REDQW21-UNLOCKED SEP EVENT OVER STAY TUNED

09/30 REDQW21(UNLOCKED) v2 发布 TI2I 多参图生图版本（沿用T2I工作流即可）

天冷了注意保暖！穿衣加强 模型仅用于学习和研究 Built with Qwen
The model is FREE. The Key is NOT. CAN YOU FIND IT ?

免費下載 KEY-REDQW21 Fine-tune，自行研究隱藏的採樣方案，成功返圖並分享你的研究即可獲得 888 BUZZ；如果不想自己「解謎」，也可以購買完整 ComfyUI 對齊工作流。

KEY-REDQW21 Fine-tuneを無料でダウンロードし、隠されたサンプリング設定を自分で研究してください。成果画像と研究内容を共有すれば 888 BUZZ を獲得できます。自分で「謎解き」をしたくない場合は、完全な ComfyUIアライメントワークフローを購入することもできます。

Download the KEY-REDQW21 Fine-tune for free, research the hidden sampling setup yourself, and earn 888 BUZZ by sharing your results and research; if you don't want to solve the puzzle yourself, you can also purchase the complete ComfyUI Alignment Workflow

THE ‘SECRET’ KEY

Qwen-Image 2.1 · Alignment Challenge · 888/Post

模型非商业用户免费下载。配套的采样设置并不公开
我们发布一个经过特殊训练与推理设计的 Qwen-Image 2.1 Fine-tune，但不会公开完整的「Alignment Key」——包括特殊采样器、采样步数、Sigma 曲线、APG / FreSca 参数、Guidance、权重偏移以及后期处理流程等。

你可以免费下载模型，并使用 ComfyUI、Diffusers 或任何其他推理环境自行研究、实验和尝试复现我们的参考效果。>付费工作流<，定价 88800 Buzz，活动开始后会有折扣。它是一套已经调好的 ComfyUI 流程，用来稳定出图。页面上的例图大多是随机种子生成的，几乎没有 Cherry pinking。

工作流里不含 Qwen 模型，也不含这个微调。

This fine-tune is based on Qwen Image 2.1. The training steps, the sampler, and the weight shift at inference were all set separately. The sampling method, step count, curve, and finish are not listed. ComfyUI or any other setup is fine. From testing so far, matching the result takes some trial and error.

There is also a paid workflow at 88,800 Buzz, with a discount after the event starts. It is a tuned ComfyUI setup for more stable output. Most of the sample images were generated directly from random seeds. The workflow does not include the Qwen model or this fine-tune.

以上内容只用于学习和研究。工作流的收入是我个人的劳动所得。这个项目没有接受 Qwen Image 开发团队的资金或技术支持，沉默与傲慢是大厂的统一回复。Qwen Image 2.1 目前是研究许可，不能用于商用。

这个微调基于 Qwen Image 2.1。训练步数、采样器，以及推理时的权重偏移，都做过单独处理。采样方式、步数、sigmas曲线和后期没有写出来。用 ComfyUI 或其他环境都可以试。以目前的经验，要把结果对齐，需要自己多试几次。

这是一场关于 Qwen-Image 2.1 推理与采样 的公开挑战

我发布的是一个经过特殊训练、特殊步数、特殊采样器，以及特殊 APG / FreSca 权重偏移处理的 Fine-tune。完整的对齐钥匙不公开。采样方案、采样步数、Sigma 曲线、APG / FreSca 参数、Guidance、权重偏移、后期方案，都在那把钥匙里。缺一环，结果就会偏。

你可以免费下载模型，并使用 ComfyUI、Diffusers 或任何其他推理环境自行研究、实验和尝试复现我们的参考效果。

这并不简单。

但这正是游戏的一部分。

如果你通过自己的研究找到了优秀的参数组合，欢迎把作品和你的实验结果发布到模型页面。我们会通过 Buzz 回馈那些真正投入时间进行研究、实验并取得有价值成果的挑战者。购买商业 Workflow 的用户同样可以参与并获得回馈。（回馈BUZZ以Posts帖子为单位）

This project and its challenge are intended solely for learning, experimentation and research purposes.

The commercial Workflow is my own work and its BUZZ revenue represents compensation for my independent workflow development and research. I have received no financial or technical support from the Qwen-Image development team for this project. My work on this project has been conducted independently, without official technical assistance, funding, or endorsement from the Qwen-Image team.

Qwen-Image and this project are not affiliated with, sponsored by, or endorsed by the Qwen-Image development team unless explicitly stated otherwise.

付费对齐工作流

不想猜的人，走短的那条路，付费购买工作流。活动启动后会有促销折扣。
Recommended to purchase BUZZ

House rules

1. Download the fine-tune and do the research yourself.

2. Post the result on this model page and write down what you tried.

3. A qualifying post pays 888 BUZZ. Workflow buyers included.

4. Publishing what you found is part of the research.

5. If you want the stable state without the hunt, buy the workflow.

技术解释 Technical explanation

APG / FreSca 这两种做法都不改模型文件，只在采样时处理「引导」这股推力。平时说的 CFG，就是沿着提示词的方向推：推得越紧，画面越贴提示词，也越容易过。

APG 管的是引导的方向。它同样往提示词上推，但会先把这股推力拆开，把容易推过头的部分压下去。说得具体一点，就是在推理时对引导向量做投影，再按情况抑制。

FreSca（Frequency Scaling Guidance）管的是引导信号里的低频和高频。正向条件和负向条件相减，得到引导，再按频率拆开处理，然后加回去。低频大致对应构图、大块色调、明暗和物体的整体形状；高频大致对应纹理、边缘和细小结构。两边可以分开调节。例如构图保持原来的力度，只把细节加强一点；或者少推细节，把构图的引导留下来。

APG / FreSca ,Neither of these changes the model file. Both only adjust the guidance push during sampling. CFG is the usual version of that push: it moves the result toward the prompt. A stronger push follows the prompt more closely, and it also overshoots more easily.

APG controls the direction of that guidance. It still pushes toward the prompt, but it splits the push first and holds back the part that tends to go too far. More precisely, it projects the guidance vector during inference and then suppresses it where needed.

FreSca (Frequency Scaling Guidance) controls the low and high frequencies inside the guidance signal. It takes the difference between the positive and negative conditions, splits that guidance by frequency, and adds it back afterward. Low frequency mostly corresponds to composition, broad color, light and dark, and the overall shape of objects. High frequency mostly corresponds to texture, edges, and small detail. The two can be adjusted separately. Composition can stay as it is while detail is guided a little harder, or detail can be pushed less while the guidance for composition is kept.

大量例图来自 @haotian236 和 @Hatolic 的努力更新和持续开放，也感谢社区里认真做作品的各位。

只用于学习和研究

这场游戏只用于学习和研究。

Fine-tune: Built with Qwen, Applicable to Qwen-Image-2.1
Workflow: Built with ComfyUI

Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.

https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE

Learning and research only

This game is for learning and research.

Fine-tune: Built with Qwen, Applicable to Qwen-Image-2.1
Workflow: Built with ComfyUI

Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.

https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE

REDKREA2-Alternating Evaluation
REDGPT2-Turbo-逼真剧创

The Grand Pussy Truth 普世真理

KREA2 GPT2 逼真剧创 Alternating Evaluation | MIST3 9/2/2026

采用 SDA（语义方向对齐）训练——基于教师指导的多样性对齐损失，结合 Forward XM 最佳-5候选者探索方法，在单节点高噪声（σ = 0.9567）上完成训练。本方案已取得 @fok3827 商业使用授权，采用双模型高低区 sigmas（特指降噪参数序列）交叉去噪，4H+6L 交叉。

本模型隶属于 AiMetatron Paid Access 模型服务包，所有购买了 红潮/黑兽 H3 服务的用户可以找我获取授权下载地址。单独在本站购买 REDKREA2 仅获得本页面模型的Access。

Recommended to purchase BUZZ

免费商业使用的条件

Krea 2 是 Krea.ai 公司发布的 AI 图像生成模型，提供两个可下载变体：Krea 2 Raw 和 Krea 2 Turbo。其商业使用受 Krea 2 Community License Agreement（社区许可协议）​ 约束。在社区许可协议下，免费商业使用 的前提是：你（包括所有共同所有或控制的关联实体）的全公司年总收入低于 100 万美元（$1,000,000 USD）​，按过去 12 个月滚动计算。

输出的所有权

• 你拥有你生成的输出（Outputs）​，前提是你遵守本协议。

• Krea 对输出不主张所有权。

• 你需对输出的内容、准确性、合法性以及使用或分发的一切后果承担全部责任。

KREA 2 社区许可协议全文阅读

Trained with SDA (Semantic Directional Alignment) — a teacher-guided diversity alignment loss — wrapped in Forward XM best-of-5 candidate exploration, on a single high-noise sigma node (σ = 0.9567). This method is commercially licensed by @fok3827 . It employs two models with dedicated high-noise and low-noise sigma schedules (i.e., denoising parameter sequences), alternating between them for cross-denoising in a 4H + 6L configuration.

This model belongs to the AiMetatron Paid Access model service package. All users who have purchased the RedCraft/Dark Beast H3 service can contact me to obtain an authorized download address. Purchasing REDKREA2 separately on this page will only provide access to the REDGPT2 Krea2 fine-tune files.

Conditions for Free Commercial Use

Krea 2 is an AI image generation model released by krea.ai, available in two downloadable variants: Krea 2 Raw and Krea 2 Turbo. Its commercial use is governed by the Krea 2 Community License Agreement.Under the Community License Agreement, free commercial use is permitted only if your total company-wide annual revenue (including all affiliated entities under common ownership or control) is less than one million US dollars ($1,000,000 USD), calculated on a trailing twelve-month basis.

Ownership of Outputs

You own the Outputs you generate, provided that you comply with this Agreement.

Krea claims no ownership of Outputs.

You are solely and exclusively responsible for the content, accuracy, legality, and all consequences of the use or distribution of the Outputs.

KREA 2 COMMUNITY LICENSE AGREEMENT

这个版本的 Krea2 Fine-Tune 由高噪权重和低噪权重两个文件组成，并通过 “Alternating Evaluation” 方案进行交叉采样。
This Krea2 Fine-Tune consists of two separate weight files: a high-noise model and a low-noise model. They are used in an “Alternating Evaluation” sampling scheme, where the two models are interleaved throughout the generation process.

The goal of this approach is to incorporate the high-noise-region diversity recovered from Krea2 RAW in @fok3827 ’s SDA (Semantic Directional Alignment) training results. To better preserve the details obtained through diversity-oriented sampling, the two models are not simply used in a sequential handoff. Instead, they are alternated according to the original training concept, helping prevent useful variations introduced in the high-noise region from being lost during subsequent inference steps.

@fok3827 treats Semantic Directional Alignment treats diversity collapse as a direction problem. For one training image x0 we draw two noises (z1, z2) and noisify both to σ = 0.9567 — the highest learnable step of the 8-step Turbo schedule, where composition is decided. The frozen teacher (Krea 2 RAW, the non-distilled parent) and the student (Turbo + LoRA) each predict x0 for both noises; both predictions are decoded and embedded by a frozen CLIP stack. The teacher's feature delta ΔT records which direction in perceptual space this noise swap should move the image; the loss L_div = 1 − cos(ΔS, ΔT) teaches the student's delta ΔS to point the same way instead of collapsing all noises onto one template. An SFT self-anchor keeps the student's own trajectory stable.

It is important to note that restoring diversity generally comes with a certain trade-off in generation stability and may increase the probability of artifacts such as anatomical or limb-related errors. Therefore, this approach is not intended to maximize randomness indiscriminately. Instead, dual-model alternating sampling is designed to produce richer and more personalized expressions of cinematic and drama-oriented visual elements, reduce the repetitive “one-style-fits-all” look often seen in generative outputs, and recover more of the stochastic sampling behavior characteristic of DiT-based Flow Matching models.

The ultimate goal is to bring back the creative unpredictability of Flow Matching sampling while maintaining a usable level of stability, allowing the model to produce more diverse, less predictable, and more original visual details.
需要注意的是，恢复多样性通常会以一定程度的生成稳定性作为代价，并可能增加肢体结构等问题出现的概率。本方案目的在于通过双模型交替采样来获得影视剧主题元素的更丰富个性化表达方式，减少千人一面的作品风格，还原 Dit flowmatch 模型抽卡的乐趣，表现出更多的原创性。

其设计目的是借鉴 @Fok 在 SDA（Semantic Directional Alignment） 训练中，基于 Krea2 RAW 所恢复的高噪区域多样性。为了更好地保留通过多样性采样获得的细节，本方案并非让两个模型进行简单的阶段接力，而是根据原作者的训练思路，在采样过程中交错使用高噪与低噪权重，从而尽可能避免高噪区域中的推理结果在后续过程中被抹平或丢失。

女性尿道口随机抽卡多样性比较。

Compare the random sampling of female urethral meatus via Krea2-REDGTA2 and Krea2-RedCraft.

Dark Beast | 黑兽 H3 Director Edition 自动短剧生产线 08/28/2026
https://civarchive.com/models/2242173/dark-beast-or-h3-director-edition
是原生H3单次采样直出2k分辨率的高清方案（ 768P/VSR/RIFE TensorRT ）

RedCraft | 红潮 | REDMIX Hybrid A2A beta2 + LTX25 2k 16-bit HDR- 08/25/2026
https://civarchive.com/models/958009/redcraft-or-or-hybrid-h3-a2a-beta2-ltx25-2k
beta2 版本融合了 LTX2.5 distilled 作为2k高清化方案，并且消除了音频bug。

REDGraft | LTX 2.5 | 老同学 Fast 2K 16bit-HD| sulphur2 ported 移植版 08/21/2026
https://civarchive.com/models/1295569/redgraft-ltx-25-fast-2k-or-sulphur2-ported
LTX 16 bit HDR & Nvidia RTX VSR

以上付费用户提供一对一配置服务

One-on-one configuration services are provided to paid users.

Recommended to purchase BUZZ

The "RedCraft" "Black Beast" and "REDGraft LTX" model series

No Mosaics. 无码 影视专用

KREA2 GPT 逼真版 Grand PUSSY Truth | MIST2 7/13/2026

文生图版：RedCraft | 红潮 | KREA 2 赤佬2 Bastard Edition （INT8/INT4）

硬核版本：Dark Beast | 黑兽 🐱‍👤Krea2赤佬无码版 已发布 06/28/2026

MIST XL Character Style Model 角色风格模型 AiARTiST

训练底模使用的是：Pony Diffusion V6 XL

开源模型 UNIT-lib XL（已包含1GIRL权重，叠加使用可加强人像）：

AiARTiST XL 基础单元 CADS2 LoRA 兼容版

UNIT-lib XL 开放会员下载，支持融合，会员同时享有在线加速生成

----------------------------------------------------------------------

内置加速器 Accelerator：Hyper-SDXL  | 快过闪电的设计渲染

Hyper-SD是最新扩散模型加速技术之一，无损高清加速，CN适配良好

经过多个底模的实测，Hyper-SD可推导出比加速前更多的画面信息！

Hyper-SD is one of the latest diffusion model acceleration technologies, with lossless high-definition acceleration and good CN adaptation.

After actual measurements on multiple base models, Hyper-SD can derive more screen information than before acceleration!

----------------------------------------------------------------------

触发词 Triger Word：1GIRL,Any_Girl

生成参数参考：

parameters

This photo shows a chinese woman wearing a warm and textured sweater,holding a steaming cup in her hand. She sat in the dimly lit room,with soft golden lights illuminating her side. The lady is braiding her hair,seemingly fully focused on this moment,perhaps savoring the aroma of the drink. The background has a rural feel,with a plant in the vase and textured wall

Negative prompt: NSFW

Steps: 10, Sampler: Euler a, CFG scale: 2, Seed: 3573086999, Face restoration: CodeFormer, Size: 640x1280, Model hash: db0525a2bc, Model: AiARTiST-MIST-HYPER.fp16, Denoising strength: 0.4, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer version: 23.11.1, Hires upscale: 1.5, Hires upscaler: Latent, kohya_hrfix_enabled: True, kohya_hrfix_block_number: 3, kohya_hrfix_downscale_factor: 2, kohya_hrfix_start_percent: 0, kohya_hrfix_end_percent: 0.35, kohya_hrfix_downscale_after_skip: True, kohya_hrfix_downscale_method: bicubic, kohya_hrfix_upscale_method: bicubic, Mask blur: 4, Inpaint area: Only masked, Masked area padding: 32, Version: f0.0.17v1.8.0rc-latest-276-g29be1da7

后期处理

Postprocess upscale by: 2, Postprocess upscaler: R-ESRGAN 4x+

It is recommended to use the sample picture prompt words directly, and then add modified detail words as needed.

「 建议直接使用样图提示词，然后根据需要增添修改细节词 」

推荐采样步数：10  CFG scale: 1，VAE Automatic 内置VAE，Clip Skip：1

推荐采样方法：Euler A，DDIM，DPM++ 2M Turbo

Hyper sample steps：10，CFG scale: 2，Clip Skip 1

Hyper sampler：Euler A

推荐分辨率：任意分辨率

2 Steps：Text2img long side 1280，Send to img2img 0.4-0.5 resize x1.5

如果图片尺寸用于出版，可发送到「Tiled Diffusion」或「后期处理」扩大

图片后期处理高清化设置： 8x_NMKD-Superscale_150000_G 可叠加

                                           4x-UltraSharp 或 R-ESRGAN 4x+ Anime6B

----------------------------------------------------------------------

More words

As of 2024, there are still many friends who don’t know much about XL. They think that online image generation is slow and that you need to record tags when publishing images.

In fact, these are concepts left over from outdated tutorials. First of all, I tested the online drawing today and found that the drawing speeds of XL and 1.5 are almost the same.

Secondly, XL benefits from a larger number of parameters and a double text layer design. In fact, it can generate most drawing styles based on the base model color.

Even if some concepts are locked through LoRA, the required screen elements can be changed arbitrarily through prompt words. There is also an automatic translation tool in lib online.

SDXL is a form between the commercial model and the 1.5 community model. It does not require too many LoRA combinations or complex prompt words.Negative words, all you need is your creativity and aesthetic taste. Even by redrawing from graph to graph, you can get very good results.

In addition, I always hear a voice saying that XL's ControlNet model is not easy to use. As of May 2024, the SDXL ecosystem has more than 20 mature CN models, and the number is still increasing. 

Soft and hard edges, Openpose, depth and normal maps, line drawings, straight lines, graffiti, various face-changing plug-ins, and high-definition redrawing are all available. 

For problems such as artifacts and image quality degradation that are prone to occur in XL's CN controller model, it can be optimized by appropriately reducing CFG and increasing the number of steps.

后话

截止到2024年，还是有很多朋友对XL不是很了解，以为在线生图慢，出图需要记标签。

其实这些都是过时教程遗留下来的观念，首先今天我实测了Lib在线生图，XL和1.5出图速度相差无几。

其次，XL得益于较大的参数量和双文本层设计，其实根据底模特色，自身已经可以生成绝大多数绘图风格。即便是通过LoRA锁定了部分概念，也可以通过提示词任意变更所需的画面元素，lib在线还有自动翻译工具。SDXL就是介于商业模型和1.5社区模型之间的一种形态，不需要过多的LoRA组合，也不需要复杂的提示词、负面词，需要的只是你的创意和审美品味。即便是通过图生图重绘，也可以获得非常好的结果。

另外，还总听到一种声音，说XL的ControlNet模型不好用什么的。截至2024年5月，SDXL生态已经有20多款成熟的CN模型，数量还在不断增加。软硬边缘、Openpose、深度与法线贴图、线稿、直线、涂鸦、各种换脸插件、高清重绘一应俱全。对于XL的CN控制器模型容易出现的伪影，画质下降等问题，可以通过适当降低CFG，提高步数来优化。

----------------------------------------------------------------------

模型用途声明：

1. 您不得将此模型及其衍生版本（如融合模型版本）托管于计划赚取收入或捐赠的网站/应用程序。

2. 您不得直接售卖此模型及其衍生版本（如融合模型版本），除非您对此模型进行了足够程度的人工修改，使其在法律意义上可以被完全判定为您的个人作品。如果您违反本条，所造成的一切法律后果由您个人承担，请恕本人概不负责。

3. 您不能使用该模型故意制作或共享非法或有害的内容传播和输出，请您遵守公序良德，将此模型用于积极正面的用途。

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

做好工具人 服务艺术家

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