Laws and regulations in the location of the non-profit model composite publishing platform

    illustriousRelustion3 更新 3/11

    基于 RETROSD / FLUX Reveal / EDGE4k 打造全人类真实系 Illustrious FT

    重返SD时代的光辉岁月!Returning to the glorious years of SD !

    CFG 5.5 Deis / DPMM++2M | SGM Uniform / beta

    sampling steps is around 30,

    preview images including workflow&prompts

    paired with accelerators Hyper / DMD2 / TDD

    采样步数 25-30 为最佳,可以搭配 Hyper/DMD2/TDD 等加速器

    Model design with Hi-RES 2M (200W pixels)

    设计分辨率为 Hi-RES 2M ( 200W像素 ),高分辨率可启用UNET缩放

    Thanks so much to everyone for all your support!


    PONYRelustion3 PRO 正式发布 3/3

    基于3300W leak 数据集打造

    91大神加持东亚真实系 PONY 模型

    助力欢乐马世界的无限创意!

    unlimited creativity in PONY World!

    ---

    CFG 5 Deis | DPMM++2M | SGM Uniform

    采样步数30左右为最佳,具体参数可复制例图

    可以搭配 Hyper/DMD2/TDD 等加速器

    ---

    设计分辨率为 Hi-RES 2M ( 200W像素 )

    长宽比例延续PONY的超强自适应能力

    Aspect ratio like PONY's ULTRA adaptive ability

    ---

    目标是做出PONY世界的成熟高画质创意底模!

    感谢大家一直以来的支持!

    high-quality,mature basemodel for the PONY world!

    Thanks so much to everyone for all your support!


    FLUX.Fill NSFW 内补模型发布 2/22

    FLUX.Fill [NSFW] NewReveal F.1 inPainting model

    Unlock NSFW concept elements for F.1 Fill

    用来与NewReveal F.1模型匹配的Fill [NSFW]

    ULTRA一样的解锁NSFW概念元素


    The inpainting model needs to be loaded using a specialized inpainting workflow and inpainting sampler, and the image should have a mask for specialized purposes such as image editing or expansion.


    【 请注意 】inPainting内补模型Fill.NSFW是要用专门的inpainting内补工作流例图包含 和内补采样器加载的,图片要有遮罩,【 用来做为修图或者扩图的专门用途 】

    常规作图请使用:RedCraft | 红潮 | Commercial & Advertising Design System - 🌹NewReveal[F.1]ULTRA🌹


    Mainly used for repairing female anatomy and human organs

    The expression of male genitalia is still not ideal

    主要是用来修复女性肢体和人体器官

    [ 对男性生殖器表达还是不太理想 ]


    RED.epicus BIG Movie (FP8) 2/23/2025

    Boring! Repetitive! Spamness!

    RED[创意] Epicus 史诗级 BIG Movie Model

    由于文生图模型的各种加密蒸馏技术普及

    社区新模型作品越来越缺乏创意...

    如果那么喜欢写实照!为什么不去约私拍?!

    乏味,重复,无休止

    Due to the widespread use of various encryption and

    distillation techniques in the T2I model.

    F.1 community works are increasingly lacking in creativity

    If you love realistic photos so much! Why not schedule a realshoot?!

    ---

    所以,夕阳红组织(Sunset Red Squad)定制了这个基于反蒸馏技术的创意类FT Model

    没有锁定的训练集,没有过拟合的风格,一切为了创意而生,

    让扩散模型呈现该有的样子(虽然失败率很高

    ---

    NSFW Unlocked

    ---

    RedCraft DRD(De-Re-Distilled) NewReveal.4M for 20Steps

    CFG 1 | Sampler Deis / DPM++2M / Euler | SGM_uniform


    "May this Valentine's Day fill your hearts with love and joy. Wishing everyone a day surrounded by affection and cherished moments." 🌹🌹🌹🌹🌹🌹

    Happy Valentine's Day ! 情 人 节 快 樂 2/14/2025


    新 年 快 樂 Happy New Year of the Snake

    Get ready for the all-new F.1 Schnell FT model — RUSHReveal·Schnell 「绝情 · 抽卡机」

    BEST Refiner for IL / PONY / XL / MJ / SD15

    RedCraft | 红潮 CADS | UPdated-Feb08 | Commercial & Advertising Design System - RASCH.3 (RUSH·Reveal)🔥 | Flux Checkpoint | Civitai


    New Reveal ULTRA 2/08/2025

    反蒸馏超高材质(可训练)

    CFG Restore to 1 Ensure the same speed as F.1 DEV

    LoRA Fully Compatible De-Distill UH Material (trainable)

    [ POWERED BY FLUX Aesthetics Enhancement LoRA ]

    NSFW Unlocked

    ---

    RedCraft DRD(De-Re-Distilled) NewReveal.4M for 20Steps

    CFG 1 | Sampler Deis / DPM++2M / Euler | SGM_uniform


    Dr Wikeeyang`s latest research

    Flux1-Dedistilled 3.0

    F.1 Distilled2PRO leaked 🏴‍☠️

    https://civarchive.com/models/941929/flux1-dedistilledmixtuned

    NewREVE[A]L 偷跑发布 1/22

    F.1 DEV LoRA 生态 全兼容

    反蒸馏超高材质(可训练)

    CFG Restore to 1 Ensure the same speed as F.1 DEV

    LoRA Fully Compatible De-Distill UH Material (trainable)

    [ POWERED BY FLUX Aesthetics Enhancement LoRA ]

    ---

    本模型基于 F.1 Distilled2PRO 反蒸馏 '破解版' 偷跑(现已公开):

    Flux1-DedistilledMixTuned V3 escape version (Published)

    Flux1-DedistilledMixTuned - v3.0 fp8 | Flux Checkpoint | Civitai

    ---

    追加FLUX DEV版本4k超高清底模权重 EOR v3:

    Flux.1 Dev Edge of Reality 真实边缘 - v3 | Flux Checkpoint

    追加 RED.2 [ ArtAUG ] BF16 美学基础模型权重:

    RedCraft | 红潮 CADS | RED.2 BF16 (ArtAug) | Flux Checkpoint

    ---

    NSFW Unlocked

    ---

    RedCraft DRD(De-Re-Distilled) NewReveal.4M for 20Steps

    CFG 1 | Sampler Deis / DPM++2M / Euler | SGM_uniform

    ---

    SOTA adaptive ability for All-F.1-LoRA !


    Special thanks

    SHM_AI works of excellence :

    SHM Realistic - v4.0 | Stable Diffusion Checkpoint | Civitai

    HudujnikBezKisty works of excellence :

    The Super Realistic - TSR 2.0 | Stable Diffusion Checkpoint | Civitai

    Astraali works of excellence :

    AstrAnime - AstrAnime_V6 | Stable Diffusion Checkpoint | Civitai

    And everyone who has quietly contributed to SD1.5


    SD15RelustionHD 发布 1/18

    Relustion1.5HD 高清发布 1/18/2025

    基于 RETROSD 素材与 HD4K 重铸 SD1.5

    RETRO!重返SD时代的光辉岁月!

    Returning to the glorious years of SD !

    CFG 5-7 DPM++2M /EulerA | SGM Uniform

    采样步数 25~30 steps 为最佳 内置VAE

    设计分辨率为 Hi-RES 0.9M ( 92W像素 )

    高分辨率直出请启用UNET缩放分块脚本

    [ 基于全精度FP32制作,首发FP16版本 ]


    RETRORelustion2 光辉发布 1/16

    基于 RETROSD / FLUX Reveal / EDGE4k 打造东亚复古真实系 Illustrious FT

    RETRO!重返SD时代的光辉岁月!Returning to the glorious years of SD !

    CFG 5 Deis | EulerA | SGM Uniform

    采样步数 25-30 为最佳,可以搭配 Hyper/DMD2/TDD 等加速器

    设计分辨率为 Hi-RES 2M ( 200W像素 ),高分辨率启用UNET缩放

    Thanks so much to everyone for all your support!


    PONYRelustion2 欢乐发布 1/11

    基于FLUX Reveal / EDGE4k 系列打造的 32-bit 东亚真实系 PONY 模型

    Full Model fp32 (12.92 GB) 全精度无蒸馏无污染,BNB显存占用7GB

    First release of FP32 version, unlimited creativity in PONY World!

    助力欢乐马世界的无限创意!

    CFG 5 Deis | EulerA | SGM Uniform

    采样步数25左右为最佳,可以搭配 Hyper/DMD2/TDD 等加速器

    设计分辨率为 Hi-RES 2M ( 200W像素 ),长宽比延续超强适应

    Aspect ratio like PONY's ULTRA adaptive ability

    目标是做出PONY世界的成熟高画质创意底模!

    感谢大家一直以来的支持!


    Special thanks

    Thank you to Freepik and Ostris

    for their outstanding contributions to parameter optimization!


    RED.2 15.2 GB(BF16) 1/8/2025

    POWERED BY FLUX Aesthetics Enhancement LoRA

    Computational Intelligence Lab at ECNU

    RED.2 美学评估基于 DiffSynth-Studio 生成理解交互训练项目 ArtAug

    Paper: https://arxiv.org/abs/2412.12888

    The training process of ArtAug consists of the following steps:

    1. Synthesis-Understanding Interaction: After generating an image using the image generation model, we employ a multimodal large language model (Qwen2-VL-72B) to analyze the image content and provide suggestions for modifications, which then lead to the regeneration of a higher quality image.

    2. Data Generation and Filtering: Interactive generation involves long inference times and sometimes produce poor image content. Therefore, we generate a large batch of image pairs offline, filter them, and use them for subsequent training.

    3. Differential Training: We apply differential training techniques to train a LoRA model, enabling it to learn the differences between images before and after enhancement, rather than directly training on the dataset of enhanced images.

    4. Iterative Enhancement: The trained LoRA model is fused into the base model, and the entire process is repeated multiple times with the fused model until the interaction algorithm no longer provides significant enhancements. The LoRA models produced in each iteration are combined to produce this final model.

    This model integrates the aesthetic understanding of Qwen2-VL-72B into FLUX.1[dev], leading to an improvement in the quality of generated images.

    Usage

    CFG 1 | Sampler Deis / DPM++2M / Euler | SGM_uniform

    Generation without accelerator in 25steps

    Recommended combination accelerator for RED.2 : RED-AIGC / TDD

    Target-Driven Distillation: Consistency Distillation with Target Timestep Selection and Decoupled Guidance

    Generation with TDD-distilled RED.2 in only 4-8 steps

    Distillation accelerator weight scaling to 0.12~0.13

    [ Using Re-sampling with Sampler LCM/EulerA yields better results ]


    The models in this link are in a parallel relationship, not version upgrades

    本链接内的模型是并列的平行关系,并非全部是版本推进

    The differents can be found in the 'About this version' section on the right

    不同版本的说明在右侧的 ‘About this version’ 清单内

    以下是模型列表 list of models


    TURBO Reveal2 圣诞首发!Merry Christmas!

    ( HOTFix v2.1 uploaded - Enhance Lora adaptation )

    是在Reveal NSFW的基础上结合更多人像做出的尝试!

    Reveal2 Turbo 8-10steps

    Combine more portraits based on Reveal NSFW

    RedCraft | 红潮 CADS - Reveal2 TURBO

    祝大家节日快乐!Wishing everyone a happy holiday!


    本模型没有采用任何反蒸馏权重

    Does not use any De-distilled weights


    PONY Relustion 冬至首发 Winter Solstice Festival

    基于高画质写实风格的创意优先设计

    RedCraft | 红潮 CADS - PONY Relustion

    Creative priority design based on high-definition realistic style


    距离上一个PONY模型发布已经过去七个月了:

    MIST XL Hyper Character Style Model 角色风格模型加速版

    感兴趣的朋友可以回顾一下全网首个 Hyper-PONY 超高速模型


    Reveal NSFW

    是FLUX.1 DEV规格的FP8 FT模型,主打男女爱情动作与人体艺术:

    RedCraft | 红潮 CADS - Reveal NSFW

    FLUX. 1 DEV FP8 FT model, featuring romantic actions and body art

    Reveal3 ULTRA

    ( HOTFix v3.2 - PENIS uploaded )

    是FLUX.1 DEV与反蒸馏技术结合实现的高画质版本迭代:

    RedCraft | 红潮 CADS - Reveal3 uncensored

    Reveal‘s high-definition update by De-Re-Distillation Quality Optimality

    Relustion IL NSFW

    是基于SDXL规格的全量训练模型 Illustrious XL 的写实化FT版本:

    RedCraft | 红潮 CADS - Relustion IL NSFW

    Realistic FT of Illustrious XL,full optimization based on SDXL

    Relustion ULTRA

    是在 Relustion IL 的基础上,进一步加强写实化的高清版本:

    RedCraft | 红潮 CADS - Relustion ULTRA

    a high-definition version that enhances realism on Reliability IL

    Relustion XL

    是在 SDXL CADS3 的基础上,结合NSFW训练集制作的高清量化版本:

    RedCraft | 红潮 CADS - Relustion XL

    It is a high-definition quantized version based CADS3 and combined

    with the NSFW training set,Used for HD refine of FLUX and IL models

    RASCH.1 / 2

    是两个不同 Schnell 反蒸馏FT模型上,结合RED.1风格的高速模型:

    RedCraft | 红潮 CADS - RASCH.2

    RedCraft | 红潮 CADS - RASCH.1 Forge

    ReFLEX NSFW

    Schnell NF4版本的二次元绘画模型,主打结构稳定与提示词还原准确:

    RedCraft | 红潮 CADS - REFLEX NSFW

    high-speed model mixed RED.1 on different Schnell De-distillation FT models


    在保证提示词准确的基础上,De-Re-Distilled(DRD) Schnell模型兼顾了速度与质量

    而且蒸馏模型具有天然的肢体稳定性与画风稳定性,即便是4bit量化版本效果也很好

    6~10GB显存占用,4-8步出图,速度飞快(特别适合于画风训练与建筑装修模型


    the De-Re-Distilled (DRD) Schnell model balances speed and quality

    Moreover, the distillation model has natural stability , even the 4-bit quantified

    6-10GB of video memory usage, 4-8 steps of image output, fast speed

    (Especially suitable for artistic style creation and architectural decoration)


    以下是RedCraft系列,基础美学模型Red.1的简介

    RedCraft RED.1

    BF16 CADS Commercial & Advertising Design System

    可能是目前快速出图(10步以内)版 BF16 模型中,出图质量较好、细节较丰富的基础模型。

    Fine Quality Steps 10-20 Model, In some details, it surpasses the Flux series models and approaches the 20B Parameters models.

    Based on METAFILM AI - Commercial & Advertising Design System, Merge of flux-dev-de-distill, finetuned by ComfyUI, Block_Patcher_ComfyUI, ComfyUI_essentials and other tools. Recommended 10-20 steps. Greatly improved quality compared to other 12B models.

    基于

    De-Distill & CADS商业素材 FP16

    支持在线生成 ComfyUI WebUI

    10-20 STEPS Euler / DPM++2M | beta / SGM_Uniform

    CFG 3-3.5

    Real CFG needs to be set (ignore guidance, or set to 0)

    ——————————————————————————

    The versions with AIO (All in one) in the name include UNET + VAE + CLIP L + T5XXL (fp8). Also known as Checkpoint or Compact version.

    Using BNB NF4 & GGUF quants in ComfyUI requires installing custom nodes that add special model loaders:

    For using UNET versions, you also need to have the TEXT ENCODERS and VAE.

    If you don't have them, download them from here:

    Place the model in "models/diffusion_models" or "models/unet", both text encoders in "models/clip" and vae in "models/vae" folder.

    In ComfyUI, use the standard flux workflow or add 'Load Diffusion Model', 'DualClipLoader' and 'Load VAE' nodes to replace the checkpoint loader and complete the setup.

    In Forge, set the option "Diffusion in low bits" to "bnb-nf4"

    Thanks to city96 for gguf quantization script. 
    Thanks to reddit user a_beautiful_rhind for bnb quantization script.

    ——————————————————————————

    同样推荐选择大杨老师的 8bit finetune 版本:

    Flux1-DedistilledMixTuned-V1 - v1.0 fp8 | Flux Checkpoint | Civitai

    可能是目前能找到最符合官方基准风格并且出图效果最好的加速模型

    Recommend:

    UNET versions (Model only) need Text Encoders and VAE, I recommend use below CLIP and Text Encoder model, will get better prompt guidance:

    Simple workflow: a very simple workflow as below, needn't any other comfy custom nodes(For GGUF version, please use UNET Loader(GGUF) node of city96's):


    Thanks for:

    https://huggingface.co/wikeeyang, Wikee Yang for carefully finetune the 8-bit model and providing the model informations,You can find it here:

    wikeeyang/Flux.1-Dedistilled-Mix-Tuned-fp8 · Hugging Face

    https://huggingface.co/Anibaaal, Flux-Fusion is a very good mix and tuned model.

    https://huggingface.co/nyanko7, Flux-dev-de-distill is a great experimental project! thanks for the inference.py scripts.

    https://huggingface.co/MonsterMMORPG, Furkan share a lot of Flux.1 model testing and tuning courses, some special test for the de-distill model.

    https://github.com/cubiq/Block_Patcher_ComfyUI, cubiq's Flux blocks patcher sampler let me do a lot of test to know how the Flux.1 block parameter value change the image gerentrating. His ComfyUI_essentials have a FluxBlocksBuster node, let me can adjust the blocks value easy. that is a great work!

    https://huggingface.co/twodgirl, Share the model quantization script and the test dataset.

    https://huggingface.co/John6666, Share the model convert script and the model collections.

    https://github.com/city96/ComfyUI-GGUF, Native support GGUF Quantization Model.

    https://github.com/leejet/stable-diffusion.cpp, Provider pure C/C++ GGUF model convert scripts.


    境内已发布在 Modelscope 魔搭社区,感受极速下载!

    RedCraft | 红潮 CADS Commercial & Advertising Design System · 模型库

    全网唯一支持反蒸馏FLUX模型在线生成的平台(社群免费):

    AIGC 专区 - 图片生成 · 魔搭社区 Model: qijitech/RedCraft-12b-10steps-FP16-AIGC

    The only platform (community-free) that supports the online generation of De-distillation FLUX models.

    Also on Huggingface.co soon

    ————————————————————————————————————————

    测试问题请留言,业务合作看个人首页 +V Zyuan980

    做好工具人 服务艺术家 更多资料:https://x1f3ewlrcf.feishu.cn/wiki/BjJ1waQaLitPB4k7Lbvc0MaVnzb?fromScene=spaceOverview&open_tab_from=wiki_home

    ————————————————————————————————————————

    Description

    null

    FAQ

    Comments (55)

    J1BFeb 27, 2026· 2 reactions
    CivitAI

    The DPO version seems very good from my few tests.. What CFG setting do you recommend with it? I found raising it slightly to 1.5 seems to give slightly better/less turbo looking images.

    AiMetatron
    Author
    Feb 27, 2026· 1 reaction

    Thanks for the feedback! I've also noticed that CFG 1.5 can indeed help reduce that "overprocessed" look on smaller canvases.

    Personally, I usually stick to CFG 1.0 for general generation because it offers the best speed. However, pushing it to 1.5 on smaller samples can provide more breathing room for composition and help the negative prompts take effect. The trade-off, though, is that it sometimes leads to overly saturated colors.

    I'll release a LoRA version of Veris DPO shortly (or export it yourself). Using it as a LoRA will allow you to run higher CFG values and step counts with lower weights, which should unlock even better detail and quality.

    BIG_AFeb 27, 2026· 4 reactions
    CivitAI

    妈呀!新模型太好用了!比老子鸡巴还好用!操!

    顺便说一句,在forge NEO上运行良好!

    AiMetatron
    Author
    Feb 27, 2026· 1 reaction

    别别。。还是你比较长

    EuphloryFeb 27, 2026· 2 reactions
    CivitAI

    So, where do i have to put this model?

    AiMetatron
    Author
    Feb 27, 2026· 4 reactions

    ComfyUI / models / diffusion_models for "UNET" loader

    738586719206Feb 27, 2026· 2 reactions
    CivitAI

    哥们儿有SD能用的吗?我没有comfyUI

    AiMetatron
    Author
    Feb 27, 2026

    这个版本 Forge(NEO) 能用,刚才大老A不是测试了么,不过模型和TE/VAE得分开加载,我没传AIO版本,C站以后就只传bf16 单文件了,太多太多人找不到下面的文件清单,找到了也对应不上所需版本。

    HF 抱脸仓库:https://huggingface.co/GuangyuanSD/Z-Image-Distilled

    AIO版本比较占内存,等明天我重启一下再导出。

    foxgotFeb 27, 2026· 6 reactions
    CivitAI

    You're very cool and your models are great!!! Are you planning to create an anime model? Or add more data for anime drawings. Thank you for your hard work!

    AiMetatron
    Author
    Feb 27, 2026· 6 reactions

    Regarding your question, I've actually been keeping an eye on the Anima model recently. I think it's a fantastic choice—a very cost-effective DiT mini model with 2 billion parameters and a text encoder (TE) of only 0.6b.

    Here is my take on the current landscape:

    Z-Image's Focus: While Z-Image does have some capability to generate anime-style images, its categorization and tagging aren't strictly aligned with the Danbooru tags library. It wasn't the primary focus.

    Specialization: When it comes to specific artists, styles, and deep anime generation, a dedicated model like Anima is definitely more professional and specialized.

    So for now, I think pairing the right tool with the right job is the best approach!

    DevilSShadoWFeb 27, 2026· 3 reactions

    @AiMetatron thank you for shouting out Anima, it's a great model and a great foundation to build on for those interested in anime generation using a combination of booru tags and natural language. It works surprisingly well out of the box and is very fun to play around with.

    I agree, let Zimage focus on what it does best, which is realism.

    AiMetatron
    Author
    Mar 4, 2026· 2 reactions

    @DevilSShadoW I completely agree. That's precisely why I've mostly abandoned the anime style in my new ZIT version. The accuracy and performance of 'Z' for natural language expression in that domain weren't ideal.

    While the ZIB line will continue to maintain its capabilities in illustration and ink-wash expression, approaches like Anima—which focus on abstraction and brushwork—are the ideal direction. This is especially true considering its efficiency with smaller parameters and how the parameter limits prevent it from becoming too 'realistic,' thereby avoiding user base fragmentation.

    Zakman99Feb 27, 2026· 6 reactions
    CivitAI

    ZImage DPO “Veris” It follows prompts perfectly, produces more realistic images, and handles skin much better, but I feel that using Lora spoils the picture, so I prefer to use it without Lora.

    LetTheBassDropFeb 27, 2026· 4 reactions

    Thgat's because so many LoRAs are being trained on SDXL/Illustrious images so you end up getting the same stupid face and plastic skin in new models.

    Zakman99Feb 27, 2026· 1 reaction

    @LetTheBassDrop I use Lora on the basic ZIT model and I have no problems with it, but that's not the point. The model basically conveys what I want perfectly well anyway.

    AiMetatron
    Author
    Feb 27, 2026· 3 reactions

    @Zakman99 That is a very astute, expert-level observation.

    You are absolutely right that the ZI Base (ZIB) architecture demonstrates far better LoRA (restoration/fidelity) capabilities compared to the ZITurbo line. However, this is precisely why ZITurbo is less affected by LoRAs—it resists the "injection" of new data.

    The distilled models (like Veris) are the opposite. They are highly sensitive to LoRAs trained on the base model. The more sensitive they are, the more they are influenced by the LoRA's training methodology. The issue is that LoRAs trained on consumer hardware often don't match the precision and resolution quality of the original pre-trained models from developers. As a result, these learned weights can disrupt the quality of the base model.

    AiMetatron
    Author
    Feb 27, 2026· 8 reactions

    The current solution I can think of is to reduce the weight of the previous version of the training model and find a balance between quality and restoration. Another is to look forward to a more efficient and compatible trainer solution with Zimage (I heard that the Z project team and developers are tackling this). Finally, we look forward to Tongyi& Z project returning to the construction of the open source community.

    Let's silently recite "Sesame! Sesame! Open the Door!"

    Zakman99Feb 27, 2026· 1 reaction

    @AiMetatron Thank you for your model and for allowing many of us to experience it. It is truly unique among the entire range of ZitTurbo user models.

    6vidit9Feb 28, 2026· 7 reactions
    CivitAI

    The skin detail is exceptional, and the scene composition is executed with remarkable precision. Overall, this release demonstrates in quality across all aspects🤘

    AiMetatron
    Author
    Mar 2, 2026· 3 reactions

    Yes,The more you play, the more ZIB's finetuning edge shines through. Its openness unlocks a world of possibilities.

    furricely8088334Feb 28, 2026· 2 reactions
    CivitAI

    ZImage DPO “Veris” fp8 please

    AiMetatron
    Author
    Feb 28, 2026· 2 reactions

    I have uploaded more quantification and LoRA editions to HF to avoid causing confusion for users in the community,CivitAI only retains BF16 edition:
    https://huggingface.co/GuangyuanSD/Z-Image-Distilled/tree/main

    furricely8088334Mar 1, 2026· 1 reaction

    @AiMetatron thanks

    jexshotFeb 28, 2026· 3 reactions
    CivitAI

    Veris Images are too saturated with color. Please improve Color accuracy.

    AiMetatron
    Author
    Mar 2, 2026· 2 reactions

    There is a Lora exported here, where you can adjust the intensity and number of steps appropriately to find your own balance point.

    https://huggingface.co/GuangyuanSD/Z-Image-Distilled

    6vidit9Mar 1, 2026· 4 reactions
    CivitAI

    Genuinely the best ZIB checkpoint here on CIVITAI

    AiMetatron
    Author
    Mar 2, 2026· 3 reactions

    Thank you too much, let's continue to think of ways to achieve the precision of Turbo and a more natural color tone transition in ZIB's distillation method.

    Nuwanda_Mar 2, 2026· 2 reactions
    CivitAI

    Is it possible to get the ZImage Base version as a lora ? As it loses variation compared to the original BF16 version.

    kennedysworksMar 2, 2026· 3 reactions
    CivitAI

    I went to hf and found that there was a version of fp32. Is it better than bf16 in detail? I'm inquiring because I've been very satisfied with the details of the image when I used zit as fp32

    t2kardyyMar 2, 2026· 3 reactions
    CivitAI

    I feel like my results with Veris are a bit worse than with Lucis, when combining them with my self trained LoRAs. I did use 15 steps on Lucis vs. 8 steps on Veris tho.
    I do train my LoRAs on Z-Image Base. Is that wrong? Will training on Veris directly yield me better results? I thought that it would destroy the acceleration, so I have not tried yet.

    AiMetatron
    Author
    Mar 2, 2026· 3 reactions

    I still insist on training on the main branch of ZIB, even though there are many excellent fintunes, for the sake of weight universality, I still recommend using the base model for training and then blending it with various ft models.

    t2kardyyMar 3, 2026· 2 reactions

    @AiMetatron First of all, thank you for the reply. Greatly appreciated. Second: What do you mean with "blending it with various finetune models"? Do you mean training with base and using finetunes for inference? Because this is what I do. Anyways, I might have found the problem to Lucis not responding too well, to my already trained LoRAs. I think it is actually quite the opposite. I think it responds too well. I checked my dataset again, and a lot of my images were a bit grainy. Really high quality, but still shot with high iso or added film-grain. Other finetunes did not really capture on that, but Veris does, making the outputs result in having sharp edges and being a bit noisy. Like also having high quality but just feeling a bit unreal due to visible, weird looking grain. I think I am going to re-train with a cleaner dataset and try if that fixes it.

    wangf35979Mar 3, 2026· 3 reactions
    CivitAI

    Veris非常好用,但对提示词的敏感度有些过于高了,一般只要出现breast相关的提示词就偏向NSFW图像了,不知道会不会再调整。
    另外个人仍旧钟爱REDZ1.5,希望能继续更新,它出来的图片细腻柔和,而后续几个版本模型相对来说调子太重了。

    3527286645845Mar 8, 2026· 1 reaction

    这个模型就是偏向nsfw的把 zi官方开源对于私处的数据集太少了

    MOD215Mar 3, 2026· 1 reaction
    CivitAI

    The absolute best by far! Thank you for sharing it.

    ciggMar 3, 2026· 4 reactions
    CivitAI

    Excellent work. The only problem is that how can I generate smaller breasts?

    AiMetatron
    Author
    Mar 8, 2026· 1 reaction
    790374882120Mar 4, 2026· 3 reactions
    CivitAI

    请问大佬为什么提示词里只要出现和胸部相关的提示词就会露点啊

    AiMetatron
    Author
    Mar 4, 2026· 1 reaction

    因为特殊部位的权重比较大,平衡点拿捏的还不够好。

    2360045981119Mar 5, 2026· 2 reactions
    CivitAI

    大佬,你是不是删模型了?以前我记得有一款黑兽的,宏大和微观世界,还有什么纹理材质啥的都能画出来的,现在找不到了

    AiMetatron
    Author
    Mar 5, 2026· 1 reaction
    kennedysworksMar 5, 2026· 2 reactions
    CivitAI

    In my case, I find a case of a beard on a woman's face in veris. I worked with a resolution of 2048_1152

    A bright, photorealistic lifestyle portrait of a 16 years old Japanese girl outdoors by a calm waterfront, about to eat ice cream from a spoon. She is framed in a tight medium close-up within a 16:9 widescreen composition. She holds a small red-handled spoon toward the camera with a neat dollop of white ice cream, hovering just in front of her lips. Her gaze is steady and direct, calm and slightly inquisitive.

    A gentle breeze lifts and sweeps strands of her dark hair across her face, adding a candid, spontaneous feeling. She wears a crisp white collared shirt with dark backpack straps visible on both shoulders, suggesting a casual school-day moment.

    Lighting: soft natural daylight with a slight back/side rim light that outlines her hair and cheekbones. Bright, airy exposure with smooth highlight roll-off; a subtle sun haze and faint flare glow on the right side of the frame.

    Background: shimmering water reflections and distant greenery/structures rendered as creamy bokeh. Strong subject separation with shallow depth of field, keeping the environment recognizable but soft and unobtrusive.

    Camera & optics: eye-level to slightly above-eye perspective, intimate portrait distance, 35–50mm equivalent look, wide aperture (f/1.8–2.8). Focus crisply on her eyes, with the spoon/hand slightly softer in the foreground to emphasize depth.

    Composition: her eyes centered as the visual anchor; the spoon in the lower foreground leads the viewer’s attention toward her mouth. Clean, minimal framing with bright open water area on the right for a light, cinematic balance. Ultra-realistic skin texture, natural color, gentle filmic grading, subtle grain, 8K detail.

    kennedysworksMar 6, 2026· 1 reaction

    @AiMetatron It's a problem with fun distil lora. I applied distil lora to the zibfp32 version and did the same prompt, and I got a beard again.

    AiMetatron
    Author
    Mar 8, 2026· 3 reactions

    This is a really scary problem, I will now test if I can reproduce it locally.

    zartoMar 9, 2026· 3 reactions

    Maybe the "dark hair across her face" can be misleading for the model

    kennedysworksMar 9, 2026· 1 reaction

    @zarto Thank you for your interest. I do the fp32 version for precision, and even if I modify the prompt in that part, I get the same beard on a woman's face,

    zartoMar 9, 2026· 1 reaction

    @kennedysworks Do you get a beard with only the first paragraph (ending at "inquisitive") ? If not, you can then try adding paragraphs one by one for each generation to find the guilty one.

    kennedysworksMar 10, 2026

    @zarto 

    https://huggingface.co/alibaba-pai/Z-Image-Fun-Lora-Distill/discussions/9


    Thank you so much for your interest. I feel the warmth of the community. I posted that on alibaba-pai and got a response. Something to do soon.

    carlosurportable419Mar 6, 2026· 1 reaction
    CivitAI

    so we are supposed to get good results?

    because i tried a bit and i don't,

    first i used z_image_turbo-Q8_0.gguf before (my gpus has small vram) so i downloaded the redcraftFeb2626LatestZib_zibDistilledDX3Lucis

    on a classic workflow it works i use 1536x2048 as usual , first thing i notice it's unable to do grass ?? i tested 1 of my lora, the skin was dark, and the character has a n o r e x i a , some part of the hair is transparent, got 2x 6 fingers, even the bannas trees was less good than in the Q8.gguf

    but where i got into problem is when i try to use freefuse ( to load multiple lora's)

    i get out of memory , redcraftFeb2626LatestZib_zibDistilledDX3Lucis.safetensors isn't that supposed to be fp8 ? like the Q8.0.gguf , file size is similar but looks like it use far more memory ?

    AiMetatron
    Author
    Mar 8, 2026

    is not a GGUF quantization specification (unless mainstream inference engines have built-in GGUF support, I cannot find a standardized GGUF format standard)

    DanzelususMar 6, 2026· 5 reactions
    CivitAI

    Every image overexposed, oversaturated, pixelated edges. Very bad model

    CupRunethOverWithLSDMar 8, 2026· 2 reactions

    Then you are doing something wrong. 😂

    AiMetatron
    Author
    Mar 8, 2026· 1 reaction

    Perhaps there are incompatible environments, do you use Forge (NEO)? It's still ComfyUI. If it's the latter, you can use the workflow shown in the example diagram, which only requires 8+2 inference steps.