Flux.3 Image Klein in ComfyUI: Text-to-Image and Edit

​

About FLUX.3 [klein]

• Text-to-image and image editing from the same weights; edits take an instruction prompt plus reference images
• Two checkpoints: full Flux.3 Image Klein, and Distilled for guidance-free sampling in 4–16 steps
• Multi-reference prompting via <ref_image_0>, <ref_image_1>, and so on
• Optional Qwen3.5 9B prompt upsampler that expands short requests into full captions before generation
• Single-file checkpoints in bf16, int8 (convrot), and nvfp4, with a Qwen3-VL-4B text encoder and the Flux 2 VAE

​

Flux.3 Image Klein Workflows

​

Flux.3 Image Klein: Text to Image

1. Enter your prompt in the subgraph’s positive prompt widget
2. Set width and height, or pick a frame from ResolutionSelector (1:1 square by default)
3. Adjust steps (25 by default) and cfg (3.5 by default). The template carries a note to try 50 steps for better quality
4. Leave enable_prompt_upsampler off to generate from your own words, or turn it on to expand the prompt with the Qwen3.5 9B upsampler first
5. Click Queue. The image is saved by the SaveImage node to ComfyUI/output/

​

Flux.3 Image Klein Distilled: Text to Image

1. Enter your prompt. The template upsamples it first, so a short description is enough
2. Set steps between 4 and 16 (15 by default). The distilled checkpoint runs with cfg 1, so raising cfg does not help
3. Pick a frame from ResolutionSelector
4. Click Queue to generate

​

Flux.3 Image Klein: Image Edit

1. Load your source image in the LoadImage node
2. Connect it to Set Reference Multi Latent. The first reference is <ref_image_0> in the prompt, the next is <ref_image_1>, and so on
3. Write the edit as a full description of the result. In the shipped example the positive prompt is change the dress color to red
4. Adjust the canvas with Empty Flux 2 Latent (1024 x 1024 in the template) if you want a different output size
5. Click Queue. ImageCompare shows the input and the result side by side

​

Flux.3 Image Klein Distilled: Image Edit

1. Load the image you want to edit and connect it to Set Reference Multi Latent
2. Write the edit prompt. The shipped example replaces an earring and describes the new one in detail, including what must not appear
3. Keep steps in the 4 to 16 range and leave cfg at 1
4. Click Queue and compare the input with the result in ImageCompare

​

Flux.3 Image Klein Model Downloads

📂 ComfyUI/
├── 📂 models/
│   ├── 📂 diffusion_models/
│   │      ├── flux3_image_klein_int8_convrot.safetensors
│   │      └── flux3_image_klein_sd_int8_convrot.safetensors
│   ├── 📂 text_encoders/
│   │      ├── qwen3vl_4b_int8_convrot.safetensors
│   │      └── qwen3.5_9b_flux3_prompt_upsampler_int8_convrot.safetensors
│   └── 📂 vae/
│          └── flux2-vae.safetensors

​

Prompting and reference tips

• Write the result, not the change. The image editing templates expect a full caption of the finished image with the edit folded in, which is why the examples describe the new object in detail instead of saying “change this”.
• References are added to both conditionings. Connect extra images to Set Reference Multi Latent in order, address them as <ref_image_0>, <ref_image_1>, and so on, and add each one to both the positive and the negative prompt, the way the reference pipeline does.
• Reference size is a cost dial. The templates resize each reference to roughly 1 megapixel and the output latent takes that size. Keeping references at their native size, which Black Forest Labs’ own pipeline does up to about 16 megapixels, keeps more detail in the edit at a higher cost.
• Step count depends on the checkpoint. The full checkpoint runs at 25 steps and guidance 3.5 in the templates, with a note to try 50 steps for better quality. The step-distilled checkpoint runs at 15 or 16 steps with guidance fixed at 1, and it is the one to use when you want a fast pass.
• The templates ship a Create Bounding Boxes node. It draws boxes over the source image for planning regional edits, and its list of boxes is shown in the Preview Any node so you can fold the regions into the prompt.