{"schema":1,"mode":"public","generatedAt":"2026-10-07T16:25:46.835Z","site":{"title":"AI雷达","by":"","description":"Civitai 新模型早晚报：MiniMax H3 与 Qwen Image 2.1 每天北京 09:00 / 21:00 两期，附编辑推荐理由。","url":""},"policy":{"nsfw":"removed","dropped":["explicit"],"keywordBlock":true,"strict":false},"boards":[{"id":"minimax-h3","name":"MiniMax H3","short":"H3","baseModel":"MiniMax H3","core":true},{"id":"qwen-image-2-1","name":"Qwen Image 2.1","short":"Qwen 2.1","baseModel":"Qwen 2.1","core":false}],"latestId":"2026-10-06-AM","range":{"first":"2026-09-21","last":"2026-10-06"},"totals":{"entries":459,"editions":29},"types":["LoRA","VAE","其他","姿势","工作流","底模 / Checkpoint","文本编码器"],"months":[{"id":"2026-10","count":174},{"id":"2026-09","count":285}],"editions":[{"id":"2026-10-06-AM","date":"2026-10-06","half":"AM","boards":{"minimax-h3":{"since":"2026-10-05T14:34:50.302Z","cutoff":"2026-10-06T01:01:02.363Z","backfill":false,"late":false,"editor":"claude-cowork-scheduled","reviewedAt":"2026-10-06T01:23:03Z","note":"本期审阅 9 个 H3 新版本。首推 5tar5ystem 角色设定图生成器 v2.0：给它人物的照片 / 视频 / 音频并标明每份参考的用途，它在 ComfyUI 里一次渲染出整张多视图角色设定图，可直接回喂 ref2va 参考工作流；v2.0 改成“一次出整张”，各面板身份、光线、配色天然一致。其他亮点：wushu_action 武术动作 LoRA（触发词、权重区间、步数、帧数、分辨率、负面词都写全，但没有可见示例），以及 ComfyUI-MinimaxH3-Studio 节点包（结构化提示词、官方工作流一键生成、最长 60 秒分镜拼接和内置剪辑器）。Lora-Daddy H3 Studio 的工作流只是占位、要另买节点包，降为仅收录；Alek Deng、Hope Belwyn 是常规角色 LoRA，AsianMix 为付费实验版，Latex 服装 LoRA 示例全被隐藏，均为仅收录。","excluded":[],"count":9,"tiers":{"S":1,"A":1,"B":2,"C":5},"edited":9,"top":"2982722:3386699"},"qwen-image-2-1":{"since":"2026-10-05T14:34:50.302Z","cutoff":"2026-10-06T01:01:02.363Z","backfill":false,"late":false,"editor":"claude-cowork-scheduled","reviewedAt":"2026-10-06T01:23:03Z","note":"本期审阅 3 个 Qwen 2.1 新版本。首推 led_xiii 的裁切式快速局部重绘工作流 Qwen 2.1 版：只把蒙版附近的区域裁出来送进模型重绘再贴回原图，不把整张大图塞进显存，Qwen 2.1 版把重绘分辨率提到 768x768。EditFlow 图像编辑工作流说明只有两个插件下载链接，3d skin 风格 LoRA 没有触发词和用法说明，都降为仅收录。","excluded":[],"count":3,"tiers":{"S":0,"A":1,"B":0,"C":2},"edited":3,"top":"2407694:3386215"}}},{"id":"2026-10-05-PM","date":"2026-10-05","half":"PM","boards":{"minimax-h3":{"since":"2026-10-05T06:12:20.334Z","cutoff":"2026-10-05T15:04:50.302Z","backfill":false,"late":false,"editor":"claude-cowork-scheduled","reviewedAt":"2026-10-05T15:09:00Z","note":"本期审阅 8 个 H3 新版本。首推 TP_AI_63 的 Frozen World：基础 H3 写“时间静止”时人和物还会继续动，这个 LoRA 让整个场景定住、只留一个人朝镜头走过来，配方（触发词、强度、分辨率、帧数、步数、提示词模板）写得很全，作者还配了免费工作流 H3 Camera Director。其他亮点：Motion Speed 滑杆用正负权重控制动作快慢（快进 / 慢动作），Darksidewalker 的 Turbo 合集新出 PDMD+DMAD 4 步混合版，HCSM 定格动画微缩风格出了 v2，Minecraft 光影方块风格 LoRA，以及 YuE2 × H3 音乐 MV 工作流 v2.19 新增音效模式。电影角色 Chel 的移植 LoRA 只是常规角色 LoRA，降为仅收录；Camera Director 主要是给 Frozen World 配的工作流，降为可关注。","excluded":[],"count":8,"tiers":{"S":1,"A":4,"B":2,"C":1},"edited":8,"top":"2986051:3385432"},"qwen-image-2-1":{"since":"2026-10-05T06:12:20.334Z","cutoff":"2026-10-05T15:04:50.302Z","backfill":false,"late":false,"editor":"claude-cowork-scheduled","reviewedAt":"2026-10-05T15:09:00Z","note":"本期审阅 8 个 Qwen 2.1 新版本。首推 TP_AI_63 的 Glamour Realism：一个约 80 MB 的轻量风格 LoRA，用来减少人像的塑料皮肤感和发灰的颜色，同时尽量保留提示词里的服装、姿势和构图，强度 0.6、无需触发词，用法写得很清楚。其他亮点：Disco Elysium 风格 LoRA 移植到 2.1，偏通用插画 / 油画而不只是头像；Character Sheet Designer 开源节点从一张参考图出多视图角色设定图；isHeSatoshi 的 6 步加速 LoRA 兼容版；Sidney Sime 与 Dream Loom 两个奇幻插画风格；以及一套五种照片编辑的教学工作流。Qwen image 2.1 Beautiful v2.0B 没说明它到底改变什么，降为仅收录。","excluded":[],"count":8,"tiers":{"S":1,"A":1,"B":5,"C":1},"edited":8,"top":"2986100:3385496"}}},{"id":"2026-10-05-AM","date":"2026-10-05","half":"AM","boards":{"minimax-h3":{"since":"2026-10-04T12:31:01.966Z","cutoff":"2026-10-05T06:42:20.334Z","backfill":false,"late":false,"editor":"claude-cowork","reviewedAt":"2026-10-05T06:48:43Z","note":"本期审阅 20 个 H3 新版本。首推 AusBoss 的 360 Orbit v1.1：一张照片生成绕主体一整圈、首尾相接可循环的环绕镜头，新版加了“周围环境”输入，纯色背景下不再像原地自转。长视频工作流集中更新：SEED HUNTER v2.5 加入台词 / 声音克隆，LongForge v4.7 可复用参考素材并新增接缝检查，Dual Segment 让 16 GB 显卡一次跑两段，Dasiwa Xtra v1.5 一键切换 Native / Turbo。风格类有片场幕后花絮、80 年代奇幻片和迪士尼文艺复兴画风 V2。几个只有角色设定或说明近乎为空的角色 LoRA、付费且无可见示例的潜水面罩 LoRA 降为仅收录。","excluded":[],"count":20,"tiers":{"S":1,"A":9,"B":3,"C":7},"edited":20,"top":"2976490:3384201"},"qwen-image-2-1":{"since":"2026-10-04T12:31:01.966Z","cutoff":"2026-10-05T06:42:20.334Z","backfill":false,"late":false,"editor":"claude-cowork","reviewedAt":"2026-10-05T06:45:48Z","note":"本期审阅 9 个 Qwen 2.1 新版本。首推 AusBoss 的换装工作流 v1.1：用第二张图里的衣服给照片中的人换装，不用画遮罩，脸、发型、姿势和背景不动，新版加了衣服图裁剪；同作者的 Edit + Realign 出了 6 步 Fast 版，另有漫画转真人编辑 LoRA、古典油画质感风格 LoRA 和一套从脸 + 体型 + 服装生成角色训练数据集的大型工作流。纹身底稿 LoRA 说明偏少，降为可关注；12 合 1 工作流只有功能目录、没有模型清单；古风人物 LoRA 只有一段角色描述，仅收录。","excluded":[],"count":9,"tiers":{"S":1,"A":4,"B":3,"C":1},"edited":9,"top":"2983126:3383775"}}},{"id":"2026-10-04-PM","date":"2026-10-04","half":"PM","boards":{"minimax-h3":{"since":"2026-10-04T03:06:45.347Z","cutoff":"2026-10-04T13:01:01.966Z","backfill":false,"late":false,"editor":"claude-cowork","reviewedAt":"2026-10-05T06:46:07Z","note":"本期审阅 4 个 H3 新版本。首推 Turbo LoRA 合集新加的 DMAD 4 步加速 LoRA：上传者把原作权重转成 ComfyUI 原生 LoRA 加载器能读的格式，作者给的设置是 6–8 步、强度 1.0，并称音频质量好于 larryvrh v4。另有一个用单张参考图出角色设定表的 H3 节点 + 最小工作流，刚发布、示例少，列为可关注。付费且说明只有一句的 Disney Style 和只有人设自述的 Francesca 人物 LoRA 仅收录。","excluded":[],"count":4,"tiers":{"S":0,"A":1,"B":1,"C":2},"edited":4,"top":"2837571:3382801"},"qwen-image-2-1":{"since":"2026-10-04T03:06:45.347Z","cutoff":"2026-10-04T13:01:01.966Z","backfill":false,"late":false,"editor":"claude-cowork","reviewedAt":"2026-10-05T06:46:07Z","note":"本期审阅 4 个 Qwen 2.1 新版本。首推 AusBoss 同时发布的换装套件：Outfit Swap Consistency LoRA 让 Qwen 2.1 双图换装时只改衣服、脸和背景不动（作者给了 16 组对比数据），Outfit Swap 工作流内置这个 LoRA，再加上自动写服装指令、按人物照取景决定换多少。另有给编辑提示词加短语权重、给参考图单独调权重的 Enhancer 节点包（上调为推荐），以及参数齐全的《十字军之王 3》风格 LoRA。","excluded":[],"count":4,"tiers":{"S":2,"A":2,"B":0,"C":0},"edited":4,"top":"2983126:3381878"}}},{"id":"2026-10-04-AM","date":"2026-10-04","half":"AM","boards":{"minimax-h3":{"since":"2026-10-03T12:31:02.040Z","cutoff":"2026-10-04T03:36:45.347Z","backfill":false,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-04T04:10:00Z","note":"本期审阅 13 个 H3 新版本。首推用 ref2va 从自有照片 / 视频 / 音频生成角色设定表的 Character Sheet Builder 节点包；另有按参考视频动作和镜头换角色的 BUNNY 动作迁移工作流、迪士尼黄金时代动画风格 LoRA、微缩超现实风格 Tiny Things，以及一句提示词出原创歌曲加两场景 MV 的 YuE2 × H3 工作流（依赖重）。一个只有角色自述的人物 LoRA 和一个缺 H3 细节的通用工作流合集已降为仅收录。","excluded":[],"count":14,"tiers":{"S":1,"A":3,"B":6,"C":4},"edited":12,"top":"2982722:3381358"},"qwen-image-2-1":{"since":"2026-10-03T12:31:02.040Z","cutoff":"2026-10-04T03:36:45.347Z","backfill":false,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-04T04:05:00Z","note":"本期两条移植 LoRA 最值得看：QR Code Monster 光学错觉生成器的 Qwen 2.1 编辑版（输入图的轮廓转成云、风景等错觉画面），以及参数齐全的苏联动画《Ну Погоди!》赛璐璐风格 LoRA；另有一个把底模、文本编码器、VAE 和若干 LoRA 合成单文件的 AIO 底模，但未列出融合了哪些 LoRA。","excluded":[],"count":5,"tiers":{"S":2,"A":0,"B":1,"C":2},"edited":5,"top":"2432921:3381026"}}},{"id":"2026-10-03-PM","date":"2026-10-03","half":"PM","boards":{"minimax-h3":{"since":"2026-10-03T03:11:04.383Z","cutoff":"2026-10-03T13:01:02.040Z","backfill":false,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-04T03:50:00Z","note":"本期审阅 5 个 H3 新版本。最实用的是一个面向 12 GB 显存的 ref2va / fl2va 二合一快速工作流，作者给了与普通 4 步流程的耗时对比；另有把 H3 文生 / 图生 / 参考生视频接进 Telegram 机器人的 ComfyUI 工作流。同一作者的 3 个 AI 网红角色 LoRA 只有角色自述、没有触发词和用法，已降为仅收录。","excluded":[],"count":5,"tiers":{"S":0,"A":1,"B":1,"C":3},"edited":5,"top":"2981319:3379615"},"qwen-image-2-1":{"since":"2026-10-03T00:30:00.000Z","cutoff":"2026-10-03T13:01:02.040Z","backfill":false,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-04T03:50:00Z","note":"本期亮点是两个说明极完整的编辑工作流：AusBoss 的人物/换头工作流（可指定替换多人中的某一位）和只用原生节点的去水印工作流（只回贴改动区域、保留原分辨率）；另有 HDR VAE 的 ComfyUI 格式转换，以及 Qwen Image 2.1 Fix 的两个新版本，但后者没有任何效果说明。","excluded":[],"count":7,"tiers":{"S":2,"A":0,"B":4,"C":1},"edited":7,"top":"2980685:3378825"}}},{"id":"2026-10-03-AM","date":"2026-10-03","half":"AM","boards":{"minimax-h3":{"since":"2026-10-02T03:33:02.220Z","cutoff":"2026-10-03T03:41:04.383Z","backfill":false,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T03:42:28.769Z","note":"本期 17 个 H3 新版本。最值得试的仍是爆炸图拆解镜头和 2D↔写实风格滑杆；新增一套三版本角色替换工作流（标准 / 面部修复 / 静音），作者在 Civitai 把基模标成了 Other，由搜索通道找回。","excluded":[],"count":17,"tiers":{"S":2,"A":2,"B":7,"C":6},"edited":17,"top":"964312:3378342"},"qwen-image-2-1":{"since":"2026-10-02T13:00:00.000Z","cutoff":"2026-10-03T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T09:20:00Z","note":"本期亮点是两个参数齐全的风格 LoRA：Danrisi 的 2000s Analog Core（VHS / 胶片摄像机质感）移植到 Qwen 2.1，以及仙侠古风 LoRA 的 v3.0 均衡版；另有从参考图生成姿势人台的 Mannequins v2.0，但暂无可展示示例。","excluded":[],"count":10,"tiers":{"S":2,"A":1,"B":5,"C":2},"edited":8,"top":"1134895:3378081"}}},{"id":"2026-10-02-PM","date":"2026-10-02","half":"PM","boards":{"minimax-h3":{"since":"2026-10-02T01:00:00.000Z","cutoff":"2026-10-02T13:00:00.000Z","backfill":true,"late":false,"editor":"claude-cowork","reviewedAt":"2026-10-05T06:46:07Z","note":"唯一的条目是 RedCraft 付费套餐里的 H3 checkpoint（ASOP H3 Director），模型卡以付费访问和私有推理框架宣传为主，没说明这个 H3 权重改了什么，维持仅收录。本期无推荐。","excluded":[],"count":1,"tiers":{"S":0,"A":0,"B":0,"C":1},"edited":1,"top":null},"qwen-image-2-1":{"since":"2026-10-02T01:00:00.000Z","cutoff":"2026-10-02T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T09:10:00Z","note":"本期只有 1 个新版本：角色设定图工作流 V3.0，作者称这是其 Qwen Image 2.1 系列的最终更新，把提示词增强模型换成 Qwen 3.5 4B 并合并了两套工作流。","excluded":[],"count":1,"tiers":{"S":0,"A":1,"B":0,"C":0},"edited":1,"top":"2960750:3377047"}}},{"id":"2026-10-02-AM","date":"2026-10-02","half":"AM","boards":{"minimax-h3":{"since":"2026-10-01T13:00:00.000Z","cutoff":"2026-10-02T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T09:15:00Z","note":"本期 18 个 H3 新版本，编辑审阅 10 条。首推 Sketch IRL：真实场景里的人物变成手绘素描风，触发词、强度、示例提示词和避坑提示齐全；QuantFunc 真 4-bit（A4W4）量化底模给出了速度对比，但需专用后端且许可不允许商用；另有三段续写拼接长视频工作流和带运镜控制 / 自动提示词的俄语 I2V 工作流。","excluded":[],"count":12,"tiers":{"S":1,"A":3,"B":4,"C":4},"edited":9,"top":"2977672:3375087"},"qwen-image-2-1":{"since":"2026-10-01T13:00:00.000Z","cutoff":"2026-10-02T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T09:05:00Z","note":"本期以基础设施为主：QuantFunc 的 A4W4 INT4 量化底模（需专用加载器）、Qwen_VL_8B 文本编码器多精度量化，以及角色数据集制作工作流 v2.0；Asian Mix 人像 LoRA 更新到 v3.0。","excluded":[],"count":6,"tiers":{"S":0,"A":1,"B":3,"C":2},"edited":5,"top":"1905553:3374888"}}},{"id":"2026-10-01-PM","date":"2026-10-01","half":"PM","boards":{"minimax-h3":{"since":"2026-10-01T01:00:00.000Z","cutoff":"2026-10-01T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T09:05:00Z","note":"本期 12 个 H3 新版本，编辑审阅 9 条。首推一键 360° 环绕首尾帧循环工作流，依赖清单和质量/速度取舍写得非常完整；长视频工作流 LongForge 4.0 新增音频软衔接，时序超分概念工作流和 1950 年代西部片风格 LoRA 也有清楚的用法；另有一套在 Civitai 标成 Other 的 H3 换人工作流。","excluded":[],"count":10,"tiers":{"S":1,"A":3,"B":3,"C":3},"edited":9,"top":"2976490:3373562"},"qwen-image-2-1":{"since":"2026-10-01T01:00:00.000Z","cutoff":"2026-10-01T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:55:00Z","note":"本期两条强烈推荐：照片转 360° 全景的编辑 LoRA（参数完整、可照做）和 MoreReal 二次元转写实 LoRA；另有去焦外 LoRA 和 SinoX 转写实 v1.1，转写实方向本期有两个可对比的选择。","excluded":[],"count":5,"tiers":{"S":2,"A":2,"B":1,"C":0},"edited":5,"top":"2904424:3374264"}}},{"id":"2026-10-01-AM","date":"2026-10-01","half":"AM","boards":{"minimax-h3":{"since":"2026-09-30T13:00:00.000Z","cutoff":"2026-10-01T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:55:00Z","note":"本期 9 个 H3 新版本，编辑审阅 5 条。首推手工定格动画微缩场景风格 HCSM v1.9，触发词、权重和提示词结构齐全、示例多；其次是把纯文字分镜转成 H3 Continuum 长视频时间线的 Beatbox 工具，以及“镜头锁定跟随主体”的稳定 LoRA。","excluded":[],"count":5,"tiers":{"S":1,"A":2,"B":1,"C":1},"edited":5,"top":"2971871:3371156"},"qwen-image-2-1":{"since":"2026-09-30T13:00:00.000Z","cutoff":"2026-10-01T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:45:00Z","note":"本期重点是图像修复与编辑工具：去高斯模糊 LoRA 发布约 61 小时 824 次下载；AusBoss 一次更新了编辑、编辑+对齐、扩图、旋转扩图 4 个 Qwen 2.1 工作流，参数和依赖都写清了。","excluded":[],"count":11,"tiers":{"S":0,"A":3,"B":4,"C":4},"edited":8,"top":"2975487:3372351"}}},{"id":"2026-09-30-PM","date":"2026-09-30","half":"PM","boards":{"minimax-h3":{"since":"2026-09-30T01:00:00.000Z","cutoff":"2026-09-30T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:50:00Z","note":"本期 12 个 H3 新版本，编辑审阅 4 条。最值得试的是 ORB360 CardSpin 第 2 版：一个 LoRA 两种镜头（照片卡片翻转 / 360° 环绕），用法和提示词写得很完整；另有一个 BFS 换头工作流和 Pocket RPG 卡通风格的 Ref2Vid 版本。","excluded":[],"count":5,"tiers":{"S":0,"A":1,"B":2,"C":2},"edited":3,"top":"2969628:3370581"},"qwen-image-2-1":{"since":"2026-09-30T01:00:00.000Z","cutoff":"2026-09-30T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:35:00Z","note":"本期安全/混合级只有 1 个可评条目：LadyBug 角色 LoRA 的 Qwen 2.1 移植，参数给全但属于常规角色 LoRA，降为仅收录。","excluded":[],"count":3,"tiers":{"S":0,"A":0,"B":0,"C":3},"edited":1,"top":null}}},{"id":"2026-09-30-AM","date":"2026-09-30","half":"AM","boards":{"minimax-h3":{"since":"2026-09-29T13:00:00.000Z","cutoff":"2026-09-30T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:45:00Z","note":"本期 10 个 H3 新版本，编辑审阅 3 条。最值得看的是从 Krea 2 移植的点彩（Stipple Dot）风格 LoRA，示例视频多；另有一个 INT6 量化底模给显存卡在 INT8 边缘的人，以及一套作者写了详细步骤的视频换脸工作流。","excluded":[],"count":5,"tiers":{"S":0,"A":1,"B":2,"C":2},"edited":3,"top":"2834973:3370010"},"qwen-image-2-1":{"since":"2026-09-29T13:00:00.000Z","cutoff":"2026-09-30T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T09:30:00Z","note":"本期以工作流和整模为主：Ray 的编辑强化整模参数给得最全，角色数据集工作流把设定图扩展成成批训练图；另有 MXFP8 量化和 Studio 工作流 v0.2 等更新。","excluded":[],"count":6,"tiers":{"S":0,"A":2,"B":3,"C":1},"edited":6,"top":"2972899:3369126"}}},{"id":"2026-09-29-PM","date":"2026-09-29","half":"PM","boards":{"minimax-h3":{"since":"2026-09-29T01:00:00.000Z","cutoff":"2026-09-29T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T11:45:00Z","note":"本期审阅 5 个 H3 新版本，以长视频工作流更新为主：Endless v1.5 加了项目目录和过曝修复精修，Continuum v3.9 支持按片段指定参考图（@R1–@R9）；风格方面有手工定格微缩世界「HCSM」和 2003 版《少年泰坦》画风 + 角色 LoRA。","excluded":[],"count":8,"tiers":{"S":0,"A":4,"B":1,"C":3},"edited":5,"top":"2971649:3367577"},"qwen-image-2-1":{"since":"2026-09-29T01:00:00.000Z","cutoff":"2026-09-29T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T09:20:00Z","note":"本期值得看的是 SinoX 的动漫转写实 LoRA（触发词明确、7 张示例）和 Qwen Image 2.1 的 INT6 量化整模；一个“Beautiful”LoRA 说明过短，难以判断用途。","excluded":[],"count":4,"tiers":{"S":0,"A":1,"B":2,"C":1},"edited":3,"top":"2972320:3368391"}}},{"id":"2026-09-29-AM","date":"2026-09-29","half":"AM","boards":{"minimax-h3":{"since":"2026-09-28T13:00:00.000Z","cutoff":"2026-09-29T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T11:30:00Z","note":"本期审阅 7 个 H3 新版本，以更新为主。最值得用的是 BUNNY 动作连续性修复 V2（专攻空翻、旋转、倒立等 H3 最容易崩的动作，作者给了单用 / 搭配 / 二采三档权重），以及 MICW 角色设定图工作流 Rev3（新增画质更好的紧凑模式）；黄金太阳像素风 V2 把训练视频增加到 50 段。","excluded":[],"count":10,"tiers":{"S":1,"A":2,"B":3,"C":4},"edited":7,"top":"2890788:3366092"},"qwen-image-2-1":{"since":"2026-09-28T13:00:00.000Z","cutoff":"2026-09-29T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T09:10:00Z","note":"本期可推荐的只有一套 Studio 工作流：一次加载模型，集成文生图、单图/多图编辑和姿势编辑，配置清单写得完整。","excluded":[],"count":3,"tiers":{"S":0,"A":1,"B":0,"C":2},"edited":1,"top":"2968132:3363156"}}},{"id":"2026-09-28-PM","date":"2026-09-28","half":"PM","boards":{"minimax-h3":{"since":"2026-09-28T01:00:00.000Z","cutoff":"2026-09-28T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T11:15:00Z","note":"本期审阅 5 个 H3 新版本，两个新能力最值得看：「Real Stereo Depth」从单张图生成左右眼立体视频（交叉眼观看），「ORB360 CardSpin」用一个 LoRA 在 Ref2VA 里做照片卡片翻转和 360° 环绕两种镜头；另有墨线赛璐珞「ta86_inkriot」和 80 年代模拟质感科幻风格的移植版。","excluded":[],"count":5,"tiers":{"S":1,"A":3,"B":1,"C":0},"edited":5,"top":"2969930:3365446"},"qwen-image-2-1":{"since":"2026-09-28T01:00:00.000Z","cutoff":"2026-09-28T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T09:00:00Z","note":"本期亮点是 AnyAngle：借 3D/高斯泼溅在 Blender 里摆机位，再用 LoRA 把任意视角迁回原图画风；另有带“回对齐”的编辑工作流和角色设定图生成工作流 v2.1。","excluded":[],"count":5,"tiers":{"S":0,"A":3,"B":1,"C":1},"edited":4,"top":"2970214:3365794"}}},{"id":"2026-09-28-AM","date":"2026-09-28","half":"AM","boards":{"minimax-h3":{"since":"2026-09-27T13:00:00.000Z","cutoff":"2026-09-28T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T11:05:00Z","note":"本期审阅 12 个 H3 新版本。最值得试的是同一作者的两款 1920 年代默片 LoRA：「Aged Celluloid」把画面做成带划痕、闪烁、片门抖动的老胶片，「Aged Intertitle」专做默片字幕卡，两者可半强度叠用；另有暖金色低光风格 Afterlight 和带动作与背景训练的日式校园天台场景 LoRA。几个说明很短的角色 / 概念 LoRA 已降为仅收录。","excluded":[],"count":12,"tiers":{"S":1,"A":4,"B":5,"C":2},"edited":11,"top":"2968359:3363446"},"qwen-image-2-1":{"since":"2026-09-27T13:00:00.000Z","cutoff":"2026-09-28T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:50:00Z","note":"本期以编辑类 LoRA 为主：Consistency LoRA 解决 Qwen Image 2.1 改图时画面整体漂移，并给了有/无 LoRA 的量化对照；另有物体融合、漫画上色和真人化移植版。","excluded":[],"count":9,"tiers":{"S":2,"A":2,"B":2,"C":3},"edited":6,"top":"2548669:3363920"}}},{"id":"2026-09-27-PM","date":"2026-09-27","half":"PM","boards":{"minimax-h3":{"since":"2026-09-27T01:00:00.000Z","cutoff":"2026-09-27T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T10:50:00Z","note":"本期审阅 6 个 H3 新版本。最实用的是女性第一人称 POV 辅助 LoRA（与早报的男性 POV 同一作者）和线框网格「3D World」风格 LoRA；另有一个借助 Ollama 视觉大模型自动写 Ref2VA 提示词的工作流。","excluded":[],"count":8,"tiers":{"S":0,"A":2,"B":2,"C":4},"edited":5,"top":"2967915:3362906"},"qwen-image-2-1":{"since":"2026-09-27T01:00:00.000Z","cutoff":"2026-09-27T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:40:00Z","note":"本期亮点是 AusBoss 的扩图 LoRA v2.0：流程、触发指令和采样参数写得最完整；另有两个 8 步加速合并（LoRA 与整模）和几套工作流（角色数据集、姿势迁移、自动提示词人像）。","excluded":[],"count":14,"tiers":{"S":1,"A":0,"B":7,"C":6},"edited":8,"top":"2967375:3362269"}}},{"id":"2026-09-27-AM","date":"2026-09-27","half":"AM","boards":{"minimax-h3":{"since":"2026-09-26T13:00:00.000Z","cutoff":"2026-09-27T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T09:10:00Z","note":"本期审阅 16 个 H3 新版本，工具类尤其多。最值得试的是「Anime IRL」（写实世界里只把人物脸和头发变成赛璐珞动漫风，触发词、强度、示例提示词齐全），以及两个提示词 / 参考工具：Beatbox 分镜提示词生成器和 Look Sheets 角色换装设定图；另有 3D 动画、黏土动画、赛璐珞等风格 LoRA。几个角色 LoRA 说明薄或没有可见示例，已降为仅收录。","excluded":[],"count":15,"tiers":{"S":1,"A":7,"B":5,"C":2},"edited":15,"top":"2966267:3360987"},"qwen-image-2-1":{"since":"2026-09-26T13:00:00.000Z","cutoff":"2026-09-27T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T07:35:00Z","note":"本期工具类最多：AusBoss 扩图 v1.1（缺 LoRA 直接报错）和新的“旋转 + 裁切 + 扩图”工作流、不依赖 LLM 的四视图角色设定表；LoRA 方面有给出文生图 / 编辑两套强度的妆容滑杆和肤色滑杆。","excluded":[],"count":10,"tiers":{"S":0,"A":4,"B":3,"C":3},"edited":8,"top":"2966369:3361108"}}},{"id":"2026-09-26-PM","date":"2026-09-26","half":"PM","boards":{"minimax-h3":{"since":"2026-09-26T01:00:00.000Z","cutoff":"2026-09-26T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T09:00:00Z","note":"本期审阅 2 个 H3 新版本。「LipSync V4」口型同步长视频工作流新增歌词字幕和磁盘分块时间线，参数写得很全；「1920s Horror」默片恐怖风格 LoRA 从多个底模移植到 H3，H3 上无需触发词。","excluded":[],"count":3,"tiers":{"S":0,"A":2,"B":0,"C":1},"edited":2,"top":"1371819:3359486"},"qwen-image-2-1":{"since":"2026-09-26T01:00:00.000Z","cutoff":"2026-09-26T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T07:25:00Z","note":"本期推荐 AusBoss 的 Qwen Image 2.1 任意方向扩图工作流（Qwen3-VL 自动写提示词 + 作者的扩图 LoRA 固定原图位置）。","excluded":[],"count":4,"tiers":{"S":0,"A":1,"B":1,"C":2},"edited":3,"top":"2964810:3359252"}}},{"id":"2026-09-26-AM","date":"2026-09-26","half":"AM","boards":{"minimax-h3":{"since":"2026-09-25T13:00:00.000Z","cutoff":"2026-09-26T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:55:00Z","note":"本期审阅 5 个 H3 新版本。最值得试的是「Wanted dead or alive」通缉令特效（句式、强度、17 段带工作流示例都齐）；行星环境 LoRA 附了可复现的提示词文件，头戴相机步行晃动 LoRA 从 Wan 2.2 移植；Turbo 加速 LoRA v2 和长视频重叠拼接工作流方向实用，但没有可展示示例。","excluded":[],"count":5,"tiers":{"S":1,"A":2,"B":2,"C":0},"edited":5,"top":"2963411:3357454"},"qwen-image-2-1":{"since":"2026-09-25T13:00:00.000Z","cutoff":"2026-09-26T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T07:15:00Z","note":"本期重点是角色设定表工作流 V2.0：改成面向视频模型的“表情 / 姿态 / 装备”分格设定表，并新增官方 PE 执行方式；另有可写胸牌文字的林地迷彩作训服 LoRA。其余多为角色 LoRA 或说明很少的工作流更新。","excluded":[],"count":9,"tiers":{"S":0,"A":1,"B":2,"C":6},"edited":4,"top":"2960750:3358399"}}},{"id":"2026-09-25-PM","date":"2026-09-25","half":"PM","boards":{"minimax-h3":{"since":"2026-09-25T01:00:00.000Z","cutoff":"2026-09-25T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:50:00Z","note":"本期审阅 6 个 H3 新版本。风格类最值得试的是黑白木刻风「Dark Life」（12 段示例、触发词完整）和毛绒世界「Plushee World」（整个场景一起毛绒化）；「Anime to Realism」动漫转真人更新了训练数据，作者给了多镜头提示词要点；FastH3 加速工作流新增带 Ultimate Upscale 的放大版。","excluded":[],"count":8,"tiers":{"S":0,"A":4,"B":0,"C":4},"edited":6,"top":"2961693:3355337"},"qwen-image-2-1":{"since":"2026-09-25T01:00:00.000Z","cutoff":"2026-09-25T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T07:05:00Z","note":"本期最实用的是 AusBoss 的两套入门工作流：一句话指令编辑（带前后对比滑块，给了位移测量和 Consistency LoRA 用法）和最简文生图；另有 Pruna 的 8 步加速 LoRA 镜像，作者称不需要 CFG。","excluded":[],"count":6,"tiers":{"S":0,"A":2,"B":2,"C":2},"edited":4,"top":"2962616:3356480"}}},{"id":"2026-09-25-AM","date":"2026-09-25","half":"AM","boards":{"minimax-h3":{"since":"2026-09-24T13:00:00.000Z","cutoff":"2026-09-25T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:45:00Z","note":"本期审阅 6 个 H3 新版本。工作流是亮点：「Simply Advanced」修复了强制音频和 <Picture 2> 断线，「一张照片 40 秒三段串联」把三段 Ref2VA 严格按顺序接成一个 MP4；风格 LoRA 中 80 年代动画开场风「Safari Eyes」和 GBA《黄金太阳》像素风说明详细。","excluded":[],"count":10,"tiers":{"S":0,"A":4,"B":1,"C":5},"edited":6,"top":"2099833:3355007"},"qwen-image-2-1":{"since":"2026-09-24T13:00:00.000Z","cutoff":"2026-09-25T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T06:55:00Z","note":"本期安全/混合内容只有两条且都偏配套资源：一个给姿态导演工作流用的“日本不良少年”姿态包，和一个说明很少的综合工作流更新，没有新的 Qwen Image 2.1 权重值得推荐。","excluded":[],"count":2,"tiers":{"S":0,"A":0,"B":1,"C":1},"edited":1,"top":null}}},{"id":"2026-09-24-PM","date":"2026-09-24","half":"PM","boards":{"minimax-h3":{"since":"2026-09-24T01:00:00.000Z","cutoff":"2026-09-24T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T08:40:00Z","note":"本期审阅 6 个 H3 新版本。最值得试的是「Anatomy Reveal」解剖透视特效（句式固定、示例多、带工作流）和 FastVideo 的 FastH3 加速工作流（T2V / I2V 两版，给了 4/6/8 步 sigma 表）；舞蹈 LoRA 有完整的防肢体丢失提示词，角色和光影 LoRA 说明较薄。","excluded":[],"count":6,"tiers":{"S":1,"A":3,"B":0,"C":2},"edited":6,"top":"2960135:3353413"},"qwen-image-2-1":{"since":"2026-09-24T01:00:00.000Z","cutoff":"2026-09-24T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T06:45:00Z","note":"本期以 Qwen Image 2.1 的基础组件和工作流为主：原版文本编码器（含 8G 显存用的 w4a8）、原版 VAE，以及从单张参考图出角色设定表的工作流；风格类可看天宫仙境建筑 LoRA。多个 LoRA 作者都强调 CFG 不能低于 3.0、不能叠加速 LoRA。","excluded":[],"count":9,"tiers":{"S":0,"A":3,"B":4,"C":2},"edited":8,"top":"2960962:3354438"}}},{"id":"2026-09-24-AM","date":"2026-09-24","half":"AM","boards":{"minimax-h3":{"since":"2026-09-23T13:00:00.000Z","cutoff":"2026-09-24T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T06:35:00Z","note":"本期审阅 8 个 H3 版本。最值得试的是用 Ref2VA 从一张参考图批量出角色设定图的 MICW Character Sheet Max，以及强度区间写得清楚的 Cartoon World 怪诞绘本风和参数齐全的温泉场景 LoRA「ROOMS : Onsen」；其余是 MisterMR「World」系列材质风格移植和重训的万圣节幽灵 LoRA。","excluded":[],"count":12,"tiers":{"S":0,"A":3,"B":5,"C":4},"edited":8,"top":"2926613:3352611"},"qwen-image-2-1":{"since":"2026-09-23T13:00:00.000Z","cutoff":"2026-09-24T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T06:45:00Z","note":"本期 Civitai 官方上架了 Qwen Image 2.1 基础模型文件包（bf16 / int8 扩散模型和多种精度的文本编码器）；Danrisi 继续移植 UltraReal 系列（GrainScape 胶片颗粒、Canon 1Ds 相机色彩），另有集成缓存加速与 VOSR 放大的 Lonecat 工作流。","excluded":[],"count":9,"tiers":{"S":0,"A":3,"B":5,"C":1},"edited":8,"top":"2960068:3353339"}}},{"id":"2026-09-23-PM","date":"2026-09-23","half":"PM","boards":{"minimax-h3":{"since":"2026-09-23T01:00:00.000Z","cutoff":"2026-09-23T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T06:25:00Z","note":"本期审阅 7 个 H3 版本。最值得试的是 h3image 图片编辑 / 局部重绘工作流 v2.0：同时支持 Mac 和 NVIDIA，依赖、耗时和显卡要求都写清楚了；其次是用 2700 段奇幻电影片段训练的 Cinematic Fantasy 风格 LoRA 和一套长视频串接工作流。","excluded":[],"count":8,"tiers":{"S":0,"A":3,"B":3,"C":2},"edited":7,"top":"2857494:3350771"},"qwen-image-2-1":{"since":"2026-09-23T01:00:00.000Z","cutoff":"2026-09-23T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T06:45:00Z","note":"本期工具向条目多：Viggle 4 步蒸馏加速 LoRA 的转载、NVFP4 量化底模、RunComfy 的换头和换装工作流；概念类亮点是 4 色 × 2 种光泽的不透明裤袜 LoRA，触发词和提示写法是本期最完整的。","excluded":[],"count":10,"tiers":{"S":0,"A":1,"B":5,"C":4},"edited":6,"top":"2958760:3351749"}}},{"id":"2026-09-23-AM","date":"2026-09-23","half":"AM","boards":{"minimax-h3":{"since":"2026-09-22T13:00:00.000Z","cutoff":"2026-09-23T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T06:15:00Z","note":"本期审阅 7 个 H3 版本。最实用的是把 H3 当图像编辑模型用的参考图编辑工作流 v2.0，提示词写法和支持分辨率都写清楚了；其余是长视频 / 综合工作流的小版本更新，以及 MisterMR「World」系列的三款材质风格移植。","excluded":[],"count":11,"tiers":{"S":0,"A":1,"B":6,"C":4},"edited":7,"top":"2833301:3349226"},"qwen-image-2-1":{"since":"2026-09-22T13:00:00.000Z","cutoff":"2026-09-23T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T06:45:00Z","note":"本期以 Danrisi 的 UltraReal 系列移植为主（Lenovo 写实风格与两版 NiceGirls 人像），另有图生图细节增强 LoRA 和视觉小说背景风格移植；“Qwen Image 2.1 Fix”没说清修复什么，暂列可关注。","excluded":[],"count":10,"tiers":{"S":0,"A":3,"B":3,"C":4},"edited":6,"top":"1662740:3349565"}}},{"id":"2026-09-22-PM","date":"2026-09-22","half":"PM","boards":{"minimax-h3":{"since":"2026-09-22T01:00:00.000Z","cutoff":"2026-09-22T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T06:05:00Z","note":"本期审阅 5 个 H3 版本。最值得试的是参数写全的草稿线条绘本风 LoRA「Rough Sketch Nature」和能做续写 / 前插 / 桥接 / 无缝循环的时间线工作流 OBVPM Timeline；BUNNY SelfLift 给出了 12GB 显存直出 1080P 的具体耗时。","excluded":[],"count":8,"tiers":{"S":1,"A":2,"B":2,"C":3},"edited":5,"top":"2956315:3348767"},"qwen-image-2-1":{"since":"2026-09-22T01:00:00.000Z","cutoff":"2026-09-22T13:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T06:45:00Z","note":"本期亮点是 lilylilith 的 Blender 视口系列：一个把图片转成材质预览 / 工作台 / 白模视图，另一个把工作台视图反推成完整 3D 渲染，可串成一条流程；另有可正可负的低清滑杆和一个作者称暂时只支持 diffusers 的 HDR VAE。","excluded":[],"count":10,"tiers":{"S":0,"A":1,"B":2,"C":7},"edited":4,"top":"2953686:3345569"}}},{"id":"2026-09-22-AM","date":"2026-09-22","half":"AM","boards":{"minimax-h3":{"since":"2026-09-21T13:00:00.000Z","cutoff":"2026-09-22T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T05:55:00Z","note":"本期审阅 14 个 H3 版本。最值得试的是视频局部编辑 LoRA「VFX-Edit」和参数写得很全的 90 年代港式武侠胶片风格；工具类里 Ref2VA 自带音频口型工作流和 Scene Builder 提示词节点用途明确、依赖少。","excluded":[],"count":20,"tiers":{"S":2,"A":4,"B":7,"C":7},"edited":14,"top":"2954203:3346199"},"qwen-image-2-1":{"since":"2026-09-21T13:00:00.000Z","cutoff":"2026-09-22T01:00:00.000Z","backfill":true,"late":false,"editor":"cursor-agent","reviewedAt":"2026-10-03T06:45:00Z","note":"本期入选一个把头像重绘成三视图角色设定图的单图编辑 LoRA，说明很短。","excluded":[],"count":3,"tiers":{"S":0,"A":1,"B":0,"C":2},"edited":1,"top":"2954346:3346353"}}}]}