license: apache-2.0
base_model: black-forest-labs/flux-3-image-klein-pup
base_model_relation: finetune
library_name: gguf
pipeline_tag: image-text-to-text
tags:
- sakura
- sakura-media
- gguf
- llama.cpp
- flux
- flux-3
- flux-3-image-klein
- prompt-upsampling
- qwen3_5
- uncensored
- abliterated
- heretic
- not-for-all-audiences
Sakura — FLUX.3 Image Klein prompt upsampler, uncensored (GGUF)
Only the prompt upsampler is edited. black-forest-labs/flux-3-image-klein-pup is a 9B vision-language model (fine-tuned from Qwen3.5-9B by Black Forest Labs) that expands a short image request into the detailed caption the FLUX.3 image model works best with. This repository has it in a version whose language weights we edited with an abliteration run so that it refuses less, as GGUF with importance-matrix quantizations. The image model and its text encoder are not part of this repository: see Sakura-FLUX.3-Image-Klein-GGUF (image model, official text encoder) and Sakura-FLUX.3-Image-Klein-TextEncoder-Uncensored-GGUF (uncensored text encoder).
For adults only; you are responsible for what you generate and for the laws that apply to you. Independent work, not made or endorsed by Black Forest Labs. Apache-2.0 (copy in LICENSE).
The upsampler is optional. FLUX.3 Klein takes any prompt directly; all example images in the FLUX repositories were made without an upsampler. We did not measure whether the upsampler improves images.
What you need to run it
Pick the column of the program you use (🟦 llama.cpp, 🟧 ComfyUI):
| Part | 🟦 llama.cpp | 🟧 ComfyUI |
|---|---|---|
| Model (pick one) | Sakura-FLUX.3-Image-Klein-PromptUpsampler-Uncensored-Q8_0-8.87GiB.gguf (best), …-Q6_K-6.85GiB.gguf, …-Q5_K_M-6.02GiB.gguf or …-Q4_K_M-5.24GiB.gguf (smallest), passed with -m |
the official FLUX.3 template loads the upsampler with CLIPLoader (type flux3) and TextGenerate; we have not tested our files there, and the BF16 safetensors file for that node is not in this repository yet |
| System prompt | system_prompts.json (t2i, i2i, t2i_bbox, i2i_bbox; unchanged copy of the original file) as the system message |
the same file |
Reference images (i2i) |
Sakura-FLUX.3-Image-Klein-PromptUpsampler-mmproj-F16-0.86GiB.gguf with --mmproj (made from the unedited original; we did not test image input) |
– |
| Program | llama-server of llama.cpp, we used release b11594 (Vulkan), not included |
not tested |
Memory: with Q8_0 about 8.9 GiB of weights (plus 0.9 GiB for the projector and the context); we did not measure peak memory.
Run
llama-server -m Sakura-FLUX.3-Image-Klein-PromptUpsampler-Uncensored-Q8_0-8.87GiB.gguf --jinja -ngl 99 -c 8192
Send the short request as the user message and the t2i text of system_prompts.json as the system message; the answer is the caption for the image model.
What is "uncensored" here, exactly
| Part | What we did |
|---|---|
| Language weights (9B) | Directional ablation with Heretic 1.4.0 on the original checkpoint (revision 8b069751): 60 trials, the tool's default prompt sets (mlabonne/harmless_alpaca, mlabonne/harmful_behaviors) and default settings, seed 42, without a system prompt; trial 23: by the tool's own count refusals on its 100 evaluation prompts 77 → 45, first-token KL divergence 0.1144. |
Projector (mmproj) |
Not changed, converted from the original checkpoint. |
| Image model | Not part of this repository. |
What we did not measure: caption quality and the effect of the edit on the captions of ordinary requests beyond reading some answers.
Files
| File | Size | What it is |
|---|---|---|
Sakura-FLUX.3-Image-Klein-PromptUpsampler-Uncensored-Q8_0-8.87GiB.gguf |
8.87 GiB | 8-bit |
Sakura-FLUX.3-Image-Klein-PromptUpsampler-Uncensored-Q6_K-6.85GiB.gguf |
6.85 GiB | 6-bit K-quant |
Sakura-FLUX.3-Image-Klein-PromptUpsampler-Uncensored-Q5_K_M-6.02GiB.gguf |
6.02 GiB | 5-bit K-quant with importance matrix |
Sakura-FLUX.3-Image-Klein-PromptUpsampler-Uncensored-Q4_K_M-5.24GiB.gguf |
5.24 GiB | 4-bit K-quant with importance matrix |
Sakura-FLUX.3-Image-Klein-PromptUpsampler-mmproj-F16-0.86GiB.gguf |
0.86 GiB | vision projector for reference images, from the original checkpoint, not edited |
system_prompts.json |
small | unchanged copy of the original system prompts |
A BF16 safetensors file (the merged result in one file, 17.5 GiB, for transformers or the ComfyUI TextGenerate node) exists on our side and will be added later.
What we measured
Refusals (104 prompts of the test split of mlabonne/harmful_behaviors, one user turn each with the original t2i system prompt, temperature 0, 900 tokens, llama-server b11594; the first 100 are the evaluation prompts Heretic optimized against, so they are not held out; a refusal is the strict English phrase regex, markers = Heretic's broader list):
| Model | strict refusals |
|---|---|
| unmodified (our Q8_0 of the original checkpoint) | 0 / 104 (markers 32 / 104, empty 0) |
| uncensored, Q8_0 | 0 / 104 (markers 3 / 104, empty 0) |
| uncensored, Q4_K_M | 0 / 104 (markers 5 / 104, empty 0) |
With the original system prompt the unmodified model already does not refuse by this regex, so this probe cannot show how much the edit reduces refusals; the broad markers fall from 32 to 3, and Heretic's own count without a system prompt falls from 77 to 45 of 100. Whether an answer that is not a refusal still writes a proper caption was not checked beyond reading some answers; we did not measure caption quality.
Quantization loss (KL divergence against the uncensored Q8_0, llama-perplexity b11594, 20 chunks of 2048 tokens of English and web text, no system prompt):
| File | KLD en | KLD web | same top token en / web |
|---|---|---|---|
| Q6_K | 0.0186 ± 0.0017 | 0.0044 ± 0.0001 | 95.8 / 95.9 % |
| Q5_K_M | 0.0170 ± 0.0013 | 0.0054 ± 0.0001 | 95.6 / 95.8 % |
| Q4_K_M | 0.0386 ± 0.0021 | 0.0157 ± 0.0002 | 92.6 / 92.9 % |
Not measured: caption quality, the i2i and bounding-box modes, image input, other languages, speed, the ComfyUI template with these files.
What we changed (Apache License 2.0, section 4)
Sakura-FLUX.3-Image-Klein-PromptUpsampler-Uncensored-*.gguf: the original checkpoint with the Heretic edit of the attention output projections and the MLP down projections merged into the weights, converted to BF16 GGUF withconvert_hf_to_gguf.py(llama.cpp commit9c2e0e491), an importance matrix computed on it (96 chunks of 512 tokens of ourcalib_mix_v1mix, English and German) and quantized withllama-quantizeof llama.cpp b11594. Modified files.…mmproj-F16-….gguf: converted from the original checkpoint, not edited.system_prompts.json: unchanged copy.
License and use
Apache-2.0 as the original model (LICENSE); the files are modified (edited and quantized) versions of the original weights, see NOTICE.
Part of the Sakura Media line.
Credits
- Model: black-forest-labs/flux-3-image-klein-pup (Black Forest Labs), fine-tuned from Qwen/Qwen3.5-9B.
- Abliteration tool: p-e-w/heretic (AGPL-3.0; used as a tool, nothing of it is distributed here).
- Conversion, quantization and measurement: ggml-org/llama.cpp (MIT).
中文说明 · 樱花 (Simplified Chinese)
English above。完整的独立中文版见 README_zh.md。
只有提示词扩写器被编辑。 black-forest-labs/flux-3-image-klein-pup(Black Forest Labs 基于 Qwen3.5-9B 微调的 9B 视觉语言模型)把简短的图像请求扩写成 FLUX.3 效果最好的详细描述。本仓库提供其一个版本:语言权重经我们用消融(abliteration)编辑,使其拒绝更少,格式为带重要性矩阵量化的 GGUF。图像模型和文本编码器 不在 本仓库中(见 Sakura-FLUX.3-Image-Klein-GGUF 与 Sakura-FLUX.3-Image-Klein-TextEncoder-Uncensored-GGUF)。仅限成年人使用。扩写器是可选的:FLUX.3 Klein 可以直接使用任何提示词,FLUX 仓库的示例图像都是不经扩写器生成的。
测量: 使用原始 t2i 系统提示词时,未改动的模型按严格正则本来就不拒绝(0 / 104),我们的版本同样是 0 / 104,因此该探测无法显示编辑减少了多少拒绝;宽泛标记从 32 降到 3,Heretic 自己在无系统提示词下的计数从 100 条中的 77 降到 45。没有测量描述质量、i2i 模式、图像输入、速度和 ComfyUI 模板。BF16 safetensors 文件(17.5 GiB)将在之后补充。