NodeWise Studio
RUDownload for Linux

Flux.2 Klein: a sampler built from parts

<!-- lesson: flux2-klein -->

Lesson "flux2-klein", in the group Pictures: models of the lesson list. Steps: 15.

What you will be able to do

You build the graph of a small Flux.2 model, where the sampler is put together from five nodes.

What you need

  • It goes on from Flux.1: a picture from separate files. The card says so; it is a hint, not a lock.
  • Model files the lesson checks:
    • a Flux.2 Klein model: models/diffusion_models, for example flux-2-klein-base-4b-fp8.safetensors;
    • qwen_3_4b, or qwen_3_8b for the 9B model: models/text_encoders, for example qwen_3_8b_fp8mixed.safetensors;
    • flux2-vae: models/vae, for example flux2-vae.safetensors.
  • Nodes: Load Diffusion Model, Load CLIP, Load VAE, CLIP Text Encode (Prompt), CFG Guider, Empty Flux 2 Latent, Flux2Scheduler, KSamplerSelect, RandomNoise, SamplerCustomAdvanced, VAE Decode, Save Image / Preview Image.
  • The app downloads and installs nothing. The step texts below say which files and folders, and which ComfyUI template names them.

The steps

The words of each step, as the lesson shows them. Node names are in bold.

  1. Flux.2 and its small models (reading)

    Flux.2 by Black Forest Labs comes as one large model, dev, and small ones, Klein, in two sizes: 4B and 9B. Each Klein comes as a base model and as a distilled one.

    Flux.2 is not Flux.1 made bigger. Its text encoder is a language model, its VAE and empty latent are its own, and no file of Flux.1 fits it.

    The official templates build its sampler from parts: KSampler takes everything in one node, while here the noise, the guide, the sampling method and the schedule are separate nodes. This is the graph of the template "FLUX.2 [Klein] 4B: Text to Image".

  2. Which files your family needs (reading)

    For Klein 4B:

    • the model: flux-2-klein-base-4b.safetensors (base) or flux-2-klein-4b.safetensors (distilled) in models/diffusion_models;
    • the text encoder: qwen_3_4b.safetensors in models/text_encoders, the same file as for Z-Image;
    • the VAE: flux2-vae.safetensors in models/vae.

    Klein 9B takes flux-2-klein-base-9b-fp8.safetensors and the larger encoder qwen_3_8b_fp8mixed.safetensors. Flux.2 dev takes a Mistral encoder and is many times larger.

    If a file is not in the list, your ComfyUI does not have it. The app downloads nothing. Open the template "FLUX.2 [Klein] 4B: Text to Image" in ComfyUI itself: it shows which files are missing and where to put them. Take files only from the model's official page. Then choose "Sync ▸ Sync now" in this app's menu, and the files appear in the lists.

  3. Load the model

    Add a Load Diffusion Model node and choose the Klein file in unet_name.

    This lesson sets the numbers of the base model. For the distilled one, the step "Base and distilled" says what changes.

  4. Load the text encoder

    Add a Load CLIP node. Choose the Qwen file in clip_name and set type to flux2.

    The same file serves Z-Image with another type: the type tells ComfyUI which model the encoder feeds.

  5. Load the VAE

    Add a Load VAE node and choose flux2-vae.safetensors in vae_name.

    The VAE of Flux.1 does not fit Flux.2: the compressed form of the picture is another one.

  6. Write the prompt and the negative

    Add two CLIP Text Encode (Prompt) nodes and connect the output of Load CLIP to the clip input of each. In the text field of the first, describe the picture in sentences: a hedgehog in a tiny party hat among confetti, flash photo of an old digital camera.

    The base model reads the negative, so the second node may hold what is not wanted, or stay empty.

  7. The guide: how closely to follow the prompt

    Add a CFG Guider node and connect:

    • to model, the output of Load Diffusion Model;
    • to positive and negative, the two text nodes.

    Set cfg to 5. This node is the part of KSampler that weighs the prompt against the negative.

  8. Add the empty latent of Flux.2

    Add a Empty Flux 2 Latent node and set width and height to 1024.

    Its latent has 128 channels at a sixteenth of the size, and this is the node of ComfyUI's template. Both sides must divide by 16.

  9. The schedule of the noise

    Add a Flux2Scheduler node. Set steps to 20, and width and height the same as in Empty Flux 2 Latent.

    This node works out how much noise is left before each step. For Flux.2 it depends on the size of the picture, so the size here must be that of the empty latent.

  10. The sampling method and the noise

Add a KSamplerSelect node and choose euler in sampler_name. Add a RandomNoise node: its noise_seed field is the seed, the number the starting noise is made from.

In KSampler these are two fields; here they are two nodes.

  1. Put the sampler together

Add a SamplerCustomAdvanced node and connect its five inputs:

  • noise: the output of RandomNoise;
  • guider: the output of CFG Guider;
  • sampler: the output of KSamplerSelect;
  • sigmas: the output of Flux2Scheduler;
  • latent_image: the output of Empty Flux 2 Latent.

This node does the work. Its first output, output, is the finished latent.

  1. Turn the result into a picture

Add a VAE Decode node. Connect the output output of SamplerCustomAdvanced to its samples input and the output of Load VAE to its vae input.

  1. Save the picture

Add a Save Image node and connect the output of VAE Decode to its images input.

Without it ComfyUI works the picture out and puts it nowhere.

  1. Base and distilled (reading)

The distilled Klein is made for speed. In the same template its numbers are: steps 4 in Flux2Scheduler, cfg 1 in CFG Guider, and an empty conditioning node instead of the negative prompt: at cfg 1 the negative is not read.

Flux.2 dev is built the same way with two changes: a Basic Guider without a negative instead of CFG Guider, and a Flux Guidance node with 4 on the prompt.

  1. The graph is ready (reading)

Save the graph with "File ▸ Save workflow for ComfyUI…", open the file in ComfyUI and run it there. This app runs nothing in ComfyUI itself.

If the result is bad:

  • grey noise or an error: the VAE is not the one of Flux.2;
  • a blurred or strange picture: the size in Flux2Scheduler differs from that of Empty Flux 2 Latent;
  • burnt colours on the distilled model: cfg is above 1.

LoRAs of Flux.1 do not fit Flux.2, and those of Klein 4B do not fit Klein 9B.

What the lesson checks

A reading step is done when you press Got it, next. A doing step ticks itself off when the canvas holds:

  • Step 3, Load the model:
    • a Load Diffusion Model node.
    • unet_name of Load Diffusion Model holds a Flux.2 Klein model; such a file lies in models/diffusion_models, for example flux-2-klein-base-4b-fp8.safetensors.
  • Step 4, Load the text encoder:
    • a Load CLIP node.
    • clip_name of Load CLIP holds qwen_3_4b, or qwen_3_8b for the 9B model; such a file lies in models/text_encoders, for example qwen_3_8b_fp8mixed.safetensors.
    • type of Load CLIP is flux2. Why: a text encoder of the wrong type reads the prompt wrongly, and the model gets nonsense.
  • Step 5, Load the VAE:
    • a Load VAE node.
    • vae_name of Load VAE holds flux2-vae; such a file lies in models/vae, for example flux2-vae.safetensors.
  • Step 6, Write the prompt and the negative:
    • a CLIP Text Encode (Prompt) node.
    • a CLIP Text Encode (Prompt) node.
    • CLIP of Load CLIP into clip of CLIP Text Encode (Prompt) (prompt).
    • CLIP of Load CLIP into clip of CLIP Text Encode (Prompt) (negative prompt).
    • text of CLIP Text Encode (Prompt) (prompt) holds at least 20 characters.
  • Step 7, The guide: how closely to follow the prompt:
    • a CFG Guider node.
    • MODEL of Load Diffusion Model into model of CFG Guider (a LoraLoaderModelOnly in between is accepted).
    • CONDITIONING of CLIP Text Encode (Prompt) (prompt) into positive of CFG Guider.
    • CONDITIONING of CLIP Text Encode (Prompt) (negative prompt) into negative of CFG Guider.
    • cfg of CFG Guider is from 3 to 6; Do this for me sets 5. Why: the Klein base model is guided at 5, as its template does.
  • Step 8, Add the empty latent of Flux.2:
    • a Empty Flux 2 Latent node.
    • width of Empty Flux 2 Latent is from 512 to 2048; Do this for me sets 1024. Why: flux was trained near one megapixel: 1024 by 1024 or the same number of pixels in other proportions.
    • height of Empty Flux 2 Latent is from 512 to 2048; Do this for me sets 1024. Why: flux was trained near one megapixel: 1024 by 1024 or the same number of pixels in other proportions.
  • Step 9, The schedule of the noise:
    • a Flux2Scheduler node.
    • steps of Flux2Scheduler is from 10 to 50; Do this for me sets 20. Why: the schedule of the Klein template has 20 steps.
    • width of Flux2Scheduler equals width of Empty Flux 2 Latent. Why: the schedule depends on the picture size; it must equal the size of the latent.
    • height of Flux2Scheduler equals height of Empty Flux 2 Latent. Why: the schedule depends on the picture size; it must equal the size of the latent.
  • Step 10, The sampling method and the noise:
    • a KSamplerSelect node.
    • a RandomNoise node.
    • sampler_name of KSamplerSelect is euler.
  • Step 11, Put the sampler together:
    • a SamplerCustomAdvanced node.
    • NOISE of RandomNoise into noise of SamplerCustomAdvanced.
    • GUIDER of CFG Guider into guider of SamplerCustomAdvanced.
    • SAMPLER of KSamplerSelect into sampler of SamplerCustomAdvanced.
    • SIGMAS of Flux2Scheduler into sigmas of SamplerCustomAdvanced.
    • LATENT of Empty Flux 2 Latent into latent_image of SamplerCustomAdvanced.
  • Step 12, Turn the result into a picture:
    • a VAE Decode node.
    • output of SamplerCustomAdvanced into samples of VAE Decode.
    • VAE of Load VAE into vae of VAE Decode.
  • Step 13, Save the picture:
    • a Save Image or Preview Image node.
    • IMAGE of VAE Decode into images of Save Image or Preview Image.

See also

This page is imported from the application repository's wiki (docs/wiki/lessons/Flux2-Klein.md, revision b8e3f71). The same guide is built into the app: Help ▸ User guide.

To refresh it, run npm run import and then npm run build in the website folder.