webmp3/Sakura-MiniMax-H3-GGUF

Sakura — MiniMax-H3 (FL2VA, pruned), GGUF quantizations

Community quantizations of the pruned FL2VA denoiser (the diffusion transformer, about 20 B parameters) of MiniMax-H3, the model that generates video with a synchronized stereo audio track. Made with stable-diffusion.cpp. Independent work, not made or endorsed by MiniMax or Comfy-Org.

This repository holds the denoiser and unmodified copies of the two VAEs. The Qwen3-VL text encoder is a separate file: an abliterated ("uncensored") build is available as a separate repository, Sakura-MiniMax-H3-Uncensored-GGUF; unchanged encoders in GGUF exist from other community members (for example leejet/MiniMax-H3-GGUF).

What you need to run it

Part	File	Source
Denoiser (pick one)	Sakura-MiniMax-H3-FL2VA-Pruned-Q4_K_M-11.71GiB.gguf (smallest), …Q5_K_M… (13.78 GiB), …Q6_K… (15.42 GiB) or …Q8_0… (19.94 GiB); passed with --diffusion-model	this repository
Video VAE	minimax_h3_video_vae_fp16.safetensors (4.85 GiB), passed with --vae	unmodified copy of the file of Comfy-Org/MiniMax-H3
Audio VAE	minimax_h3_audio_vae_fp32.safetensors (0.56 GiB), passed with --audio-vae; without it the video is generated without a decoded audio track	unmodified copy of the file of Comfy-Org/MiniMax-H3
Text encoder	one GGUF of the Qwen3-VL-32B encoder (50 layers), passed with --llm	see above
Program	sd-cli of stable-diffusion.cpp with MiniMax-H3 support (we used master-956-1b0ba10, Vulkan; documentation: docs/minimax_h3.md)	not included

ComfyUI: not tested by us; the community GGUF loader fork named in leejet's repository is the one to try.

Example

sd-cli -M vid_gen --diffusion-model Sakura-MiniMax-H3-FL2VA-Pruned-Q4_K_M-11.71GiB.gguf --vae minimax_h3_video_vae_fp16.safetensors --audio-vae minimax_h3_audio_vae_fp32.safetensors --llm <encoder>.gguf -p "A red fox walks through fresh snow in a quiet pine forest at sunrise, soft golden light, cinematic tracking shot, realistic motion, gentle wind in the trees." --cfg-scale 1.0 --steps 8 -W 480 -H 288 --video-frames 22 --fps 24 --diffusion-fa --rng cpu -s 42 -o frames/f_%03d.png

We only tested the PNG-sequence output shown here (-o with a %03d pattern); the documentation of stable-diffusion.cpp lists .avi, .webm and animated .webp as video outputs.

On our machine (AMD Strix Halo, Vulkan, 32 GiB graphics memory plus 31.6 GiB system memory) one such run took 57 to 62 seconds with the Q4_K_M denoiser and 94 seconds with Q8_0 (wall-clock, including model loading). Memory: the Q8_0 denoiser together with a Q8_0 encoder (sd-cli reported 20,506 + 26,221 MB of weights, plus 5,558 MB for the VAEs) did not fit and failed with ErrorOutOfDeviceMemory; the Q8_0 denoiser with a Q4_K_M encoder ran, as did the other combinations we tested (every denoiser with a Q4_K_M encoder; the Q4_K_M denoiser with Q5_K_M or Q6_K encoders).

Example clips

One prompt (the one in the example above), seed 42, 22 frames at 480x288 (less than one second at 24 fps), 8 steps, text encoder fixed at our Q4_K_M build, one clip per denoiser. The GIFs show the decoded frames only; there is no audio in them. They are tiny test renders for comparison, not a statement of the full quality of the model.

Q4_K_M	Q5_K_M
Q6_K	Q8_0

What we measured (and what not)

Test: one prompt (the one in the example), seed 42, 22 frames at 480x288, 8 steps, --cfg-scale 1.0, text encoder fixed at our abliterated Q4_K_M build (made without a working importance matrix), denoiser changed; PSNR and SSIM of the decoded frames against the same run with the Q8_0 denoiser (mean over the 22 frames, SSIM on luma in 8x8 windows):

Denoiser	PSNR	SSIM
Q6_K	18.16 dB	0.713
Q5_K_M	15.55 dB	0.656
Q4_K_M	13.88 dB	0.595

The same Q5_K_M and Q4_K_M quantizations without the protected tensors measured 15.36 dB / 0.645 and 13.13 dB / 0.544; those unprotected files are not published.

How to read this: the absolute values are low because every change in the weights changes the sampled video, not only its sharpness; use the table for the ordering (Q6_K is closest to Q8_0, and protecting the sensitive tensors, which the 4- and 5-bit files do, helped a little: 0.75 dB at 4 bit, and 0.19 dB at 5 bit, which is within what a single seed can show by chance). It is one prompt and one seed, and we looked at single frames of two runs only; we did not assess motion, audio, longer clips or higher resolution, and we did not compare against the unpruned BF16 model or against other people's quantizations.

Files and how they were made

Files: Q4_K_M, Q5_K_M, Q6_K, Q8_0 and the two VAEs.

File	Size	What it is
…-Q8_0-19.94GiB.gguf	19.94 GiB	sd-cli -M convert --type q8_0 from the pruned FL2VA BF16 file
…-Q6_K-15.42GiB.gguf	15.42 GiB	--type q6_K --imat-in <imatrix>
…-Q5_K_M-13.78GiB.gguf	13.78 GiB	--type q5_K --imat-in <imatrix> --tensor-type-rules 'condition_proj\.weight$=bf16,token_refiner\.=q8_0,blocks\.(46|47|48|49)\.=q8_0'
…-Q4_K_M-11.71GiB.gguf	11.71 GiB	--type q4_K --imat-in <imatrix> --tensor-type-rules with the same rules
minimax_h3_video_vae_fp16.safetensors, minimax_h3_audio_vae_fp32.safetensors	4.85 / 0.56 GiB	unmodified copies
LICENSE, NOTICE, SHA256SUMS		MiniMax H3 Community License Agreement, notices, checksums

• Importance matrix: the published H3 matrix minimax_h3_fl2va_pruned.imatrix of MarxistLeninist, used as published. A control conversion of Q4_K_M without it gives a different file (different sha256), so it is applied at least to part of the tensors (checked for Q4_K_M only); we did not measure how much quality it adds. The file names Q4_K_M and Q5_K_M follow common practice; sd-cli quantizes all large weights uniformly to q4_K or q5_K (no mixed Q6_K layers as in llama.cpp), small tensors stay F16, BF16 or F32.
• Protected tensors (Q4_K_M and Q5_K_M): three rules taken from the same author's QF2 recipe: condition_proj.weight in BF16, token refiner and the last four blocks (46 to 49) in Q8_0 (this also stores their small q/k-norm weights as Q8_0 instead of BF16). Their low-bit rules for mlp.fc1, attn.qkv_proj/attn.out_proj and mlp.fc2 are not used. All other weights, including the norm1/norm2 weights, are quantized like the rest of the file. The unprotected versions of these two files were built too, measured as listed above and left out of the repository; the gain in our single-seed test is small.
• Source: diffusion_models/minimax_h3_fl2va_pruned_bf16.safetensors of Comfy-Org/MiniMax-H3 (revision e5eb578a89295337b8ff433a035929ce0279e0b6, sha256 checked against the Hub), a pruned derivative of MiniMaxAI/MiniMax-H3 (reference revision we checked: 42ed227ee7df40d41602854ae760620d6eb651fe). We did not verify that the pruning leaves the output unchanged.
• Not included: the REF2VA denoiser, LoRAs, ControlNet.

What we changed (license section III.2)

All .gguf files are modified files: weights of the pruned FL2VA denoiser converted from BF16 safetensors to GGUF and quantized as listed above. The VAE files are unmodified copies.

License and use

MiniMax H3 Community License Agreement (LICENSE), including the Acceptable Use Policy (Exhibit A) which also binds you: among other things no use outside the Applicable Territory, no use to circumvent safeguards, no use to exploit or harm minors, no use for impersonation or deception without disclosure, and machine-generated content must be disclosed in public environments. Commercial use above 20 million US dollars of yearly revenue needs a separate authorization from MiniMax; commercial products must display "MiniMax H3". Section V.3 also forbids using the models or their outputs to improve any other AI model, and Section V.5 requires safeguards if you offer generation as a service. By downloading or using these files you agree to LICENSE, including Section V and Exhibit A; these restrictions apply to you. You must pass LICENSE and NOTICE on with any redistribution. "Powered by MiniMax H3" is encouraged by the license.

Credits

• Model: MiniMaxAI/MiniMax-H3 (MiniMax); pruned BF16 file and VAEs: Comfy-Org/MiniMax-H3.
• Importance matrix and the QF protection pattern: MarxistLeninist.
• Conversion and inference: leejet/stable-diffusion.cpp (MIT).

---

中文说明 · 樱花 (Simplified Chinese)

English above. 本节为上文的中文翻译(Sakura = 樱花 yīnghuā);完整的独立中文版见 README_zh.md。

Sakura — MiniMax-H3(FL2VA,剪枝版),GGUF 量化

这是 MiniMax-H3(可生成带同步立体声音轨的视频的模型)剪枝版 FL2VA 去噪器(扩散 Transformer,约 20 B 参数)的社区量化版本,使用 stable-diffusion.cpp 制作。属于独立工作,并非 MiniMax 或 Comfy-Org 制作或认可。

本仓库包含 去噪器 以及两个 VAE 的未改动副本。Qwen3-VL 文本编码器是单独的文件:经消融(“去审查”)的版本以单独仓库提供:Sakura-MiniMax-H3-Uncensored-GGUF;未改动编码器的 GGUF 可由其他社区成员提供(例如 leejet/MiniMax-H3-GGUF)。

运行所需文件

部分	文件	来源
去噪器(任选一个)	Sakura-MiniMax-H3-FL2VA-Pruned-Q4_K_M-11.71GiB.gguf(最小)、…Q5_K_M…(13.78 GiB)、…Q6_K…(15.42 GiB)或 …Q8_0…(19.94 GiB);通过 --diffusion-model 传入	本仓库
视频 VAE	minimax_h3_video_vae_fp16.safetensors(4.85 GiB),通过 --vae 传入	Comfy-Org/MiniMax-H3 中该文件的未改动副本
音频 VAE	minimax_h3_audio_vae_fp32.safetensors(0.56 GiB),通过 --audio-vae 传入;不传时生成的视频没有解码出的音轨	Comfy-Org/MiniMax-H3 中该文件的未改动副本
文本编码器	Qwen3-VL-32B 编码器(50 层)的一个 GGUF,通过 --llm 传入	见上文
程序	支持 MiniMax-H3 的 stable-diffusion.cpp 的 sd-cli(我们使用 master-956-1b0ba10,Vulkan;文档:docs/minimax_h3.md)	不包含

ComfyUI:我们没有测试;请尝试 leejet 仓库中提到的社区 GGUF 加载器分支。

示例

sd-cli -M vid_gen --diffusion-model Sakura-MiniMax-H3-FL2VA-Pruned-Q4_K_M-11.71GiB.gguf --vae minimax_h3_video_vae_fp16.safetensors --audio-vae minimax_h3_audio_vae_fp32.safetensors --llm <encoder>.gguf -p "A red fox walks through fresh snow in a quiet pine forest at sunrise, soft golden light, cinematic tracking shot, realistic motion, gentle wind in the trees." --cfg-scale 1.0 --steps 8 -W 480 -H 288 --video-frames 22 --fps 24 --diffusion-fa --rng cpu -s 42 -o frames/f_%03d.png

我们只测试了这里给出的 PNG 序列输出(带 %03d 模式的 -o);stable-diffusion.cpp 的文档列出 .avi、.webm 和动画 .webp 作为视频输出。

在我们的机器上(AMD Strix Halo,Vulkan,32 GiB 显存加 31.6 GiB 系统内存),这样一次运行用时:Q4_K_M 去噪器 57 到 62 秒,Q8_0 去噪器 94 秒(实际耗时,含模型加载)。内存: Q8_0 去噪器与 Q8_0 编码器同时使用(sd-cli 报告权重 20,506 + 26,221 MB,另有 VAE 5,558 MB)放不下,并以 ErrorOutOfDeviceMemory 失败;Q8_0 去噪器配 Q4_K_M 编码器可以运行,我们测试过的其他组合也可以(每个去噪器配 Q4_K_M 编码器;Q4_K_M 去噪器配 Q5_K_M 或 Q6_K 编码器)。

示例片段

一条提示词(上面示例中的那一条),种子 42;22 帧;分辨率 480x288(24 fps 下不到一秒),8 步,文本编码器固定为我们的 Q4_K_M 版本,每个去噪器一个片段。GIF 只显示解码后的画面,不含音频。它们是用于比较的小型测试渲染,不代表模型的完整质量。

Q4_K_M	Q5_K_M
Q6_K	Q8_0

我们测量了什么(以及没有测量什么)

测试:一条提示词(示例中的那一条),种子 42;22 帧;分辨率 480x288;8 步;--cfg-scale 1.0,文本编码器固定为我们经消融的 Q4_K_M 版本(制作时重要性矩阵并未真正生效),只更换去噪器;解码后帧的 PSNR 和 SSIM 与同样设置下使用 Q8_0 去噪器的运行相比(22 帧的平均值,SSIM 在亮度上以 8x8 窗口计算):

去噪器	PSNR	SSIM
Q6_K	18.16 dB	0.713
Q5_K_M	15.55 dB	0.656
Q4_K_M	13.88 dB	0.595

同样的 Q5_K_M 和 Q4_K_M 量化在不保护这些张量时测得 15.36 dB / 0.645 和 13.13 dB / 0.544;这些未保护的文件没有发布。

如何解读: 绝对数值偏低,因为权重的任何变化都会改变采样出的视频,而不仅仅是清晰度;请只用这张表看 排序(Q6_K 最接近 Q8_0,保护敏感张量(4 比特和 5 比特文件都这样做)有一点帮助:4 比特为 0.75 dB,5 比特为 0.19 dB,后者在单一种子下可能只是偶然)。这只是一条提示词、一个种子,而且我们只查看了两次运行中的单帧;我们 没有 评估运动、音频、更长的片段或更高的分辨率,也没有与未剪枝的 BF16 模型或其他人的量化版本比较。

文件及其制作方式

文件: Q4_K_M、Q5_K_M、Q6_K、Q8_0 和两个 VAE。

文件	大小	说明
…-Q8_0-19.94GiB.gguf	19.94 GiB	对剪枝版 FL2VA 的 BF16 文件执行 sd-cli -M convert --type q8_0
…-Q6_K-15.42GiB.gguf	15.42 GiB	--type q6_K --imat-in <imatrix>
…-Q5_K_M-13.78GiB.gguf	13.78 GiB	--type q5_K --imat-in <imatrix> --tensor-type-rules 'condition_proj\.weight$=bf16,token_refiner\.=q8_0,blocks\.(46|47|48|49)\.=q8_0'
…-Q4_K_M-11.71GiB.gguf	11.71 GiB	--type q4_K --imat-in <imatrix> --tensor-type-rules,使用同样的规则
minimax_h3_video_vae_fp16.safetensors、minimax_h3_audio_vae_fp32.safetensors	4.85 / 0.56 GiB	未改动的副本
LICENSE、NOTICE、SHA256SUMS		MiniMax H3 社区许可协议、声明、校验和

• 重要性矩阵: 使用 MarxistLeninist 发布的 H3 矩阵 minimax_h3_fl2va_pruned.imatrix,按原样使用。不使用它的 Q4_K_M 对照转换会得到不同的文件(sha256 不同),所以它至少对部分张量起了作用(只检查了 Q4_K_M);我们没有测量它带来多少质量提升。文件名 Q4_K_M 和 Q5_K_M 沿用惯例;sd-cli 把所有大权重统一量化为 q4_K 或 q5_K(没有 llama.cpp 那样的 Q6_K 混合层),小张量保持 F16、BF16 或 F32。
• 受保护的张量(Q4_K_M 和 Q5_K_M): 采用同一作者 QF2 方案中的三条规则:condition_proj.weight 用 BF16,token refiner 和最后四个块(46 到 49)用 Q8_0(其中的小型 q/k 归一化权重因此也由 BF16 变为 Q8_0)。不使用他们针对 mlp.fc1、attn.qkv_proj/attn.out_proj 和 mlp.fc2 的低比特规则。其余权重(包括 norm1/norm2)与文件其他部分一样量化。这两个文件也制作了未保护的版本,按上表测量后没有放入仓库;在我们单一种子的测试中增益很小。
• 来源: Comfy-Org/MiniMax-H3 中的 diffusion_models/minimax_h3_fl2va_pruned_bf16.safetensors(版本 e5eb578a89295337b8ff433a035929ce0279e0b6,sha256 已与 Hub 核对),它是 MiniMaxAI/MiniMax-H3(我们核对过的参考版本:42ed227ee7df40d41602854ae760620d6eb651fe)的剪枝衍生版本。我们没有验证剪枝是否不改变输出。
• 不包含: REF2VA 去噪器、LoRA、ControlNet。

我们做了哪些修改(许可第 III.2 条)

所有 .gguf 文件都是修改过的文件:剪枝版 FL2VA 去噪器的权重从 BF16 safetensors 转为 GGUF,并按上表量化。VAE 文件是未改动的副本。

许可与使用

MiniMax H3 社区许可协议(LICENSE),包括同样对您有约束力的可接受使用政策(附件 A):其中包括不得在适用地域之外使用、不得用于绕过防护措施、不得用于利用或伤害未成年人、不得用于冒充或未披露的欺骗,公开环境中的机器生成内容必须披露。年收入超过 20 million 美元的商业使用需要另行取得 MiniMax 的授权;商业产品必须显示“MiniMax H3”。第 V.3 条还禁止使用这些模型或其输出来改进任何其他 AI 模型,第 V.5 条要求在提供生成服务时落实防护措施。下载或使用这些文件即表示您同意 LICENSE,包括第 V 条和附件 A;这些限制对您有约束力。再分发时必须一并提供 LICENSE 和 NOTICE。许可鼓励标注“Powered by MiniMax H3”。

致谢

• 模型:MiniMaxAI/MiniMax-H3(MiniMax);剪枝 BF16 文件和 VAE:Comfy-Org/MiniMax-H3。
• 重要性矩阵和 QF 保护模式:MarxistLeninist。
• 转换与推理:leejet/stable-diffusion.cpp(MIT)。