Sakura — FLUX.3 Image Klein for ComfyUI (GGUF)

Built with FLUX. The files of Sakura-FLUX.3-Image-Klein-GGUF in the ComfyUI folder layout, with two ready-to-use ComfyUI workflows (text-to-image and image editing). Download the folders into ComfyUI/models/ and drag a workflow into ComfyUI. Independent work, not made, reviewed or endorsed by Black Forest Labs. The image model files, the VAE and the workflows' model content are under the FLUX Kommunity License (copy in LICENSE.md: non-commercial use, or commercial use of outputs only as a Qualifying User under US$5 million revenue with the filtering or review the license requires); the unmodified text encoder is Apache-2.0 (LICENSE-Apache-2.0.txt); the workflow graphs are based on a ComfyUI template under the MIT license (LICENSE-MIT-ComfyUI-workflow-templates.txt). See NOTICE.

The model files are the same files (byte-identical) as in Sakura-FLUX.3-Image-Klein-GGUF; quality numbers, how they were made and the stable-diffusion.cpp instructions are there. An uncensored version of the text encoder (only the encoder) is in Sakura-FLUX.3-Image-Klein-TextEncoder-Uncensored-GGUF.

Text-to-image workflow	Reference image	Image-edit workflow

Both images were made with the two workflows of this repository exactly as they are (Q8_0 image model, seed 42, 16 steps, CFG 1.0, euler/simple, 1024x1024), loaded in the ComfyUI web interface and run with its queue button. The text-to-image result is pixel-identical to the fox image published in Sakura-FLUX.3-Image-Klein-GGUF (same prompt and seed). The image edit uses the fox picture as reference (workflows/sakura_klein_reference_fox.png) and the edit prompt shown in the gallery caption. One seed per workflow, no re-rolls.

Install

Install ComfyUI-GGUF (custom node, needed for Unet Loader (GGUF); restart ComfyUI afterwards).
Download the folders of this repository into ComfyUI/models/:

Folder in this repository → `ComfyUI/models/...`	File	Size
`diffusion_models/` (pick one)	`Sakura-FLUX.3-Image-Klein-Q8_0-5.58GiB.gguf` (best, used in the workflows)	5.58 GiB
	`Sakura-FLUX.3-Image-Klein-Q6_K-4.35GiB.gguf`	4.35 GiB
	`Sakura-FLUX.3-Image-Klein-Q5_K_M-3.68GiB.gguf`	3.68 GiB
	`Sakura-FLUX.3-Image-Klein-Q4_K_M-3.05GiB.gguf`	3.05 GiB
`text_encoders/`	`Qwen3-VL-4B-Instruct-bf16-8.27GiB.safetensors` (official Qwen3-VL-4B-Instruct, unmodified, the two shards merged into one file for the standard Load CLIP node, type `flux3`)	8.27 GiB
`vae/`	`flux2-vae.safetensors` (unmodified copy of `ae.safetensors` of the official repository)	0.31 GiB

With the huggingface_hub CLI, run inside your ComfyUI/models folder (only the Q8_0 image model; change the pattern for another quant):

hf download webmp3/Sakura-FLUX.3-Image-Klein-Comfy-GGUF --include "diffusion_models/*Q8_0*" "text_encoders/*" "vae/*" --local-dir .

Drag a workflow from workflows/ into ComfyUI (or use the "Open" menu). For the image-edit workflow put workflows/sakura_klein_reference_fox.png into ComfyUI/input/, or pick your own picture in the Load Image node.

Optional: uncensored text encoder

The text encoder in this repository is the official one. Our uncensored version of the text encoder (only the encoder is changed, the image model is the same) is a separate repository: Sakura-FLUX.3-Image-Klein-TextEncoder-Uncensored-GGUF. For ComfyUI take its file Sakura-TextEncoder-Qwen3VL-4B-Uncensored-bf16-8.27GiB.safetensors (8.27 GiB), put it into ComfyUI/models/text_encoders/ and select it in the Load CLIP node of either workflow (type stays flux3):

hf download webmp3/Sakura-FLUX.3-Image-Klein-TextEncoder-Uncensored-GGUF Sakura-TextEncoder-Qwen3VL-4B-Uncensored-bf16-8.27GiB.safetensors --local-dir text_encoders

That file was tested in ComfyUI for text-to-image with the equivalent API-format graph (its repository shows the images and what we measured); we did not run it with the two workflows of this repository. It is for adults; see its repository for what was changed and what it still refuses.

Workflows

File	What it does	Tested
`workflows/Sakura-FLUX.3-Image-Klein-GGUF_text-to-image.json`	text-to-image: Unet Loader (GGUF), Load CLIP (`flux3`), Load VAE, prompt, zeroed negative, `Empty Flux 2 Latent Image` 1024x1024, KSampler 16 steps / CFG 1.0 / `euler` / `simple`, seed 42 (fixed)	yes, Q8_0, one image (above)
`workflows/Sakura-FLUX.3-Image-Klein-GGUF_image-edit.json`	image editing with a reference image: the same loaders and sampler, `Reference Multi Latent` (reference scaled to about 1 megapixel; more reference images can be plugged in and addressed as `<ref_image_0>`, `<ref_image_1>`, ... in the prompt), an image-compare view of input and result	yes, Q8_0, one image (above)
`workflows/sakura_klein_reference_fox.png`	the reference picture of the image-edit workflow (our own text-to-image output, see above)	-

Both graphs are derived from the ComfyUI template image_flux3_klein_distilled_image_edit (package comfyui-workflow-templates-json 0.1.111, Comfy Org, MIT): we swapped the standard UNET loader for the GGUF loader, set our file names, a fixed seed and our prompts, removed the template's bounding-box helper nodes and notes, and for text-to-image removed the reference-image part. Each workflow carries download links of its model files, so ComfyUI can offer to fetch missing files.

Tested with ComfyUI 0.41.1 (ComfyUI Desktop environment, frontend 1.55.17, PyTorch 2.12.0+rocm7.14.0) and the UnetLoaderGGUF node of the installed ComfyUI-GGUF custom node, Radeon 8060S. Not measured / not tested: the smaller quants with these exact workflows (they were tested for text-to-image with the equivalent graph in Sakura-FLUX.3-Image-Klein-GGUF), image editing quality beyond the one example, multiple reference images, the base (non-distilled) model, other GPUs; no timings are given here because the GPU was not idle during these runs.

License and use

Mixed: FLUX Kommunity License (LICENSE.md) for the image model files and the VAE; Apache-2.0 (LICENSE-Apache-2.0.txt) for the text-encoder file; MIT (LICENSE-MIT-ComfyUI-workflow-templates.txt) for the workflow graphs derived from the ComfyUI template. Any rights to use the FLUX model and derivatives are granted directly by Black Forest Labs Inc. to the recipient under that license. See NOTICE. Part of the Sakura Media line.

Credits

Image model: black-forest-labs/flux-3-image-klein (Black Forest Labs).
Text encoder: Qwen/Qwen3-VL-4B-Instruct (Qwen team, Alibaba).
Workflow template: ComfyUI workflow templates (Comfy Org, MIT); loading: city96/ComfyUI-GGUF; conversion and quantization: leejet/stable-diffusion.cpp (MIT).

中文说明 · 樱花 (Simplified Chinese)

English above。完整的独立中文版见 README_zh.md。

Built with FLUX. 本仓库把 Sakura-FLUX.3-Image-Klein-GGUF 的文件按 ComfyUI 文件夹结构 排好(diffusion_models/、text_encoders/、vae/),并附带两个可直接使用的 ComfyUI 工作流(文生图、图像编辑)。模型文件与主仓库逐字节相同;质量数据见主仓库。去审查文本编码器(可选)见 Sakura-FLUX.3-Image-Klein-TextEncoder-Uncensored-GGUF,在 Load CLIP 节点中选择其 …-Uncensored-bf16-8.27GiB.safetensors 即可。需要自定义节点 ComfyUI-GGUF。

两个工作流均按原样在 ComfyUI 网页界面中运行过(Q8_0,种子 42,16 步,CFG 1.0,euler/simple,1024x1024),各生成一张图(见上方)。文生图结果与主仓库中发布的狐狸图像素级一致。未测试:其他量化在这两个工作流中的效果、多参考图、基础模型、其他 GPU;GPU 测试时并不空闲,所以不给出耗时。

许可:图像模型文件和 VAE 为 FLUX Kommunity License(LICENSE.md),文本编码器为 Apache-2.0,工作流图基于 ComfyUI 模板(MIT)。详见 NOTICE。