Workflows/Create images locally with FLUX.2 Klein

Images / WORKFLOW

Create images locally with FLUX.2 Klein

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.

Verified sourceSource checked 9/29/2026

LOCALRENTED GPU

Freshness 35 out of 100. Release age: 273 days old, Jan 15, 2026. Measures age, not quality.

Popularity 76 out of 100. 423,736 downloads in 30 days; 1,043 lifetime likes. Checked Oct 15, 2026. Measures adoption, not quality.

What do these scores mean?

Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.

Freshness · release age

A new release starts at 100. The score halves every 180 days: 100 × 2^(−age in days / 180). When the release date is unavailable, a repository date is a fallback—not proof of when the model was released. Missing dates stay unscored.

Check the date source

Popularity · adoption

Downloads over the past 30 days carry 80% of the score; lifetime likes carry 20%. Both use a logarithmic scale, capped at 10 million downloads and 100,000 likes. If likes are unavailable, downloads carry all the weight. Missing download statistics stay unscored.

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

Check the statistics source

Scores as of Oct 15, 2026. Statistics checked Oct 15, 2026. Statistics are marked for refresh after 72 hours.Full scoring method

Before you begin

Hardware

Estimated full-pipeline capacity: 4B full-precision model, Qwen text encoder, VAE, batch 1 and 1024 px / about 1 MP. ComfyUI swaps components between GPU and system RAM.

Machine	FLUX.2 Klein 4B · distilled	FLUX.2 Klein 4B · base
NVIDIA GPU · 8 GB + 32 GB RAM	Heavy offload may work at 768 px; choose a 12 GB+ GPU for the complete preset. Estimated.	Heavy offload may work at 768 px; choose a 12 GB+ GPU for the complete preset. Estimated.
NVIDIA GPU · 12 GB + 32 GB RAM	Starting point with automatic offload; reduce resolution if needed. Estimated.	Starting point with automatic offload; reduce resolution if needed. Estimated.
NVIDIA GPU · 16–24 GB + 32 GB RAM	Practical 1024 px, batch-1 capacity with automatic model swapping. Estimated.	Practical 1024 px, batch-1 capacity with automatic model swapping. Estimated.
NVIDIA GPU · 32 GB + 64 GB RAM	Ample capacity for this 4B pipeline at 1024 px; no speed claim. Estimated.	Ample capacity for this 4B pipeline at 1024 px; no speed claim. Estimated.
AMD GPU · 16 GB + 32 GB RAM	Supported ROCm GPU required; 1024 px with automatic offload. Estimated.	Supported ROCm GPU required; 1024 px with automatic offload. Estimated.
Intel Arc GPU · 16 GB + 32 GB RAM	XPU backend required; 1024 px with automatic offload. Estimated.	XPU backend required; 1024 px with automatic offload. Estimated.
Apple Silicon · 24 GB unified	Tight at 1024 px; reduce to 768 px or 0.5 MP for editing. Estimated.	Tight at 1024 px; reduce to 768 px or 0.5 MP for editing. Estimated.
Apple Silicon · 32–64 GB unified	Useful headroom for full-precision weights and batch-1 image work. Estimated.	Useful headroom for full-precision weights and batch-1 image work. Estimated.
CPU only · 32 GB RAM	Not recommended for this image workflow; use a GPU route or a rented 16 GB GPU.	Not recommended for this image workflow; use a GPU route or a rented 16 GB GPU.

Memory capacity is not a speed measurement. AMD/Intel require the correct supported GPU backend; Mac uses Apple Silicon Metal. These are source-informed estimates, not YouRunAI hardware benchmarks. Multi-GPU memory is not pooled by this graph.

LOCAL IMAGE GENERATIONSOURCE-LINKED SETUP

Create images locally with FLUX.2 Klein

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. Choose your package and machine, then use Comfy Desktop to choose a built-in template or the supplied workflow JSON, download its models and save a result.

Choose a package Load the visual workflow Check the result

BEFORE YOU BEGIN

Choose your package

Select the weights you want. The instructions, downloads and settings change with your choice.

Save your machine in My Hardware to get an automatic starting choice. You can always choose any package yourself.

CURRENT SETUP

FLUX.2 Klein 4B · base · official

Estimated for the complete full-precision pipeline at 1024 px / about 1 MP, batch 1. Allow 32 GB system RAM for offload; use 24 GB minimum / 32 GB recommended Mac unified memory. An 8 GB card may need reduced resolution and substantial offload. This is capacity guidance, not a speed benchmark.

Freshness 35 out of 100. Repository age: 274 days old, Jan 14, 2026. Measures age, not quality.

Popularity 73 out of 100. 400,431 downloads in 30 days; 180 lifetime likes. Checked Oct 15, 2026. Measures adoption, not quality.

What do these scores mean?

Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.

Freshness · release age

Check the date source

Popularity · adoption

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

Check the statistics source

Scores as of Oct 15, 2026. Statistics checked Oct 15, 2026. Statistics are marked for refresh after 72 hours.Full scoring method

View weight source

THE WORKFLOW

Set up FLUX.2 Klein 4B · base on your machine

Choose your computer first, then its GPU path. Follow the visual setup and import the linked workflow.

CHOOSE YOUR SETUPWindows

Hardware type

NVIDIA (CUDA) through Comfy Desktop. An existing working ComfyUI installation can use the same JSON and model files; a second installation is unnecessary.

Open Comfy Desktop — install only if needed

Already have a working, current ComfyUI? Skip installation and continue to the workflow choice below. Otherwise: Download the Windows installer, open the .exe and follow the setup screens. Launch Comfy Desktop from the Start menu. Choose New Instance → Standalone, the Stable channel and NVIDIA (CUDA) when the hardware options appear. Let the app install Python and dependencies for you; no separate Git or Python setup is needed. Use the current NVIDIA driver. For RTX 20-series and newer, including RTX 5090, select the recommended stable CUDA build; respect any driver or GPU-compatibility warning. When ready, click the instance card to open ComfyUI. In the card’s ⋮ → Manage → About, confirm the expected GPU and PyTorch backend rather than CPU before downloading large files.

Download Comfy Desktop Official Windows installation NVIDIA driver downloads

Choose Templates or the supplied workflow JSON

For the fewest setup changes, download this package’s JSON below (use Save link as if it opens as text), then drag it onto the ComfyUI canvas. It already contains the exact model selections, preset and output path used in this guide; no custom-node packs are needed. Or use ComfyUI Templates: click Templates in the sidebar (also Workflow → Browse Workflow Templates), search for “Flux.2 [Klein] 4B: Text to Image” and open that exact template. Keep “Text to Image (Flux.2 Klein 4B)” and its Save Image enabled, with the Distilled branch bypassed. Check flux-2-klein-base-4b.safetensors, 20 steps and CFG 5. Both variants use qwen_3_4b.safetensors and flux2-vae.safetensors. Download only the selected package’s files below. Double-click the empty area of its subgraph to inspect the internal settings, then apply the prompt and fixed-seed preset below. Set Save Image filename_prefix to YouRunAI/flux-klein-base-image so the save-location instructions match. If the exact template is absent, use the JSON or update ComfyUI and its template package. Use one route per run. The supplied JSON remains available if you prefer the expanded graph or want this guide’s exact setup.

Download this exact ComfyUI workflow JSON How to open ComfyUI Templates Official Flux.2 [Klein] 4B: Text to Image template Original official graph and explanation

Download the matching model files

If you chose Templates, download only your selected 4B variant, not both branches. In Desktop’s missing-model dialog, click Download for each required file and wait for completion. Models already present can be reused. If the prompt is unavailable, use the exact links below: open ☰ → Desktop Settings → Storage → Shared Models, then click the folder icon beside Primary. The tree starts INSIDE that primary model folder; do not add another models folder. Older manual/portable installations use their existing ComfyUI/models folder instead. Confirm Load Diffusion Model selects flux-2-klein-base-4b.safetensors, Load CLIP selects qwen_3_4b.safetensors with type flux2, and Load VAE selects flux2-vae.safetensors. After downloads finish, reload the workflow and reselect a file if a loader still shows it missing.

Put these files in Your primary model folder/3 required files

diffusion_models/

flux-2-klein-base-4b.safetensorsBase 4B model; 20 denoising stepsGet file

text_encoders/

qwen_3_4b.safetensorsEncodes your text instruction; shared by both packagesGet file

vae/

flux2-vae.safetensorsEncodes reference images and decodes the final imageGet file

Desktop model downloads Find your shared model folder

Enter a prompt and check the preset

In the Image prompt node (the template’s Prompt field), enter: A matte black water bottle on a warm stone pedestal, soft studio lighting, clean product photograph, no text. Avoid changing several controls at once. Keep the fixed seed for your first comparison. Set both Width and Height PrimitiveInt nodes to 1024. They drive the latent image and scheduler together; batch_size remains 1. Flux2Scheduler uses 20 steps, CFGGuider 5, KSamplerSelect euler and RandomNoise seed 20260928 with control_after_generate fixed. On a 12 GB card start at 768 × 768 if the 1024 render exceeds memory. If using Templates, the sampler, resize and seed controls are inside the selected subgraph: double-click its empty area to open it, then use the top navigation or Esc to return. Follow the values above even when the template’s defaults differ.

Generate and inspect the image

Click Run once. Wait for the active queue item to finish; the queue shows progress. The first run can take longer while models load. Inspect the Save Image preview at full size. Confirm the requested subject appears and the image is not black or corrupt. A completed render is the success check; importing a graph alone is not.

Save your result and return later

Use the Save Image preview menu to download your PNG. ComfyUI also saves it in the configured output folder’s YouRunAI subfolder. Find that folder in Desktop Settings → Storage → Shared Output; if shared output is disabled, use the instance’s Manage → Storage → Output Directory. Save/export the workflow JSON after editing settings so you can reopen the same setup. Change one prompt or seed and click Run for another result. To stop an unwanted render, use the queue’s Cancel/Interrupt control. To shut down, return to the instance card → ⋮ → Stop; tomorrow, open Comfy Desktop, click that card and load your saved JSON. No terminal launch commands are needed.

SUCCESS CHECK · FLUX.2 Klein 4B · base

Your first image is saved

The requested subject appears in the Save Image preview, a full-size PNG exists in your configured output folder’s YouRunAI subfolder, and you have saved the matching workflow JSON. Reopen it later with the same package and seed.

Verified refers to the linked model/runtime documentation and reviewed graph structure. Hardware advice is Estimated. No render-time or YouRunAI physical-execution claim is made. This focused setup uses publisher model packages; it does not bundle a speculative uncensored LoRA or claim that every third-party derivative works.

Back to setup steps

When it doesn’t go to plan

A loader says the model file is missing

Check the exact filename in this package’s file tree and finish the Desktop download. If downloading manually, put the file inside the Primary shared model folder’s indicated subfolder; for manual/portable ComfyUI use its models folder. Reload the workflow and reselect the file. If it remains missing, stop the instance from its card menu and click the card to relaunch. A .safetensors file must not be a partial download or an HTML error page.

The GPU check fails or generation falls back to the CPU

Open the instance card → ⋮ → Manage → About and inspect Hardware and PyTorch. Confirm NVIDIA/CUDA, AMD/ROCm, Intel/XPU or Apple/MPS for your route. Update the vendor driver if needed. Desktop can change the PyTorch stack from its Manage panel (About → Change PyTorch; some versions place the picker on Update). Use an offered stable stack compatible with your actual GPU and OS. On Linux AMD, resolve any missing /dev/kfd or GPU-access warning using the vendor instructions. More VRAM does not fix an unsupported backend.

A core node is missing when the JSON is loaded

Update the active ComfyUI instance through its card → Manage → Update → Check for update, using Stable, then stop and reopen it and reload the JSON. Updating only the Desktop launcher does not update an older instance. The supplied graphs use ComfyUI core nodes; do not install unrelated custom-node packs to replace them. If a node is still absent in the current stable release, compare the linked official template requirements before changing release channels.

Generation runs out of memory

Cancel the job, keep batch 1 and close other GPU apps. For text-to-image reduce both width/height controls to 768. For editing reduce ImageScaleToTotalPixels from 1 to 0.5 megapixels. Leave automatic offload enabled; on a PC with enough system RAM, set --lowvram in the instance’s Manage → Startup Args if necessary, then stop and reopen it before retrying.

The subject does not match the prompt

Keep the seed fixed and describe one subject, setting and lighting style. Check that the selected 4-step or 20-step preset matches the downloaded distilled or base model. Change one variable at a time; neither package guarantees readable small text or exact product branding.

REFERENCE LIBRARY

Sources and files

Original instructions, model files, and compatibility notes behind this setup.

11 SOURCES

01Official FLUX.2 Klein ComfyUI guidedocs.comfy.org 02FLUX.2 Klein publisherhuggingface.co 03Official ComfyUI graph sourcegithub.com 04AMD GPU / OS compatibilityrocm.docs.amd.com 05Intel XPU setupdocs.pytorch.org 06Apple Metal setupdeveloper.apple.com 07Comfy Desktop installationdocs.comfy.org 08Comfy Desktop: manage GPU, updates and storagedocs.comfy.org 09Shared models and output foldersdocs.comfy.org 10Automatic model downloads from workflowsdocs.comfy.org 11Official built-in Flux.2 [Klein] 4B: Text to Image templategithub.com

The model behind this workflow

FLUX.2 Klein · 4B

Keep exploring

Text

SETUP GUIDE

Qwen3.5 · 9B private chat 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.

Freshness 41 out of 100. Release age: 230 days old, Feb 27, 2026. Measures age, not quality.

Build a private coding assistant with Qwen3.8 27B

Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.

Freshness 76 out of 100. Repository age: 71 days old, Aug 5, 2026. Measures age, not quality.

Arena Open Source Web coding: rank 10 of 55. Ranked model: qwen3.8-27b. Original model ranking, not a benchmark of this package. Export dated Oct 13, 2026. Source checked Oct 15, 2026. Human preference ranking, not hardware suitability.+3

SOFTWAREOllama + Visual Studio Code + Cline + Python

Generate videos locally with Wan2.1

Turn text prompts into local videos with the Wan2.1 T2V 1.3B model, complete pipeline files and a ready-to-load ComfyUI workflow.

Freshness 10 out of 100. Repository age: 597 days old, Feb 25, 2025. Measures age, not quality.

SOFTWAREComfyUI

LOCALRENTED GPU

Verified source