Load a photo into the official FLUX.3 Image Klein edit template in ComfyUI, describe one change, and compare the edited result with the original on your own NVIDIA GPU.
Estimates for one 1024 × 1024 edit with one reference image. Not measured by YouRunAI.
| Machine | FLUX.3 Image Klein Distilled · int8 | FLUX.3 Image Klein Base · int8 |
|---|---|---|
| NVIDIA RTX 3060/3070 · 8 GB VRAM + 32 GB RAM | Publisher’s target class for int8; ComfyUI offloads, so expect slower runs | Works with offloading; 50 steps make it noticeably slower |
| NVIDIA · 16 GB VRAM | Comfortable starting point | Comfortable; slower than distilled |
| NVIDIA RTX 4090/5090 · 24–32 GB VRAM | Fast; you can try the bf16 files | Fast; you can try the bf16 files |
| AMD Radeon · 16 GB VRAM | Not documented for these int8 files; use an NVIDIA GPU or RunPod | Not documented for these int8 files; use an NVIDIA GPU or RunPod |
| Intel Arc · 16 GB VRAM | Not documented for these int8 files; use an NVIDIA GPU or RunPod | Not documented for these int8 files; use an NVIDIA GPU or RunPod |
| Apple Silicon · 16 GB unified memory | Tight for the 8-bit MLX weights; close other apps | Tight; 50 steps make it slow |
| Apple Silicon · 24–32 GB unified memory | Recommended Mac start (mflux 8-bit) | Works; about three times slower than distilled |
| CPU only · 32 GB RAM | Not supported in practice: requires an NVIDIA GPU; use RunPod | Not supported in practice: requires an NVIDIA GPU; use RunPod |
Black Forest Labs also publishes NVFP4 files for RTX 50-series (Blackwell) GPUs and bf16 files for GPUs above 24 GB in the same repository.
Place three files, open the official edit template, load your photo as the reference and describe the change. The template shows the original and the edit side by side so you can check what changed.
Select the weights you want. The instructions, downloads and settings change with your choice.
The publisher has not provided a complete memory target for this task. Check its hardware guidance before starting.
Black Forest Labs says the int8 files are meant for 8 GB GPUs such as the RTX 3070 and for cards without FP8 support; ComfyUI offloads what does not fit. 16 GB is a more comfortable estimate at 1024 × 1024. Above 24 GB you can switch to the bf16 files in the same repository.
Choose your computer first, then its GPU path. Follow the visual setup and import the linked workflow.
Mac: These int8 (convrot) files target NVIDIA GPUs, and neither Black Forest Labs nor the official ComfyUI guide documents running them on Apple Silicon. On a Mac, choose the MLX 8-bit (Mac) package instead; it runs with mflux. Compatibility source
Download ComfyUI Desktop for Windows with an NVIDIA GPU and install it. If you already have it, update it: the FLUX.3 Klein templates need ComfyUI 0.41.0 or later. Desktop asks where to keep your models; this guide assumes the default Documents\ComfyUI folder.
Download ComfyUI Desktop Official FLUX.3 Klein ComfyUI guideThe files are in a gated repository. Sign in to Hugging Face, open the repository below and choose Agree and access repository. Without this step every download link returns an error. The FLUX Kommunity License allows non-commercial use, and commercial use of outputs for individuals and companies under US$5M annual revenue; read it before using images commercially.
black-forest-labs/flux-3-image-klein-comfy FLUX Kommunity LicenseWhile signed in to Hugging Face in your browser, download each file and move it into the matching folder under Documents\ComfyUI/models. Keep the exact filenames. The edit templates do not use the prompt upsampler, so it is not needed here. If you already set up the text-to-image guide, you have all three.
Restart ComfyUI so it sees the new files. Then either open Templates and search for “FLUX 3 Klein” and choose FLUX 3 Image Klein Distilled: Image Edit, or download the workflow JSON below and drag it onto the canvas. These are two ways to load the same graph; use one. If nodes are red or the template is missing, your ComfyUI is older than 0.41.0.
Download the workflow JSON Template details in the ComfyUI guideClick the Load Image node and choose your photo. It is wired to Set Reference Multi Latent, where your first image is called <ref_image_0>. Describe the finished picture and the change, naming the reference, as in the example below. Keep the template’s 4–16 steps (the template default), CFG 1, and the 1024 × 1024 canvas set in Empty Flux 2 Latent; change that size to match your photo’s shape if it is not square.
| Setting | Value |
|---|---|
| Model | flux3-image-klein-int8-convrot.safetensors |
| Canvas | 1024 × 1024 (Empty Flux 2 Latent) |
| Steps | 16 |
| CFG | 1 |
| Example prompt | In <ref_image_0>, change the colour of the jacket to deep red. Keep the face, pose, background and lighting the same. |
Click Queue (Run). The Image Compare node shows your photo and the edited result side by side; drag the slider to check that only the jacket changed. Black Forest Labs notes edits can also alter areas outside the requested region, so look at the face and background closely. To keep the result, right-click it and choose Save Image, or add a Save Image node so it is written to Documents\ComfyUI/output.
Image Compare shows your photo and the edited version; the requested change is there and the face, pose and background are unchanged or only slightly altered, and you saved the result.
Up to 10 reference images are supported; this guide uses one. Template names, files and settings come from the official ComfyUI guide, checked November 9, 2026. YouRunAI has not run this setup.
Sign in to Hugging Face and accept the license on the flux-3-image-klein-comfy page, then download again in the same browser.
Update ComfyUI to 0.41.0 or later and restart it. Check the filenames exactly match the listed files.
Name the reference as <ref_image_0>, say what must stay the same, and keep CFG at 1. Try another seed; Black Forest Labs notes references anchor content but do not lock structure.
Original instructions, model files, and compatibility notes behind this setup.
Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.
Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.
Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.