Analysis updated 2026-05-18
Run FLUX.2-klein-9B image generation with a smaller, faster text encoder to save VRAM.
Fix size mismatch errors when trying to load the pruned Klein text encoder in ComfyUI.
Try the example workflow to generate images with recommended Klein settings.
Compare image quality and memory use between the full and pruned text encoders.
| kgonia/comfyui-kleinprunedte | 0-bingwu-0/live-interpreter | 010zx00x1/faresnipe | |
|---|---|---|---|
| Stars | 2 | 2 | 2 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | moderate | easy |
| Complexity | 2/5 | 2/5 | 2/5 |
| Audience | vibe coder | general | general |
Figures from each repo's GitHub metadata at analysis time.
Requires downloading a gated model file after accepting a license on Hugging Face, plus a ComfyUI version with FLUX.2 support.
ComfyUI-KleinPrunedTE is a small add-on node for ComfyUI, a visual tool for building AI image generation workflows. It adds support for a trimmed-down version of the text encoder used by the FLUX.2-klein-9B image generation model. A text encoder is the part of an image generation system that turns a written prompt into a form the rest of the model can use to create an image. The pruned version of this text encoder has 5.1 billion parameters instead of the original 8.2 billion, which the README says makes it about 35 percent smaller and cuts peak video memory use by around 5 gigabytes during encoding, while producing images that are nearly indistinguishable from the full-size encoder. However, because this pruned version has an irregular internal structure that varies layer by layer, ComfyUI's normal text encoder loading node cannot read it correctly and fails with an error. This project exists specifically to load that pruned checkpoint correctly by figuring out the right shape for each layer directly from the file itself, without needing any extra configuration files. Installation can be done through the ComfyUI Manager by installing from a Git URL, or manually by cloning the repository into ComfyUI's custom_nodes folder and restarting ComfyUI. No additional Python packages are needed since the node only relies on ComfyUI's own built-in tools, though it does require a version of ComfyUI that already supports FLUX.2. Using it requires downloading three separate model files and placing them in specific ComfyUI folders: the pruned text encoder itself, the main FLUX.2-klein-9B image generation model (which is gated and requires accepting a license and logging in on Hugging Face to download), and a matching VAE file. The repository includes an example workflow file that already has the new node wired in correctly with the recommended settings for this particular model. The code for this ComfyUI node is released under the MIT license, but the pruned model weights themselves are a derivative of FLUX.2-klein-9B and carry that model's FLUX Non-Commercial License, meaning the weights cannot be used commercially.
A ComfyUI custom node that correctly loads a smaller, pruned text encoder for the FLUX.2-klein-9B image generation model.
Mainly Python. The stack also includes Python, ComfyUI.
The node's code is MIT and free to use, but the pruned model weights inherit the FLUX Non-Commercial License, so the weights cannot be used for commercial purposes.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly vibe coder.
This repo across BitVibe Labs
Verify against the repo before relying on details.