FLUX 3 Image Klein in ComfyUI

Use the local Klein workflow in ComfyUI

The native workflow loads downloaded weights on your hardware. The separate FLUX 3 Image Partner Node calls a hosted API and has different credentials and billing. Choose the local Klein template when your goal is to use the open weights; seeing the FLUX name in a node is not enough to identify which engine is running.

Begin with the official ComfyUI launch instructions and the BFL ComfyUI model card. Update your supported ComfyUI installation, reopen it, and load the native template before downloading files. That lets the template identify the components it expects.

Keep a working environment recoverable

Save existing workflows before updating. If you are trying a new precision format, give the new workflow its own filename. Change one setting at a time so you can tell whether a problem comes from the model, a loader or an optional branch.

Put each component in its model folder

The current publisher package uses the following layout. This example shows the distilled INT8 route; it is a file-location map, not a universal hardware recommendation. Confirm the exact filenames against the native template and repository revision you use.

ComfyUI/models/
  diffusion_models/
    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

The _sd_ filename identifies the distilled transformer in this package. Use matching Base files only when the workflow calls for Base. The VAE's filename contains “flux2”; do not replace a template's linked VAE solely because that name looks older.

Check the loaders and prompt upsampler

The native path uses Load Diffusion Model, Load CLIP with the flux3 type, and Load VAE. Keep the template's prompt-upsampling branch enabled for ordinary short requests. Its purpose is to expand the request into the detailed scene caption the model expects. A missing upsampler can affect output even when sampling succeeds.

If you still need files or gated access, use the download and access checklist before debugging nodes.

Run a small baseline before adding references

Step 1: keep the template defaults

Select the downloaded files in their respective loaders. Preserve the native sampling path; settings copied from a different FLUX generation can be incompatible. Save this configured workflow before running it.

Step 2: generate one simple image

Use one subject, an uncomplicated background and the template's smaller output setting. Keep reference inputs empty for the first test. Save both the result and workflow, then note the checkpoint and seed so later changes have a baseline.

Step 3: introduce one reference

Add a photograph you have permission to use and request one visible change. Inspect the preserved regions as well as the target. Add more references only after the single-image path works, assigning each image a clear role.

Step 4: test larger outputs or layout control

Increase resolution separately from reference count. For bounding boxes, follow the native template's coordinate conventions rather than pasting a box format from a different runtime. A successful small image does not establish that a 4K multi-reference job will fit.

Common workflow errors and fixes

Missing node or missing flux3 loader type

Confirm the running application was updated, not just a launcher or a second checkout. Reopen a fresh native template in that installation. Check the console's first relevant error before adding custom nodes.

Files do not appear in a selector

Compare the model-folder map with the application's actual model paths. Refresh or restart as required by your installation, then reselect the file. Keep a complete file in its component folder instead of placing every download in checkpoints.

Out-of-memory during sampling

Reduce resolution and reference count and close other GPU tasks. The hardware and memory discussion explains why file size is not a complete VRAM estimate.