NKD Klein Tools

ComfyUI nodes that turn a Flux Klein workflow into something simple. Plug in your model, drop an image, write a prompt, and go — no manual wiring of internal pieces. Whether you want to generate from scratch, transform an existing photo, paint over a specific area, or zoom in for a high-detail touch-up, a couple of nodes handle it all. There's also an optional node to control how much each reference image shows up in the result.

https://github.com/user-attachments/assets/f84cc919-325d-465b-8d3d-e178de5f7872

Full introduction tutorial

---

What's new in 1.10.x

• New node: 😺NKD Klein Prompt Builder — build a prompt visually by combining your own text with curated presets (style, lighting, camera angle, composition, mood, and more) from dropdowns, with a live preview of the result. Switch between flowing prose (best for Klein) and a structured JSON template for automation. Wire its prompt output into Presampling's positive input. The dropdown options are defined in klein_presets.json — edit that file to add, remove or reword presets to your taste, then restart ComfyUI.

• New (experimental): regional reference control — send a specific reference image to a specific area of the canvas. Paint a mask for the zone and the reference's influence is confined there (and can be reinforced inside it), instead of the model spreading it across the whole image. A small family of nodes on the model line, between Presampling and your sampler:

  • 😺NKD Klein Reference Region — confines one reference to a masked zone, with controls for how strong it is inside, how firmly it's held back outside, and how crisp the edge is.
  • 😺NKD Klein Reference Fit — scales a reference to sit inside the masked area, so the whole reference lands in the zone instead of only the part that happens to overlap it.
  • 😺NKD Klein Reference Control — one node that does it all: overall strength + optional per-step curve + regional confinement (the mask is optional — without it, it's just a strength control). Chain one per reference.

  Experimental: it works well in testing, but on strong settings expect a little bleed into neighbouring areas. Feedback welcome.

https://github.com/user-attachments/assets/a62cd1a4-6c2c-4ee4-89c8-e515857a4835

What's new in 1.9.x

• New node: 😺NKD Klein Reference Weight — when you use more than one reference image, this lets you decide how much each one shows up in the result, one at a time. Turn a reference up so it asserts itself, or down so it stops dominating. 1.0 leaves it as usual, lower fades it out (0 = ignored), higher makes it stronger. You can also connect a curve so a reference is strong at the start and eases off later (handy when you want it to set the mood without taking over the whole image). Optional — only add it when you want that extra control.
• Better multi-reference handling — extra reference images of a different size no longer drift or overlap; they line up cleanly with your main image now.
• New Resize toggle (Presampling) — on by default, the node picks the output size for you from Aspect Ratio and Megapixels (as before). Turn it off to let the node leave every image at its own size and work at your input image's native size — handy when you'd rather control sizing yourself with other nodes. Turning it off tidies the node by hiding Aspect Ratio and Megapixels.

<img width="2306" height="1195" alt="image" src="https://github.com/user-attachments/assets/83fdc302-073c-4d3c-afc1-6555bb7d949a" />

What's new in 1.8.x

• New Match Original Colors (Postsampling) — pulls the regenerated area's colors and lighting back toward the original image. The model often shifts the overall white balance or saturation slightly; this dial corrects that drift so the edit blends into the same scene. 0 leaves things as the model produced them; 1 matches them fully to the original.
• New Seamless Edges (Postsampling) — erases any remaining color or lighting seam at the boundary of the regenerated zone. Best for dramatic relighting or strong style changes where the edge is still visible after color matching. Off by default — it's heavier and can smear textured edges.
• New Auto-Detect Edit Region (Postsampling) — when you run an img2img edit without a mask, the node figures out which pixels actually changed and composites only those back. Keeps the rest of the image pixel-perfect across iterative edits instead of letting the model rewrite the whole canvas every time. Comes with its own fine-tuning controls:
  • Edge Softness — how gently the detected region fades into the original.
  • Region Padding — grow or shrink the detected region before blending.
  • Fill Inner Gaps — seal small holes inside the detected region.
  • Extend To Borders — extrapolate to the image edge so no frame of the original peeks through.
• Example workflow included — a ready-to-use workflow lives in example_workflows/ (with a preview image). Drag it into ComfyUI to get a working setup immediately.

What's new in 1.7.x

• Megapixels is now a slider with decimal precision (0.1 – 4.0) instead of a dropdown with fixed steps. You can pick any size that suits your needs.
• New Image Fit control — decide how the input image should be handled when the canvas you chose has a different shape than the image:
  • Native (default): the model rebuilds the canvas around your subject without distorting it. Best for changing aspect ratio or tile-based workflows.
  • Center Crop: cuts the image to fit the canvas (centered, no distortion, loses the edges).
  • Outpaint: fits the whole image inside the canvas and lets the model fill in the surrounding space.
• New Outpaint Fill — when using Outpaint, choose what goes in the empty space: Gray (neutral, default), Black, White, or Smart (a soft continuation of your image so the model has a natural starting point).
• New Slide control — shift the image off-centre instead of always centering it. With Outpaint it moves the image within the empty space; with Center Crop it chooses which part of the image is kept. 0.5 stays centered; the direction follows the canvas shape (a taller canvas slides it up/down, a wider canvas slides it left/right).
• New ref_0 output — your input image after the Image Fit / Outpaint preprocessing, at the final canvas size. Reuse it anywhere else in your workflow.

⚠️ Heads up if you're upgrading from an older version: the Megapixels widget changed from a dropdown to a numeric slider. Workflows saved with the old version will load fine — the value is migrated automatically and a notification will let you know — but it's a good idea to open the node and double-check the value is what you want.

---

What you can do with it

• Generate images from a prompt — just connect the model and write what you want.
• Transform an existing image — drop your image into the reference slot and add a prompt describing the change.
• Inpaint a specific area — paint a mask and the masked zone is the only part that gets regenerated. Everything else stays untouched.
• Detail a small area at high quality — turn on detailing to zoom into the masked zone (a face, a hand, an eye) and regenerate it with way more detail than you'd get from a full-image pass.
• Combine multiple reference images — extra slots appear as you connect them, so you can give the model several visual hints to work with.

The node figures out which mode you're using based on what you connect — no setting to flip.

---

The nodes

😺NKD Klein Presampling

The starting point. You connect your model, prompts, and reference image here. It hands off everything the sampler needs.

😺NKD Klein Postsampling

The end point. It takes the sampler's output and delivers the final image, putting everything back in its place when you've used inpainting or detailing.

😺NKD Klein Reference Weight (optional)

Sits between Presampling and your sampler. Lets you dial how much a single reference image shows up in the result, one reference at a time. Skip it if you don't need that control. See Reference Weight below.

😺NKD Klein Prompt Builder (optional)

Assembles a prompt from your own text plus curated preset dropdowns, with a live preview, and outputs a string you connect to Presampling's positive input. Choose flowing prose (best for Klein) or a structured JSON template. The dropdown presets live in klein_presets.json — edit that file (then restart ComfyUI) to customise them.

The chain looks like this:

NKD Klein Presampling → (NKD Klein Reference Weight) → [your sampler chain] → NKD Klein Postsampling

---

Inputs that matter

Image and prompt

• ref_0 — your input image. The most important slot.
• ref_1, ref_2, … — extra reference images. Slots appear automatically when you connect one, up to 8.

Output size

• Resize — on by default: the node chooses the output size from Aspect Ratio and Megapixels. Turn it off to keep every image at its own size and work at your input's native size (the node won't resize anything — handle sizing yourself elsewhere). With Resize off, Aspect Ratio and Megapixels are hidden since they no longer apply.
• Aspect Ratio — the shape of the final image. As Reference matches your input image; Custom lets you set any size; or pick one of the named ratios.
• Megapixels — how big the final image should be. Bigger = more detail, but slower.
• Custom Width / Custom Height — only used when Aspect Ratio is set to Custom.

Inpainting (when you connect a mask)

• mask — paint white over the area you want regenerated; black stays as-is.
• Mask Expand — makes the painted area a bit bigger so the new content blends naturally with its surroundings.
• Mask Blur — softens the mask edges for a smoother transition.
• Inpaint Blend — 0 = clean cut along the mask edge, higher values fade the new and old together.

Detailing — for high-quality touch-ups

• Use Detailing — turn this on to zoom into the masked area before regenerating it. Perfect for fixing faces, hands, eyes, small props. The result is composed back into the full image automatically.
• Detail Padding — how much of the surrounding area to include in the zoom. More padding = the model sees more context; less padding = tighter focus on the masked zone.

Reference Strength — the creative dial

This is the dial that controls how much creative freedom the model has versus how strictly it sticks to your reference image's layout.

| Value | Behaviour | When to use | |---|---|---| | -3 to -1 | ⚠️ Mostly experimental. More freedom, looser interpretation | When you want bigger, more imaginative changes — the model reinterprets things more | | 0 (default) | Klein's official default behaviour — balanced | Good for most edits | | 1 to 4 | Mild anchor | When you want changes but the layout to stay close to the original | | 5 to 7 | Strong anchor | Tight alignment with the reference — useful for upscales or precise edits | | 8 to 10 | Almost locked to reference | When things absolutely have to line up pixel-for-pixel with the input |

If you're getting unwanted drift between the original and the result (faces shifting, edges not quite aligning), bumping this up usually fixes it. If your generations feel too restrained or "stuck" close to the input, try negative values for more creative leeway.

Reference Strength vs Reference Weight — what's the difference? Reference Strength (above) decides how closely the result follows the layout of your reference. Reference Weight (the separate node below) decides how much a reference shows up in the image. Two different dials — you can keep a loose layout but a strong look, or the other way round.

Reference Weight — per-reference influence

The 😺NKD Klein Reference Weight node lets you turn one reference image up or down without touching the others. Add it on the model line between Presampling and your sampler. Want to control two references separately? Just chain two of these nodes.

• reference_index — which reference this node affects. 0 is your main image (ref_0), 1 is ref_1, and so on, in the order you connected them.
• Reference Weight — 1.0 leaves the reference as usual. Below 1 fades it out (0 = ignored completely); above 1 makes it show up more (up to 2).
• schedule (optional) — connect the FLOAT output of NKD Sigmas Curve to change the weight over the course of the generation. For example, a curve that starts high and drops to zero lets a reference set the overall look at the start, then step aside so it doesn't take over the fine details. Leave it unconnected to keep the same weight throughout. It behaves the same whichever scheduler you use.

Postsampling clean-up

These live on the Postsampling node and fine-tune how the regenerated area is composited back over the original.

• Uncrop Feather — softens the edge where a detailed zoom blends back into the rest of the image.
• Match Original Colors — pulls the new area's colors and lighting back toward the original to undo any white-balance or saturation drift from the model. 0 = no correction, 1 = full match. The reference is read only from the unchanged background, so the edit itself keeps its intended colors.
• Seamless Edges — turns on Poisson blending to erase any remaining color/lighting seam at the edge of the regenerated zone. Turn on for dramatic relighting or strong style changes; leave off by default (it's heavier and can smear textured edges).
• Auto-Detect Edit Region (img2img without a mask) — detects what actually changed between the input and the generated image and composites only that region back. Keeps the unchanged parts of the original pixel-perfect across iterative edits. Ignored when there's no input image, or when a mask is connected.
• Edge Softness — how softly the auto-detected region fades into the original (% of image diagonal, so it looks the same at any resolution). Higher = wider, gentler transition; lower = tight, geometric edits.
• Region Padding — grow (positive) or shrink (negative) the auto-detected region before blending. Useful when detection falls just short of the true edges, or bleeds into background that should stay untouched.
• Fill Inner Gaps — seals small holes inside the auto-detected region, for when the edit changed an object's color but not its interior contrast.
• Extend To Borders — extrapolates the detected region into any thin border void left by alignment, so no frame of the original peeks through. Leave on.

Other

• Pin Model — keeps the model loaded in your graphics card so it doesn't reload between runs. Faster, but only turn it on if you have plenty of VRAM.
• Bypass Reference — turns off the model's ability to look at your reference image, so it behaves like a traditional image-to-image model instead. Leave it off in most cases.

---

Modes (auto-detected)

| What you connect | Mode | |---|---| | No reference image | Text-to-image | | Reference image only | Image-to-image | | Reference image + mask | Inpainting | | Reference image + mask + Use Detailing on | Inpainting with detail zoom |

---

Quick recipes

Generate an image from scratch

1. Write a positive prompt.
2. Pick an Aspect Ratio and Megapixels.
3. Run.

Transform an image

1. Drop your image into ref_0.
2. Write a prompt describing what you want changed.
3. Run.

Edit a specific area of an image

1. Drop your image into ref_0.
2. Paint a mask over the area you want to change and conect it to the mask socket.
3. Write a prompt describing the change.
4. Run.

Fix a face, hand or small detail at high quality

1. Drop your image into ref_0.
2. Paint a mask over the small area and conect it to the mask socket.
3. Turn on Use Detailing.
4. Write a prompt describing what you want the detail to look like.
5. Run.

Use multiple reference images for inspiration

1. Drop your main image into ref_0.
2. As you connect, more slots will appear — add up to 8 reference images.
3. Write a prompt and run.

Make one reference show up more (or less) than the others

1. Set up your references as above.
2. Add a NKD Klein Reference Weight node on the model line, between Presampling and your sampler.
3. Set reference_index to the reference you want to adjust (0 = ref_0, 1 = ref_1, …).
4. Turn Reference Weight up (toward 2) to make it show up more, or down (toward 0) to fade it out.
5. (Optional) connect an NKD Sigmas Curve to its schedule input to make it strong early and fade later.
6. Run. Add another Reference Weight node if you want to adjust a second reference too.

Iteratively edit an image without a mask (keep the rest pixel-perfect)

1. Drop your image into ref_0.
2. Write a prompt describing the change ("make the sky stormy", "swap the shirt for a denim jacket", …).
3. On the Postsampling node, turn on Auto-Detect Edit Region.
4. Run. Only the changed region is composited back; the rest of the image stays identical, so you can chain several edits without drift.

---

Example workflow

A ready-to-use workflow is bundled in example_workflows/. Drag NKD Klein Tools.json straight into ComfyUI to load it, or check the preview image first.

---

Typical workflow

flowchart LR
    ML([Model Loader]):::input --> PRE
    PR([Prompts]):::input --> PRE
    LI([Load Image]):::input -- ref_0 --> PRE
    MP([Mask Painter]):::input -- mask --> PRE

    PRE["**NKD Klein**<br/>**Presampling**"]:::nkd

    PRE -- "model · positive<br/>negative · latent" --> SAMP([Your Sampler]):::external
    PRE -- "ref_0<br/>(preprocessed)" --> REUSE([reuse anywhere<br/>in your graph]):::output
    PRE -- bundle --> POST

    SAMP --> VAE([VAE Decode]):::external
    VAE -- image --> POST["**NKD Klein**<br/>**Postsampling**"]:::nkd

    POST --> OUT([final image]):::output

    classDef nkd fill:#3b3b6b,stroke:#8ab4ff,stroke-width:2px,color:#fff
    classDef input fill:#2d2d2d,stroke:#888,color:#eee
    classDef external fill:#2d2d2d,stroke:#888,color:#eee
    classDef output fill:#1f4a1f,stroke:#7fd97f,color:#fff

The bundle output carries everything Postsampling needs to put the image back where it belongs (crop boxes, masks, the original reference, mode…). You don't need to look inside it — just connect it straight through.

---

Requirements

• ComfyUI with Flux Klein model support
• PyTorch ≥ 2.0

---

Installation

Clone into your ComfyUI/custom_nodes folder:

git clone https://github.com/Nekodificador/ComfyUI-NKD-Klein-Tools

Or install via the ComfyUI Manager by searching for NKD Klein Tools.

---

Made by Nekodificador