Flux 3 Klein GGUF

FLUX 3 Image Klein is a 5.6B parameter open-weights diffusion transformer from Black Forest Labs for text-to-image generation and reference-based image editing. A single model accepts a text prompt and up to 10 reference images to produce a new image or edit existing inputs. It also supports bounding-box coordinates in the prompt for spatial layout control and text rendering. This repository provides GGUF quantizations of both the 16-step Distilled (sd) and 50-step Base (base) checkpoints for ComfyUI.

Hugging Face repositoryFlux-3-Klein-GGUF on Hugging Face

Example

Prompting tips for this model:

• Write full scene descriptions instead of short tag lists, or run the Qwen3.5 9B prompt upsampler to expand short instructions automatically.
• For spatial layout control or regional edits, include [top, left, bottom, right] coordinates on a 0 to 1000 grid directly in the prompt, for example place a brass lantern at [620, 180, 860, 340].

Available Quantizations

Q8_CR is recommended on GPUs with native INT8 acceleration. For lower VRAM usage under 3 GB, Q4_CR_L uses packed INT4 ConvRot weights with INT8 activations and keeps additional sensitive attention projections in Q8_CR, while Q4_CR uses INT4 weights and INT4 activations. One-dimensional normalization tensors and final_layer.image.linear stay in FP32, and the 128-wide input patch projections (emb_in) remain in source BF16.

Distilled (flux3_image_klein_sd)

Quant	Size	File
Q4_CR	2.74 GB	flux3_image_klein_sd-Q4_CR.gguf
Q4_CR_L	2.78 GB	flux3_image_klein_sd-Q4_CR_L.gguf
Q4_K_M	2.91 GB	flux3_image_klein_sd-Q4_K_M.gguf
Q6_K	4.24 GB	flux3_image_klein_sd-Q6_K.gguf
Q8_CR	5.17 GB	flux3_image_klein_sd-Q8_CR.gguf
Q8_0	5.49 GB	flux3_image_klein_sd-Q8_0.gguf

Base (flux3_image_klein_base)

Quant	Size	File
Q4_K_M	2.90 GB	flux3_image_klein_base-Q4_K_M.gguf
Q6_K	4.23 GB	flux3_image_klein_base-Q6_K.gguf
Q8_CR	5.16 GB	flux3_image_klein_base-Q8_CR.gguf
Q8_0	5.48 GB	flux3_image_klein_base-Q8_0.gguf

Model Input and Output

Variants and inputs

Variant	Inputs
Distilled (flux3_image_klein_sd)	Text prompt and up to 10 optional reference images for fast generation and image editing (4 to 16 steps, CFG 1.0)
Base (flux3_image_klein_base)	Text prompt and up to 10 optional reference images for undistilled generation and LoRA training (50 steps, CFG 3.5)

Outputs

Up to 4k resolution image.

ComfyUI Setup

Load the model with the UnetLoaderGGUF node from comfyui-gguf-reboot. The loader node gives the best performance. Set CLIPLoader to the flux3 type when loading the Qwen3-VL 4B text encoder.

Dependencies

Place the GGUF file and these companion files from the upstream repository in the ComfyUI model folders before running a workflow.

Component	File	Folder
Diffusion transformer	flux3_image_klein_sd-<quant>.gguf or flux3_image_klein_base-<quant>.gguf	models/diffusion_models/ or models/unet/
Text encoder	qwen3vl_4b_int8_convrot.safetensors or qwen3vl_4b_bf16.safetensors	models/text_encoders/ or models/clip/
Prompt upsampler (optional)	qwen3.5_9b_flux3_prompt_upsampler_int8_convrot.safetensors or qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors	models/text_encoders/ or models/clip/
VAE	flux2-vae.safetensors	models/vae/

Workflow downloads

Import these JSON files into ComfyUI.

JSON fileText-to-Image workflowJSON fileImage Edit workflow