AI image workflows often require different models for text-to-image generation, reference-based editing, and multi-image composition. Switching between models can make the workflow more complicated, especially when you need precise control over object placement, consistent subjects, or detailed edits. FLUX 3 Image Klein addresses these problems by bringing image generation and editing into one model, with support for multiple reference images, spatial guidance, and prompt enhancement.
 |
| flux3 image klein showcase |
Its a 5.6B parameter open-weights model designed for text-to-image generation and image editing. It accepts a text prompt and up to 10 reference images, then produces a single output image.
It also supports bounding-box instructions for more controlled image editing and can render text within generated images. With support for images up to 4K resolution, it targets workflows that need both creative flexibility and detailed output.
Offcially it is recommended to use Klein model with its prompt upsampler. The upsampler expands a short prompt into a more detailed scene description, helping specify framing, subjects, backgrounds, and lighting. For example, instead of manually describing every visual detail, you can start with a simple idea and let the upsampler turn it into a fuller prompt. This provides additional guidance to the image model and can help it better interpret your intended composition. The model uses Qwen3-VL-4B-Instruct as its text encoder.
FLUX 3 Image Klein is released under the FLUX Kommunity (Base License and Distilled License) License. It permits non-commercial use and includes a commercial-output exception for qualifying users, subject to the license terms. Anyone planning commercial use should review the license carefully to confirm that their intended use qualifies.
Installation
1. First you should have ComfyUI installed. Update it to its latest version from the Manager tab if already have.
2. Download these diffusion models provided below. Choose any of them as per your system VRAM resources. Save any of them inside ComfyUI/models/diffusion_models folder.
-flux3_image_klein_base_bf16.safetensors (full precision)
-flux3_image_klein_base_int8_convrot.safetensors (for 8 GB GPUs )
-flux3_image_klein_base_nvfp4.safetensors (for Blackwell GPU)
-flux3_image_klein_sd_bf16.safetensors (Full precision)
-flux3_image_klein_sd_int8_convrot.safetensors (for 8 GB GPUs )
-flux3_image_klein_sd_nvfp4.safetensors (for Blackwell GPUs)
Note- Model files with "_sd_" in its model name are the distilled variants and rest are base varaints. Models without FP8 support runs on RTX 30 series and older.
GGUF-(Alternative ) You can also use GGUF quantized variants instead of diffusion models if having low VRAMS released by community.
(a) Flux 3 Klein GGUF by Abiray
(b) Flux 3 Klein GGUF by Unsloth
(c) Flux 3 Klein GGUF by RebelAI
Download any of them and save these into ComfyUI/models/unet folder. Make sure you are using ComfyUI-GGUF custom node by City96.
3. Again, download any of the text encoder models. Choose the one thats suits your VRAM requirements. Its necessary to use prompt upsampler to get the best results. Save this into ComfyUI/models/text_encoders folder.
-qwen3.5_9b_flux3_prompt_upsampler_int8_convrot.safetensors
-qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors
-qwen3vl_4b_bf16.safetensors
-qwen3vl_4b_int8_convrot.safetensors
4. Download vae model (flux2-vae.safetensors) and save this into ComfyUI/models/vae folder.
5. Restart and refresh ComfyUI.
Workflow
1. Download Flux 3 Klein Base & Distilled workflows from our Hugging Face repository.
(a) Flux3-Klein-t2i.json
(b) Flux3-Klein-Image-Edit.json
(c) Flux3-Klein-Distilled-t2i.json
(d) Flux3-Klein-Distilled-Image-Edit.json
(e) Flux3-Klein-Distilled-Image-Edit-GGUF.json
(f) Flux3-Klein-Distilled-t2i-GGUF.json
2. Drag and drop into comfyUI.
3. Load the models into their relevant nodes.
4. Set the KSampler configuration settings:
Klein Base (Used For finetuning, lora training)
Steps-50
CFG-3.5
Klein Distilled (For Image generation and editing, self-hosting , with BFL API)
Steps-16
CFG-1
5. Add prompts into prompt box. Use prompt upsampler to rewrite the prompting. For more detailed overview you can refer to their official prompting format.
6. Hit Run to start generation.
){kind=link}