Built with Qwen. Community quantization of Qwen/Qwen-Image-2.1-Turbo (8-step distilled, 7B DiT) with an abliterated text encoder. Independent work, not made or endorsed by Alibaba/Qwen. Non-commercial use only (Qwen RESEARCH LICENSE AGREEMENT, copy in LICENSE).
For adults only; you are responsible for what you generate and for the laws that apply to you.
Q8_0 image model and our 8bit uncensored text encoder, stable-diffusion.cpp (Vulkan, Radeon 8060S), 1024x1024, 8 steps, CFG 1.0, euler, official Turbo sigmas. One seed per prompt and no re-rolls; the prompts themselves were chosen from a larger set of candidates. About 77 s per image including model loading; the GPU memory in use on our machine peaked at 18.3 GiB while generating (Windows counter for the whole GPU, 1.5 GiB of it were in use before the run). The images were made with stable-diffusion.cpp, not with ComfyUI.
The prompts are neutral. These images say nothing about what the ablated text encoder changes for restricted content; they only show that ordinary prompts still work. The same nine prompts and seeds are used in the sibling repository Sakura-Qwen-Image-2.1-Turbo-GGUF with the official encoder, so the two sets can be compared.
Seeds 1202, 1302, 1305, 1306, 1105, 1103, 1503, 1506, 1406 in the order shown; the prompts are the captions of the images above.
Everything except the runtime is in this repository. Pick the column of the program you use (🟦 stable-diffusion.cpp, 🟧 ComfyUI):
| Part | 🟦 stable-diffusion.cpp | 🟧 ComfyUI |
|---|---|---|
| Image model (pick one) | Sakura-Image-2.1-Turbo-Q8_0-7.07GiB.gguf (best), Sakura-Image-2.1-Turbo-Q5_K_M-4.60GiB.gguf (better) or Sakura-Image-2.1-Turbo-Q4_K_M-3.77GiB.gguf (smallest), passed with --diffusion-model |
the same file, loaded with Unet Loader (GGUF) |
| Uncensored text encoder (pick one) | Sakura-TextEncoder-Qwen3VL-8B-Uncensored-8bit-8.11GiB.gguf (GGUF Q8_0, recommended: closest to the official encoder in our image test) or Sakura-TextEncoder-Qwen3VL-8B-Uncensored-4bit-4.68GiB.gguf (GGUF Q4_K_M, smaller), passed with --llm |
Sakura-TextEncoder-Qwen3VL-8B-Uncensored-bf16.safetensors (BF16, 16.3 GiB), loaded with the normal Load CLIP node, type qwen_image |
| VAE | qwen_image_2.1_vae_bf16.safetensors (0.63 GiB, unmodified copy of the official file, identical sha256), passed with --vae |
the same file, loaded with Load VAE |
| Program | sd-cli / sd-server of stable-diffusion.cpp master (we used commit 228c707, Vulkan build), not included |
a recent ComfyUI build with native Qwen-Image 2.1 support and a ComfyUI-GGUF loader that supports Qwen-Image 2.1 (custom node), not included |
The text encoder file names have no quantization label on purpose, so the Hub card lists only the image models.
.gguf) into ComfyUI/models/diffusion_models/ (or unet/), the VAE into ComfyUI/models/vae/, and the text encoder (the .safetensors file) into ComfyUI/models/text_encoders/.qwen_image (text encoder), the normal Load VAE, and the TextEncodeQwenImage21 node. Sampling: 8 steps, CFG 1.0, euler, with the Turbo sigmas 1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568, 0.0 (node "ManualSigmas" into "SamplerCustomAdvanced").Qwen-Image-2.1-Turbo_ComfyUI_workflow.json (drag it into ComfyUI; it is set to Load CLIP with Sakura-TextEncoder-Qwen3VL-8B-Uncensored-bf16.safetensors for the text encoder, the Q8_0 image model, 512x512 and seed 42, the settings we tested).Tested: the BF16 safetensors encoder Sakura-TextEncoder-Qwen3VL-8B-Uncensored-bf16.safetensors with the normal Load CLIP node and the Q8_0 image model produced a correct image at 512x512 (one prompt), as did the official encoder in the same file format; the two images are very similar (42.8 dB PSNR, one prompt). The Q5_K_M and Q4_K_M image models also ran in ComfyUI at 512x512 (with an earlier GGUF text-encoder set-up that is no longer offered).
In our ComfyUI test on a Radeon 8060S, sampling at 1024x1024 was extremely slow (more than five minutes per step, not investigated); 512x512 ran in seconds. stable-diffusion.cpp on the same machine needed about 95 s for a 1024x1024 image, so the sample images were made with that.
Weights in memory for Q5_K_M with the 8bit encoder: about 4.6 + 8.1 + 0.6 = 13.3 GiB (with the 4bit encoder: 4.6 + 4.7 + 0.6 = 10 GiB) (we did not measure peak VRAM for this exact combination).
| Part | What we did |
|---|---|
| Text encoder (Qwen3-VL-8B) | Edited with an abliteration method (settings chosen by us), then BF16 safetensors (ComfyUI), GGUF Q8_0 and Q4_K_M. Measured as a chat model: refusals on the 100 harmful evaluation prompts of the optimisation 100/100 -> 8/100, KL divergence of the first-token distribution 0.047 (BF16 checkpoint, before quantization). |
| Image model (7B DiT) | Only quantized (Q4_K, Q5_K, Q8_0) from the official Turbo weights. We did not change what the DiT can draw. |
What we did not measure: how much the ablated encoder changes the images. The refusal numbers describe the encoder used as a language model; in this pipeline it only supplies embeddings. We only compared neutral prompts (see below). No sample images of restricted content are included.
| File | Size | What it is |
|---|---|---|
Sakura-Image-2.1-Turbo-Q8_0-7.07GiB.gguf |
7.07 GiB | image model (DiT), Q8_0 |
Sakura-Image-2.1-Turbo-Q5_K_M-4.60GiB.gguf |
4.60 GiB | image model (DiT), uniform Q5_K recipe of sd.cpp (file name carries the standard label Q5_K_M so the Hub recognises it) |
Sakura-Image-2.1-Turbo-Q4_K_M-3.77GiB.gguf |
3.77 GiB | image model (DiT), uniform Q4_K recipe of sd.cpp (file name carries the standard label Q4_K_M) |
Sakura-TextEncoder-Qwen3VL-8B-Uncensored-8bit-8.11GiB.gguf |
8.11 GiB | our ablated text encoder, Q8_0 (recommended) |
Sakura-TextEncoder-Qwen3VL-8B-Uncensored-4bit-4.68GiB.gguf |
4.68 GiB | our ablated text encoder, Q4_K_M (smaller) |
Sakura-TextEncoder-Qwen3VL-8B-Uncensored-bf16.safetensors |
16.33 GiB | our ablated text encoder, BF16, standard ComfyUI format (single file, loads with the normal Load CLIP node) |
The VAE qwen_image_2.1_vae_bf16.safetensors in this repository is the unmodified official file from Comfy-Org/Qwen-Image-2.1.
sd-cli \
--diffusion-model Sakura-Image-2.1-Turbo-Q5_K_M-4.60GiB.gguf \
--vae qwen_image_2.1_vae_bf16.safetensors \
--llm Sakura-TextEncoder-Qwen3VL-8B-Uncensored-8bit-8.11GiB.gguf \
--sigmas "1.0,0.978453,0.95418,0.926626,0.89508,0.845148,0.704534,0.414568,0.0" \
--steps 8 --cfg-scale 1.0 --sampling-method euler --diffusion-fa --vae-tiling \
-W 1024 -H 1024 -p "your prompt" -o out.png
We ran and measured everything at 512x512 (8 steps, about 12 to 15 s per image on a Radeon 8060S with Vulkan); 1024x1024 is the usual size of the model but we did not test it. The sigma schedule is the official Turbo schedule (sample_sigmas of the model card). Use CFG 1.0 and 8 steps. sd-cli/sd-server of stable-diffusion.cpp master (commit 228c707 was used) load these files directly.
Same prompts, same seed (42), 512x512, 8 steps, same machine; reference is the Q8_0 file the two files were made from (DogukanUrker's). Eight neutral prompts (cat, street with neon signs, portrait, mountain lake, poster with text, kitchen, marble statue, beach). Higher is closer to Q8_0; a diffusion model changes details even with tiny weight changes, so these are deviation numbers, not a quality score.
| File | Size | PSNR vs Q8_0 | SSIM vs Q8_0 |
|---|---|---|---|
| Q5_K | 4.60 GiB | 24.77 dB | 0.906 |
| Q4_K | 3.77 GiB | 21.46 dB | 0.842 |
Our own refusal measurement on the encoder used as a chat model (llama.cpp, Q4_K_M, temperature 0, 120 tokens, 104 held-out harmful prompts of the mlabonne/harmful_behaviors test split, identical prompts and settings for both rows):
| Text encoder (Q4_K_M) | strict refusals | broad refusal markers |
|---|---|---|
| official, unchanged | 103 / 104 | 103 / 104 |
| ours | 0 / 104 | 9 / 104 |
Same probe on the Q8_0 version of our encoder (temperature 0, 120 tokens, same 104 prompts): 0 / 104 strict refusals, 7 / 104 broad markers.
The optimisation's own count on its 100 evaluation prompts: 100/100 refusals before, 8/100 after, KL divergence 0.047.
Effect on images (Q8_0 image model, eight neutral prompts, seed 42; reference = the same Q8_0 with the official int8 text encoder):
| Text encoder | PSNR vs reference | SSIM vs reference |
|---|---|---|
| official encoder as GGUF Q4_K_M | 20.75 dB | 0.830 |
| ours Q4_K_M | 18.70 dB | 0.771 |
| ours Q8_0 | 25.59 dB | 0.908 |
The 4bit ablated encoder moves the images further away from the reference than plain 4bit quantization does (compositions and details change), but the images follow the prompts (we looked at the cat, street and poster images; text such as "LISBOA" is still rendered correctly). These numbers do not say anything about content restrictions.
Sakura-Image-2.1-Turbo-Q8_0-7.07GiB.gguf: made with sd-cli -M convert directly from the official BF16 weights of Qwen-Image-2.1-Turbo (sha256 of both shards checked); its tensors are identical to the Q8_0 file below (compared tensor by tensor), apart from the norm weights mentioned further down. Modified file.Sakura-Image-2.1-Turbo-Q5_K_M-*.gguf, Sakura-Image-2.1-Turbo-Q4_K_M-*.gguf: made with sd-cli -M convert from the Q8_0 GGUF of DogukanUrker/Qwen-Image-2.1-Turbo-GGUF (sha256 228d7931adade6cd..., itself a conversion of the official BF16 weights), not directly from the BF16 weights. They are plain uniform Q5_K / Q4_K files like other sd.cpp conversions; the new part of this repository is the text encoder. Modified files.Sakura-TextEncoder-*.gguf: the official text encoder of Qwen-Image-2.1-Turbo with our edit of part of its weights, converted to GGUF with llama.cpp and quantized to Q8_0 (8bit file) or Q4_K_M (4bit file). Modified files.attn.norm_q, attn.norm_k) of the image-model files are stored as F32 instead of Q8_0 (the same values after dequantization, +22 KB) so that ComfyUI can read them. A check with stable-diffusion.cpp (Q5_K_M, 4bit encoder, one prompt, seed 42, 512x512) gave a pixel-identical image before and after.qwen_image_2.1_vae_bf16.safetensors is an unmodified copy of the official VAE. Tokenizer and scheduler are not part of this repository.Qwen RESEARCH LICENSE AGREEMENT (LICENSE): use, reproduction and modification for non-commercial purposes only (research or evaluation); redistribution requires this license text and the notices in NOTICE. The agreement is governed by the laws of China. "Qwen" is the name of the original work; this is a derivative work by Sakura, "based on Qwen-Image-2.1-Turbo".
Part of the Sakura Mini line.
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。
Built with Qwen. 这是 Qwen/Qwen-Image-2.1-Turbo(8 步蒸馏,7B DiT)的社区量化版本,并附带一个 经消融(abliterated)的文本编码器。属于独立工作,并非 Alibaba/Qwen 制作或认可。仅限非商业用途(Qwen RESEARCH LICENSE AGREEMENT,副本见 LICENSE)。
仅限成年人使用;您需对自己生成的内容以及适用于您的法律负责。
Q8_0 图像模型与我们的 8bit 去审查文本编码器,stable-diffusion.cpp(Vulkan,Radeon 8060S),1024x1024、8 步、CFG 1.0、euler、官方 Turbo sigmas。每条提示词只用一个种子,没有重抽;提示词本身是从更多候选中挑选的。每张图约 77 秒(含模型加载);生成期间我们机器上的 GPU 显存占用峰值为 18.3 GiB(Windows 计数器,整块 GPU,运行前已占用 1.5 GiB)。这些图片由 stable-diffusion.cpp 生成,而不是由 ComfyUI 生成。
提示词是中性的。这些图片并不说明被消融的文本编码器对受限内容有何改变,只说明普通提示词仍然可用。同样的九条提示词和种子也用于姊妹仓库 Sakura-Qwen-Image-2.1-Turbo-GGUF(使用官方编码器),两组图片可以对照。
种子依次为 1202, 1302, 1305, 1306, 1105, 1103, 1503, 1506, 1406;提示词即上方图片的说明文字。
除运行程序外,所需文件都在本仓库中。请选择您所用程序对应的一列(🟦 stable-diffusion.cpp,🟧 ComfyUI):
| 部分 | 🟦 stable-diffusion.cpp | 🟧 ComfyUI |
|---|---|---|
| 图像模型(三选一) | Sakura-Image-2.1-Turbo-Q8_0-7.07GiB.gguf(最佳)、Sakura-Image-2.1-Turbo-Q5_K_M-4.60GiB.gguf(质量更好)或 Sakura-Image-2.1-Turbo-Q4_K_M-3.77GiB.gguf(最小),通过 --diffusion-model 传入 |
同一个文件,用 Unet Loader (GGUF) 加载 |
| 去审查文本编码器(二选一) | Sakura-TextEncoder-Qwen3VL-8B-Uncensored-8bit-8.11GiB.gguf(GGUF Q8_0,推荐:在我们的图像测试中最接近官方编码器)或 Sakura-TextEncoder-Qwen3VL-8B-Uncensored-4bit-4.68GiB.gguf(GGUF Q4_K_M,更小),通过 --llm 传入 |
Sakura-TextEncoder-Qwen3VL-8B-Uncensored-bf16.safetensors(BF16、16.3 GiB),用常规的 Load CLIP 节点加载,类型为 qwen_image |
| VAE | qwen_image_2.1_vae_bf16.safetensors(0.63 GiB,官方文件的未改动副本,sha256 相同),通过 --vae 传入 |
同一个文件,用 Load VAE 加载 |
| 程序 | stable-diffusion.cpp master 版本的 sd-cli / sd-server(我们使用的是 commit 228c707,Vulkan 构建),不包含 |
原生支持 Qwen-Image 2.1 的较新版 ComfyUI,以及支持 Qwen-Image 2.1 的 ComfyUI-GGUF 加载器(自定义节点),不包含 |
文本编码器的文件名中特意不带量化标签,这样 Hub 的硬件兼容性卡片只列出图像模型。
.gguf)放入 ComfyUI/models/diffusion_models/(或 unet/),VAE 放入 ComfyUI/models/vae/,文本编码器(.safetensors 文件)放入 ComfyUI/models/text_encoders/。qwen_image 的 Load CLIP(文本编码器)、常规的 Load VAE 以及 TextEncodeQwenImage21 节点搭建工作流。采样:8 步,CFG 1.0,euler,Turbo sigmas 为 1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568, 0.0(节点 "ManualSigmas" 接入 "SamplerCustomAdvanced")。Qwen-Image-2.1-Turbo_ComfyUI_workflow.json(拖入 ComfyUI 即可;文本编码器节点为 Load CLIP(Sakura-TextEncoder-Qwen3VL-8B-Uncensored-bf16.safetensors),图像模型为 Q8_0、512x512、种子 42,即我们测试过的设置)。已测试:BF16 safetensors 编码器 Sakura-TextEncoder-Qwen3VL-8B-Uncensored-bf16.safetensors 配常规的 Load CLIP 节点和 Q8_0 图像模型,在 512x512 下生成了正确的图像(一条提示词),同样文件格式的官方编码器也一样;两张图片非常相似(PSNR 42.8 dB,一条提示词)。Q5_K_M 和 Q4_K_M 图像模型在 ComfyUI 中也能以 512x512 运行(使用的是此前的 GGUF 文本编码器配置,现已不再提供)。
在我们于 Radeon 8060S 上的 ComfyUI 测试中,1024x1024 的采样极其缓慢(每一步超过五分钟,未做深入排查);512x512 只需数秒。同一台机器上 stable-diffusion.cpp 生成一张 1024x1024 的图约需 95 秒,因此示例图片由它生成。
使用 8bit 编码器时,Q5_K_M 在内存中的权重约为:4.6 + 8.1 + 0.6 = 13.3 GiB(使用 4bit 编码器时:4.6 + 4.7 + 0.6 = 10 GiB)(我们没有测量这一具体组合的显存峰值)。
| 部分 | 我们做了什么 |
|---|---|
| 文本编码器(Qwen3-VL-8B) | 用消融方法编辑(设置由我们选择),再转为 BF16 safetensors(ComfyUI)、GGUF Q8_0 和 Q4_K_M。作为聊天模型测得:在优化所用的 100 条有害评估提示上,拒绝次数 100/100 -> 8/100,首个 token 分布的 KL 散度为 0.047(BF16 检查点,量化之前)。 |
| 图像模型(7B DiT) | 仅对官方 Turbo 权重做了 量化(Q4_K、Q5_K、Q8_0)。我们没有改变 DiT 能画出什么。 |
我们 没有 测量的内容:被消融的编码器对 图像 改变了多少。拒绝次数描述的是把编码器当作语言模型使用时的表现;在本流程中它只提供嵌入向量。我们只比较了中性提示(见下文)。本仓库不包含任何受限内容的示例图片。
| 文件 | 大小 | 说明 |
|---|---|---|
Sakura-Image-2.1-Turbo-Q8_0-7.07GiB.gguf |
7.07 GiB | 图像模型(DiT),Q8_0 |
Sakura-Image-2.1-Turbo-Q5_K_M-4.60GiB.gguf |
4.60 GiB | 图像模型(DiT),sd.cpp 统一的 Q5_K 方案(文件名使用标准标签 Q5_K_M,以便 Hub 识别) |
Sakura-Image-2.1-Turbo-Q4_K_M-3.77GiB.gguf |
3.77 GiB | 图像模型(DiT),sd.cpp 统一的 Q4_K 方案(文件名使用标准标签 Q4_K_M) |
Sakura-TextEncoder-Qwen3VL-8B-Uncensored-8bit-8.11GiB.gguf |
8.11 GiB | 我们消融后的文本编码器,Q8_0(推荐) |
Sakura-TextEncoder-Qwen3VL-8B-Uncensored-4bit-4.68GiB.gguf |
4.68 GiB | 我们消融后的文本编码器,Q4_K_M(更小) |
Sakura-TextEncoder-Qwen3VL-8B-Uncensored-bf16.safetensors |
16.33 GiB | 我们消融后的文本编码器,BF16,ComfyUI 标准格式(单个文件,用常规的 Load CLIP 节点加载) |
本仓库中的 VAE qwen_image_2.1_vae_bf16.safetensors 是来自 Comfy-Org/Qwen-Image-2.1 的未改动官方文件。
sd-cli \
--diffusion-model Sakura-Image-2.1-Turbo-Q5_K_M-4.60GiB.gguf \
--vae qwen_image_2.1_vae_bf16.safetensors \
--llm Sakura-TextEncoder-Qwen3VL-8B-Uncensored-8bit-8.11GiB.gguf \
--sigmas "1.0,0.978453,0.95418,0.926626,0.89508,0.845148,0.704534,0.414568,0.0" \
--steps 8 --cfg-scale 1.0 --sampling-method euler --diffusion-fa --vae-tiling \
-W 1024 -H 1024 -p "your prompt" -o out.png
我们的所有运行和测量都在 512x512 下进行(8 步,在 Radeon 8060S 上使用 Vulkan,每张图约 12 到 15 秒);1024x1024 是该模型的常用尺寸,但我们没有测试。sigma 调度是官方 Turbo 调度(模型卡中的 sample_sigmas)。请使用 CFG 1.0 和 8 步。stable-diffusion.cpp master 版本的 sd-cli/sd-server(我们使用的是 commit 228c707)可直接加载这些文件。
相同提示、相同种子(42)、512x512、8 步、同一台机器;参照是制作这两个文件所用的 Q8_0 文件(DogukanUrker 的版本)。八条中性提示(猫、霓虹招牌街道、肖像、山间湖泊、带文字的海报、厨房、大理石雕像、海滩)。数值越高越接近 Q8_0;即使权重只有微小变化,扩散模型也会改变细节,因此这些是 偏离度 数值,而不是质量评分。
| 文件 | 大小 | 相对 Q8_0 的 PSNR | 相对 Q8_0 的 SSIM |
|---|---|---|---|
| Q5_K | 4.60 GiB | 24.77 dB | 0.906 |
| Q4_K | 3.77 GiB | 21.46 dB | 0.842 |
我们对把编码器当作聊天模型使用时的拒绝情况进行了自己的测量(llama.cpp,Q4_K_M,温度 0,并生成 120 个 token,mlabonne/harmful_behaviors 测试集中的 104 条留出有害提示,两行使用相同的提示和设置):
| 文本编码器(Q4_K_M) | 严格拒绝 | 宽泛拒绝标记 |
|---|---|---|
| 官方,未改动 | 103 / 104 | 103 / 104 |
| 我们的 | 0 / 104 | 9 / 104 |
对我们编码器的 Q8_0 版本做同样的测试(温度 0,生成 120 个 token,同样的 104 条提示):严格拒绝 0 / 104,宽泛标记 7 / 104。
优化过程自己在其 100 条评估提示上的计数:编辑前 100/100 次拒绝,编辑后 8/100,KL 散度 0.047。
对图像的影响(Q8_0 图像模型,八条中性提示,种子 42;参照 = 使用官方 int8 文本编码器的同一个 Q8_0):
| 文本编码器 | 相对参照的 PSNR | 相对参照的 SSIM |
|---|---|---|
| 官方编码器,GGUF Q4_K_M | 20.75 dB | 0.830 |
| 我们的 Q4_K_M | 18.70 dB | 0.771 |
| 我们的 Q8_0 | 25.59 dB | 0.908 |
4bit 的被消融编码器使图像比单纯 4bit 量化更偏离参照(构图和细节会变化),但图像仍然遵循提示(我们查看了猫、街道和海报图像;“LISBOA”这样的文字仍能正确渲染)。这些数值并不说明任何关于内容限制的情况。
Sakura-Image-2.1-Turbo-Q8_0-7.07GiB.gguf:使用 sd-cli -M convert 直接由 Qwen-Image-2.1-Turbo 的官方 BF16 权重制作(两个分片的 sha256 均已核对);其张量与下面提到的 Q8_0 文件完全相同(逐张量比较),下文提到的归一化权重除外。已修改文件。Sakura-Image-2.1-Turbo-Q5_K_M-*.gguf、Sakura-Image-2.1-Turbo-Q4_K_M-*.gguf:使用 sd-cli -M convert,由 DogukanUrker/Qwen-Image-2.1-Turbo-GGUF 的 Q8_0 GGUF(sha256 228d7931adade6cd...,其本身是官方 BF16 权重的转换)制作,而不是直接由 BF16 权重制作。它们与其他 sd.cpp 转换版本一样是普通的统一 Q5_K / Q4_K 文件;本仓库的新内容是文本编码器。已修改文件。Sakura-TextEncoder-*.gguf:Qwen-Image-2.1-Turbo 的官方文本编码器,加上我们对部分权重的编辑,用 llama.cpp 转换为 GGUF 并量化为 Q8_0(8bit 文件)或 Q4_K_M(4bit 文件)。已修改文件。attn.norm_q、attn.norm_k)由 Q8_0 改为以 F32 存储(反量化后数值相同,增加 22 KB),以便 ComfyUI 读取。用 stable-diffusion.cpp 检查(Q5_K_M,4bit 编码器,一条提示,种子 42;512x512),修改前后生成的图像逐像素完全相同。qwen_image_2.1_vae_bf16.safetensors 是官方 VAE 的未改动副本。分词器和调度器不属于本仓库。Qwen RESEARCH LICENSE AGREEMENT(LICENSE):使用、复制和修改 仅限非商业目的(研究或评估);再分发需附带此许可文本以及 NOTICE 中的声明。该协议适用中国法律。“Qwen”是原作品的名称;本作品是 Sakura 的衍生作品,“基于 Qwen-Image-2.1-Turbo”。
属于 Sakura Mini 系列。
Sakura-TextEncoder-Qwen3VL-8B-Uncensored-bf16.safetensors 18 GBqwen_image_2.1_vae_bf16.safetensors 676 MB有文件?提交种子。
提交种子Qwen/Qwen-Image-2.1-Turbo