To run FLUX 3 Image Klein locally, update ComfyUI, download the black-forest-labs/flux-3-image-klein-comfy package from Hugging Face, install its transformer, text-encoder, prompt-upsampler and VAE files in the locations used by the supplied workflow, then open the FLUX 3 Image Klein template. Choose the distilled checkpoint for normal image generation and editing, or the base checkpoint for fine-tuning, LoRA training and research. There is no published universal VRAM minimum, so memory requirements depend on precision, resolution, references and batch size.
What FLUX 3 Image Klein is and what a local setup can do
FLUX 3 Image Klein is a 5.6-billion-parameter, open-weights image model designed for both text-to-image generation and image editing. Its workflow can accept as many as 10 reference images, use written layout instructions with bounding boxes, render text, transfer a reference style and generate at native 4K resolution. Running it in ComfyUI keeps the workflow on hardware you control instead of sending prompts and source images to a hosted service.
As an Amazon Associate I earn from qualifying purchases.
The model is not a single file that you drop into a generic checkpoint folder. The ComfyUI package includes separate components: the diffusion transformers, a text encoder, a prompt upsampler and an autoencoder (VAE). The current template connects those components with the node types expected by FLUX 3 Image Klein. Installing only a transformer or only a VAE will leave the workflow incomplete.
There are two principal checkpoints. The step-distilled version is optimized for ordinary generation and editing in 16 steps without guidance. The base version runs for 50 steps with guidance set to 3.5 and is intended for fine-tuning, LoRA training and research. Quantized INT8, FP8 and NVFP4 variants can lower memory pressure, but they require loaders that understand the selected format.
#1 Best Overall
- Powered by the NVIDIA Blackwell architecture and DLSS 4 OC mode: 2640MHz/Default mode: 2610MHz (Boost Clock)
- Military-grade components deliver rock-solid power and longer lifespan for ultimate durability
- Protective PCB coating helps protect against short circuits caused by moisture, dust, or debris
- 3.125-slot design with massive fin array optimized for airflow from three Axial-tech fans
- Phase-change GPU thermal pad helps ensure optimal thermal performance and longevity, outlasting traditional thermal paste for graphics cards under heavy loads
Before you install: software, files and hardware decisions
Use a current ComfyUI installation
Update ComfyUI before loading the workflow. The supported installation routes include the current desktop and manual paths for Windows and Linux, as well as Apple Silicon macOS. Recent ComfyUI builds include model offloading and support for quantized models, both of which can affect whether a large workflow fits on your machine. An older build may not recognize the FLUX 3 Image Klein nodes, model type or quantized files.
Download the complete model package
Accept any Hugging Face access conditions shown for the repository, then download black-forest-labs/flux-3-image-klein-comfy. The package is arranged for ComfyUI and contains the base and distilled transformers, text encoder, prompt upsampler and autoencoder. Download the components you actually intend to use; the base and distilled transformers are alternatives, while the shared supporting components are required by the workflow.
Plan around workload rather than a made-up VRAM number
The official launch material describes FLUX 3 Image Klein running on a single consumer GPU, but the official sources do not publish a universal minimum VRAM requirement. A file’s size is not a reliable prediction of total GPU memory use. Your practical requirement changes with:
- Precision: BF16 uses more memory than an appropriately supported INT8, FP8 or NVFP4 build.
- Resolution: 4K generation requires substantially more working memory than a 1024×1024 image.
- References: attaching several reference images increases the conditioning workload.
- Batch size: generating multiple images at once consumes more memory than a batch of one.
- Workflow choices: offloading and other ComfyUI memory settings can change what fits.
Treat any GPU example as a starting point, not a guarantee. If a workflow runs out of memory, first lower the output size or batch size, reduce the number of references, select a supported quantized build, or enable suitable offloading.
Install FLUX 3 Image Klein in ComfyUI
- Update ComfyUI. Use the current ComfyUI desktop updater or your manual installation procedure. Restart ComfyUI after updating so the new model definitions and nodes are loaded.
- Open the Hugging Face repository. Locate
black-forest-labs/flux-3-image-klein-comfy, accept the access terms if the site requests them, and download the transformer, text encoder, prompt upsampler and autoencoder files. Keep the base and distilled transformer files identifiable; you will choose between them in the workflow. - Place files in ComfyUI’s model directories. Copy each component into the model folder expected by the current FLUX 3 Image Klein template. Do not put every file into one generic directory: the template expects the diffusion model, CLIP/text encoder, prompt upsampler and VAE to be discoverable by their respective loader nodes. If the template displays a missing-file warning, use its loader’s model-location convention rather than renaming files at random.
- Open the template. In ComfyUI, open the Templates panel and load the FLUX 3 Image Klein workflow. The supplied graph is the safest starting point because it already wires the model-specific loaders and conditioning stages.
- Check the loader types. The diffusion model should use the standard Load Diffusion Model node. The text encoder should be loaded with
Load CLIPand typeflux3. The image decoder should use Load VAE. Select the transformer and supporting files you installed; a loader showing “model not found” means the file is in the wrong directory or the ComfyUI model index has not refreshed. - Keep the prompt upsampler connected. The FLUX 3 Image Klein model card explicitly says, “Do not use the model without a prompt upsampler.” Enable the supplied upsampler in the template, or connect another language model that expands your short request into a complete scene caption.
- Run a text-to-image test. Enter a full scene description, leave references disconnected for the first test, choose a moderate output size and queue the prompt. A successful run should produce an image and no missing-model or out-of-memory error.
- Add editing inputs after the basic test works. Connect one reference image, queue again, then add more references or layout instructions. Testing incrementally makes it clear whether a failure comes from installation, a particular image, or memory pressure.
What each downloaded component does
| Component | Purpose in the workflow | What to check |
|---|---|---|
| Base transformer | Full 50-step model for research, fine-tuning and LoRA training | Select it when you need training-oriented behavior; guidance is 3.5 |
| Distilled transformer | Fast 16-step model for normal generation and editing | Use it for routine self-hosted creation; it runs without guidance |
| Text encoder | Converts the text conditioning used by the FLUX 3 workflow | Load it through Load CLIP with type flux3 |
| Prompt upsampler | Expands a short request into the detailed scene caption the model expects | Keep it enabled; the model card warns against running without one |
| Autoencoder (VAE) | Converts between the model’s latent representation and viewable images | Load it with the template’s Load VAE node |
Base or distilled: which checkpoint should you choose?
| Decision point | Distilled checkpoint | Base checkpoint |
|---|---|---|
| Primary purpose | Everyday text-to-image generation, image editing and self-hosting | Fine-tuning, LoRA training and research |
| Sampling schedule | 16 steps | 50 steps |
| Guidance | No guidance | Guidance 3.5 |
| Speed expectation | Lower step count is intended to reduce generation time | More steps generally make a run longer |
| Best first choice | Use this unless you have a training or research reason to use base | Use when your workflow specifically requires the full base model |
Distillation does not mean the model is a lower-quality preview that must be replaced for ordinary work. It is the checkpoint intended for fast generation and editing. The base checkpoint is the practical choice when you need to adapt the model with a LoRA, fine-tune it, or investigate model behavior. Switching checkpoints changes the appropriate step count and guidance settings, so do not leave a 50-step base configuration attached to the distilled model or remove guidance from a base workflow without understanding the template’s settings.
Prompting: why the upsampler is required
A short prompt such as “a red bicycle in a studio” does not provide the scene specification FLUX 3 Image Klein is designed to consume. The upsampler expands that request into a full caption describing subject, composition, environment, lighting, materials, camera viewpoint and other useful details. The model card’s instruction is categorical: do not run the model without a prompt upsampler.
Write the initial prompt as an intention, then let the upsampler elaborate it. For controllable results, include the subject, action, setting, time of day, visual style, important colors, text that must appear and any spatial relationship. Review the expanded caption when the workflow exposes it. If the expansion adds an unwanted object or changes the intended style, make the original prompt more explicit or use another compatible language model as the upsampler.
Text-to-image, editing and layout workflows
Text-to-image
Start with no reference images. Describe the complete scene, select the distilled transformer for a normal creation job, keep prompt upsampling enabled and queue a single image. Once the basic graph works, increase resolution or batch size one setting at a time so you can identify the point at which your GPU runs out of memory.
Single-reference editing
Attach one source image through the reference input in the template and describe the change in the prompt. State what must remain unchanged as well as what should be replaced. For example, identify the subject’s pose and background as fixed while requesting a new jacket color. The model can preserve reference content, but an edit prompt is not a pixel-perfect mask; inspect edges, small text and other details in the result.
Multi-reference composition
The workflow supports up to 10 reference images. Use separate references for subjects, products, environments or style cues, and explain each reference’s role in the prompt. More references provide more conditioning information and can also increase memory use. Add them gradually rather than attaching 10 images before you have confirmed that one reference works.
Text rendering
Include the exact wording, capitalization and intended placement in the prompt. Examine the output at the final delivery resolution because small lettering that looks acceptable in a preview may be distorted when enlarged or downsampled.
Recommended Free Tools
Bounding-box layout instructions
FLUX 3 Image Klein accepts bounding-box instructions on a 0–1000 coordinate grid. Use the grid to describe where an object should appear and how large its region should be. A box guides placement and scale; it is not a hard pixel mask, so the subject can extend beyond the coordinates and edges may not be exact. Combine a box with a clear textual description of the object and its relationship to the rest of the scene.
Reference-based style transfer
Supply a style reference and state which visual properties should transfer—such as palette, brushwork, lighting or material treatment—while separately describing the new subject. This separation helps prevent the reference’s original subject from being copied when you only want its appearance.
Rank #2
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5070 Ti
- Integrated with 16GB GDDR7 256bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
Native 4K generation
The model supports native 4K output, but 4K is a demanding first test. Begin at a lower resolution to validate the graph, then raise the dimensions after the prompt, references and checkpoint are working. At 4K, reduce batch size and reference count if memory becomes the limiting factor.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Quantization and performance
INT8, FP8 and NVFP4 choices
INT8, FP8 and NVFP4 builds are available to reduce memory pressure. Quantization is not a universal drop-in switch: the model card warns that each quantized package needs a loader supporting that format, and the on-disk file size alone does not predict total GPU memory consumption. Select a quantized transformer only when your ComfyUI build and loader support it, and keep a standard-precision option available for diagnosing compatibility problems.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What the published benchmark actually shows
ComfyUI’s October 15, 2026 performance report measured 1024×1024 text-to-image generation with the step-distilled model. It reported that pairing INT8 ConvRot quantization with the distilled model was 4.3× to 9.3× faster than a BF16, 50-step baseline. In the examples, an RTX 5070 run fell from 71 seconds to under 8 seconds, and an RTX 4070 run fell from 78 seconds to under 9 seconds.
Those are ComfyUI’s own published tests, not independent measurements. They describe a particular resolution, precision, quantization method, checkpoint and baseline, so they should not be treated as a promised time for every workflow. Editing with references, 4K output, a different sampler, a base checkpoint or a different driver can produce materially different latency.
Diagnosing common setup failures
“Model not found” or an empty loader list
- Confirm that the download completed and was not left as a partial archive.
- Move the transformer, text encoder, prompt upsampler and VAE into the directories used by the current template rather than one shared folder.
- Restart or refresh ComfyUI so it rescans model files.
- Check that the loader’s model type matches the file: the CLIP node must use type
flux3.
The workflow starts but fails at the prompt stage
Make sure the prompt upsampler node is present, enabled and connected to the text-conditioning path. A short prompt sent directly to the model violates the model card’s required usage. If the supplied upsampler cannot load, verify its file and use a compatible alternative language model only after confirming that the template accepts it.
Out-of-memory errors
- Lower the output resolution and set the batch size to one.
- Remove extra reference images and test with a single reference or none.
- Switch to a supported INT8, FP8 or NVFP4 build and its matching loader.
- Enable ComfyUI’s available model-offloading options.
- Use the distilled 16-step checkpoint for routine generation instead of the 50-step base model.
These changes address different causes, so test one at a time. A quantized file may still require more memory than its file size suggests, and a workflow that fits at 1024×1024 may fail at 4K.
Free tools Windows power users keep installed
One-click scans. No signup required.
Results are generic or ignore the layout
Confirm that prompt upsampling is active and rewrite the prompt as a complete scene caption. For layout, describe the object, its box coordinates on the 0–1000 grid, its intended scale and its relationship to neighboring objects. Remember that boxes guide placement; they do not cut out a hard mask. For edits, identify which elements are fixed and which are allowed to change.
The quantized model will not load
Use a loader that explicitly supports the selected INT8, FP8 or NVFP4 format and a current ComfyUI build. If the loader reports an unsupported type, return to the standard-precision transformer to verify the rest of the graph, then install or enable the appropriate quantized-loader support before trying again.
Commercial use and the FLUX Kommunity License
The official Comfy Q&A describes free non-commercial use under the FLUX Kommunity License. It says users and their affiliates with less than US$5 million in annual revenue may use generated images commercially when they filter or review outputs and disclose AI generation where the law requires it. The revenue threshold applies to the user and affiliates as described by the license, not merely to the individual project using the image.
Other commercial use, including commercial fine-tuning, requires a license from Black Forest Labs. That distinction matters if you plan to train a LoRA for a paid product, operate a commercial image-generation service or use the base checkpoint in a business whose revenue does not meet the stated condition. Check the current license text before shipping paid work because licensing terms can change.
Keep a record of the checkpoint, quantization, prompt and any reference assets used for a commercial deliverable. Review every output for unwanted copyrighted material, personal data, unsafe content and inaccurate text before publication; the license condition to filter or review outputs is a practical reason to include a human approval step in your workflow.
When local ComfyUI is the right choice
A local installation makes sense when you own a suitable GPU, need privacy for prompts or reference images, want repeatable access without a per-image hosted charge, or need to customize a graph with references, layout controls and post-processing. It also lets you choose between base, distilled and quantized files and keep the workflow available when an internet connection is not.
A hosted ComfyUI option can avoid installing model files and managing GPU drivers, but it moves the computation and source images to a provider and may charge for usage. It is an alternative, not a prerequisite for FLUX 3 Image Klein. For an offline-first setup, finish the local installation and test a complete run before disconnecting from the internet; the model files and all custom nodes required by the template must already be present.
Quick Recap
A reliable first-run checklist
- ComfyUI is updated to its current release and restarted.
- The Hugging Face access conditions are accepted if prompted.
- The complete
black-forest-labs/flux-3-image-klein-comfypackage is downloaded. - The transformer, text encoder, prompt upsampler and VAE are in the directories expected by the template.
- The CLIP loader is set to
flux3, and the correct VAE is selected. - The distilled checkpoint is selected for ordinary generation, or the base checkpoint is selected for training and research.
- The prompt upsampler is enabled and connected.
- The first test uses one image, a moderate resolution, batch size one and no reference images.
- References, bounding boxes, text rendering and 4K output are added only after the basic text-to-image run succeeds.
- Any commercial output is reviewed under the current license conditions before distribution.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →
