AI News

After MiniMax-H3-Image was open-sourced, the community quantized it into 27 versions, with Q2_K compressed to 6.4G

AYi AI NotesUpdated Nov 15 20:44

English text is being prepared.

AI summary

MiniMax's open-source 33.1B image generation model MiniMax-H3-Image has an original size of 68.65 GB; the community has already produced 27 quantized versions, with over 110,000 downloads. The author compiled actual test feedback stating that the Q4 version is compressed to 10.78G, Q2_K to only 6.4G, and Unsloth's FP8 version is 18.8G with image quality close to the original. The model supports native 4096×4096 direct output, 6-step sampling without enabling CFG, and can take 16 reference images at once to maintain character consistency; ComfyUI's GGUF nodes and Mac's MLX weights are already ready.

Why read this

AI score 71

The original article lists the specific VRAM thresholds and image-quality trade-offs of the community's quantized versions, which can be used to judge the hardware cost of running 4K image generation locally.

Topics

Tags

Quoted post

AYi

@AYi_AInotes

Holy crap, everyone stand up! Claude Fable 5.2 has completely ascended to godhood—distilling a video foundation to draw still portraits is simply a one-sided massacre of the laws of physics, holy crap!! The Nine Forms of the AI Image Generation Pure-Desire Portrait Secret Art summarized by Fable 5.2, plus the exclusive top-tier prompts it helped me optimize, are being open-sourced and shared with everyone without reservation!! I can't even resist making a hot take, brothers: in AIGC, anyone still trying image-generation prompts line by line is the most miserable wage slave of the compute era!! This afternoon I fed the image-generation secret art summarized by Claude Fable 5.2 into MiniMax's newly released H3-Image, and the set of nine-grid at-home portraits it actually produced in testing left me stunned—truly awesome, it feels like my youth has come back, ahhh! In the past, running portraits, I was most afraid of the inflatable-doll fake-skin feel and stiff poses, but this time, running according to the aesthetic skeleton broken down by Fable 5.2, the realistic texture directly leaves traditional models several tiers behind. The first rule: define the persona with a specific age plus recognizable features—write an East Asian adult female around 25 years old, then lock down a tear mole, eyebrow shape, and hair length. The second rule: describe the figure from the front only once; leave the rest to the cut and fabric, such as a fitted white ribbed cotton thin-strap camisole, round neckline, hem tucked above the waistline. The third rule: use half-finished actions to create a candid feel—stretching, brushing hair, tying a ponytail; in the nine-grid, keep only two frames looking at the camera. The fourth rule: define the tone with light position and exposure; do not write a single word about focal length or aperture, only write that diffused morning light falls on the face through a white gauze curtain, and exposure is based on the face. The fifth rule: use close-ups to convey bone structure and skin; reserve one frame separately for the collarbone and strap, specifying visible tiny pores and fine vellus hair, no skin smoothing. The sixth rule: use a white scene to bring out an airy feel—white walls, white bedsheets, white gauze curtains; white occupies more than half the frame. The seventh rule: break down the film feel and write it separately—high-key slightly overexposed, low contrast, low saturation, fine grain; do not use a single vague "film look." The eighth rule: at the end, use an avoidance list to clamp down on AI plastic feel, specifying plastic skin, excessive skin smoothing, doll face, darkened face, and graying whites. This writing approach happened to coincide with @MiniMax_AI's official release today of H3-Image, an image generation model derived from H3, with text-to-image, image reference, and image editing capabilities, and 4K image generation directly down to 10 seconds!! Looking at the attached images, you can see how exaggerated it is: stray hair strands, the faint sheen on the collarbone, the ribbing of the cotton fabric, and that tiny mole under the eye are all crystal clear, and the face basically holds up across different poses. My understanding is that it stripped the audio from its own mature video foundation and distilled it; video models are understanding the real world's three-dimensional light and shadow continuity every day, so turning back to draw a still portrait is a dimensional-reduction harvest. Top-tier large models translate the aesthetics of a photography director into plain human language, and the video foundation produces an image in ten seconds; the full set of prompts has been placed in the comments—brothers, help yourselves and go run them. #H3Image

View original

卧槽,全体起立!Claude Fable 5.2彻底封神了,视频底座蒸馏画静止人像,简直是物理规律的单方面屠杀啊卧槽!! Fable 5.2总结的AI生图纯欲写真焚诀九式和帮我优化的独家顶级提示词无保留开源分享给大家!! 我甚至忍不住说个暴论兄弟们,玩AIGC谁还在一行行试错生图提示词,谁就是算力时代最苦逼的打工人!! 下午拿Claude Fable 5.2总结的生图焚决喂给MiniMax新出的H3-Image,实测跑出来的这组九宫格居家写真直接把我看懵了,真的牛逼,感觉我的青春都回来了啊啊啊! 以前跑人像最怕充气娃娃假皮感和死板姿势,但这次按Fable 5.2拆出来的审美骨架去跑,真实质感直接甩开传统模型几个档次, 第一条,用具体年龄加认人的特征定人设,写25岁左右的东亚成年女性,再锁死一颗泪痣、眉形和发长, 第二条,身材只正面写一次,剩下的全交给剪裁和面料,比如贴身的白色罗纹棉细肩带吊带、圆领口、下摆收在腰线上方, 第三条,用做到一半的动作制造抓拍感,伸懒腰、拨头发、扎马尾,九格里只留两格看镜头, 第四条,用光位和曝光定影调,焦段光圈一个字不写,只写漫射晨光穿过白纱帘落在脸上、曝光以脸为准, 第五条,用特写交代骨相和皮肤,单独留一格拍锁骨和肩带,写明可见细小毛孔和细绒毛、不磨皮, 第六条,用白场景托出空气感,白墙白床单白纱帘,白色占画面一半以上, 第七条,胶片感拆开写,高调略过曝、低对比、低饱和、细颗粒,一句笼统的film look都不要, 第八条,在最后用避免清单卡死AI塑料感,写明塑料皮肤、过度磨皮、娃娃脸、脸部发暗、白色发灰, 这套写法正好碰上了@MiniMax_AI今天正式发布的H3-Image,这是一款衍生自H3的图像生成模型,具备文本生成图像、图像参考和图像编辑能力,4K图像生成,直接干到10秒了!! 看配图就知道有多夸张,散乱的发丝、锁骨微光、棉织物的罗纹和眼下那颗小痣都清清楚楚,不同姿势下的脸也基本咬住了, 我的理解是,它把自家成熟的视频底座拆掉音频做了蒸馏,视频模型天天在理解真实世界的三维光影连续性,回过头画一张静止人像就是降维收割, 顶级大模型把摄影指导的审美翻译成人话,视频底座十秒出片,成套提示词已放在评论区,兄弟们自取去跑。 #H3Image

View original post

Body

卧槽,开源社区这帮大神又把大厂的硬件门槛给掀翻了, MiniMax 刚开源的 33.1B 生图大模型 MiniMax-H3-Image, 原版体积足足 68.65 GB,直接劝退 99% 的家用电脑, 结果才过了几天,就被社区硬生生榨出了 27 个量化版本,下载量直接干破 11 万次。

我这两天顺手翻了一下实测反馈,对咱们普通玩家来说,这波压缩后的体验真的太香了。

先说大家最关心的显存门槛,原版 68G 连 4090 跑起来都得喘两口气, 现在社区把 Q4 版本压到了 10.78G,极限一点的 Q2_K 甚至只有 6.4G,连轻薄款 Mac 都能强行跑起来,

要是你手里显存宽裕,直接上 Unsloth 出的 18.8G FP8 版本,画质几乎跟原版没差。

更有意思的是它的原生 4K 直出能力,以前在本地跑高清大图,经常得先出小图再挂放大器,

显存动不动当场爆掉,这个模型直接硬吞 4096 乘 4096 的 1600 万像素,哪怕是 21 比 9 的电影宽屏比例也是原生一步到位,省去了二道放大的涂抹感。

速度方面也把痛点拿捏得很死,官方直接蒸馏到了 6 步采样成图,连 CFG 都可以不用开,基本告别了以前跑一张图等半天的煎熬。

除此之外它还把生图和修图揉在了一起,支持一次性喂进 16 张参考图来死死锁住角色一致性,目前 ComfyUI 的 GGUF 节点和 Mac 的 MLX 权重都已经全部就绪了。

说实话,我一直觉得开源社区最迷人的地方就在这里,大厂负责把天花板捅破,

社区大神负责把庞然大物拆成普通人家里那台 8G、12G 老显卡也能顺滑吞下的小钢炮,

手头硬件有限又想在本地白嫖顶级 4K 画质的兄弟,这波现成权重真的可以去下一份折腾试试了。

Source

AYi AI Notes