Z-Image Turbo Inpainting: Complete Guide to 3 Approaches (Native Nodes / KJNodes / SAM3)

---

"I just want to remove the stranger from my photo — not regenerate the whole image!" This is the most common scenario for Z-Image Turbo users. The official team has promised a Z-Image-Edit variant and a dedicated Inpainting ControlNet, but neither has shipped yet. The good news: Z-Image-Turbo by itself, combined with ComfyUI, already delivers solid inpainting results. Popular community tutorials (NerdyRodent / Prompting Pixels) have distilled three mature routes — pure native nodes, KJNodes enhancement, and SAM3 smart segmentation. This article breaks down the node wiring, use cases, and selection advice for all three.

1. Inpainting Basics: Masks, Latents, and Turbo's Quirks

How Inpainting Works

Inpainting (local redraw) is about telling the model: "only change this region, keep everything else." In ComfyUI, that "region" is a mask, and "only this part" is enforced through three mechanisms:

1. VAE Encode (for Inpainting): When encoding the original image into latent space, noise is injected into the masked region while the rest keeps its original latent information
2. InpaintModelConditioning: Injects mask info into the sampler so the model knows which areas to regenerate
3. Differential Diffusion (optional): Replaces hard mask boundaries with progressive mask weights for more natural edge blending

Z-Image Turbo Specifics

• 8-step distilled model: Use FlowMatchEulerDiscreteScheduler, 8 steps is enough, no CFG needed (CFG 1.0)
• The three-piece model set: Z-Image-Turbo checkpoint (fp16 or GGUF) + Qwen3-4B text encoder + flux_ae VAE
• Native resolution: 1024×1024 is optimal; keep the redraw region at original resolution

2. Option 1: Pure Native Nodes (Zero Dependencies)

The most basic, dependency-free approach — only built-in ComfyUI nodes, no custom node packs required.

Node Wiring

Load Image → Right-click → Open in Mask Editor (paint mask) → Save
    ├── → VAE Encode (for Inpainting) → InpaintModelConditioning → KSampler → VAEDecode → Preview
    └── → (mask output) ───────────────────────────↗

Steps

1. Load image: Load the target image with the Load Image node
2. Paint the mask: Right-click Load Image → Open in Mask Editor, brush over the region to redraw (object, person, text, etc.), click Save
3. Encode: VAE Encode (for Inpainting) takes image + mask, outputs noise-injected latents
4. Inject conditioning: InpaintModelConditioning takes latents + positive/negative prompts
5. Sample: KSampler, 8 steps, CFG 1.0, sampler FlowMatchEulerDiscrete
6. Decode: VAEDecode → PreviewImage

Pros & Cons

Pros	Cons
Zero dependencies, works in any ComfyUI out of the box	Mask edges can be harsh — a "patchwork" look
Fewest nodes, easy to learn and debug	Hand-painting masks is tedious; complex outlines are hard to draw precisely
Great for small fixes (blemishes, small objects)	Large-area redraws blend less well with surrounding light/shadow

3. Option 2: KJNodes Enhancement (GrowMaskWithBlur)

The harsh mask edges from Option 1 can be dramatically improved with GrowMaskWithBlur from the KJNodes custom node pack. It expands the mask by a few pixels and applies blur, letting the redrawn area blend smoothly with its surroundings.

Node Wiring

Load Image → Right-click → Open in Mask Editor → Save
    ├── → GrowMaskWithBlur (KJNodes) → VAE Encode (for Inpainting) → InpaintModelConditioning → KSampler → VAEDecode → Preview
    └── → (mask) ────────────────────────────↗

GrowMaskWithBlur Parameters

Parameter	Recommended	Notes
grow	8 - 16 px	Mask expansion; larger covers more area
blur	8 - 16 px	Edge blur radius; larger = softer transition
iterations	1 - 2	Usually 1 is enough

Advanced: Differential Diffusion

To kill the "patchwork" look entirely, add the Differential Diffusion node (from KJNodes or the Fooocus Inpaint wrapper). It uses grayscale mask weights instead of a binary mask, making the model "half-commit" near edges for much more natural blending.

Best Use Cases

• Medium-area edits: outfit changes, watermark removal, background element swaps
• Outputs where edge quality matters
• Combined with Inpaint Crop & Stitch (crop the region, sample separately, stitch back) to slash VRAM usage on large images

4. Option 3: SAM3 Smart Segmentation (Text Prompt / Point Click)

Tired of hand-painting masks? SAM3 (Segment Anything Model 3) — Meta's latest segmentation model — can isolate any object from a single sentence description. The ComfyUI ecosystem offers two routes: ComfyUI-SAM3 (community node pack by wouterverweirder) or ComfyUI's native SAM 3.1 nodes (built into recent versions).

Installation

1. Install ComfyUI-SAM3 via ComfyUI Manager
2. Download sam3.pt and place it in ComfyUI/models/sam3/
3. Restart ComfyUI

Method A: Text-Prompt Segmentation (SAM3Grounding)

Load Image → SAM3 Segmentation (prompt: "person", "car"...) → mask
    → GrowMaskWithBlur → VAE Encode (for Inpainting) → InpaintModelConditioning → KSampler → VAEDecode

Node parameters:

Parameter	Description
prompt	Text description of the object to segment, ≤32 tokens; separate multiple subjects with commas + :N for counts, e.g. eye:2, window panels:4
threshold	Confidence threshold, default 0.5; raise it if you get false positives
min_width / min_height	Minimum object size (pixels), filters small noise

Method B: Point-Click Segmentation (SAM3PointCollector)

When text isn't precise enough (e.g., "the third person"), use the SAM3 Point Collector node to click directly on the target in the image — the model generates a precise mask from point prompts. Ideal for: picking one specific object among similar ones, or complex outlines.

Best Use Cases

• Precise person/object extraction for redraws (background swap, outfit change, removing bystanders)
• Per-frame video segmentation (SAM3 supports video tracking — video inpainting)
• Point-click fallback when prompts can't describe what you mean

5. Comparison & Selection Guide

Dimension	Option 1: Native	Option 2: KJNodes	Option 3: SAM3
Dependencies	None	KJNodes pack	ComfyUI-SAM3 + sam3.pt
Mask creation	Hand-painted	Hand-painted + auto grow/blur	Text prompt / point click
Edge quality	Fair	Good	Good
Automation	Low	Medium	High
Learning curve	⭐ Easiest	⭐⭐	⭐⭐⭐
Best for	Quick small fixes	Medium areas, blend quality	Precise segmentation, batch work

Rule of thumb:

• Quick blemish fix → Option 1
• Outfit change / watermark removal / background element swap → Option 2
• Precise person/object extraction, complex outlines → Option 3
• Maximum quality → SAM3 mask + Option 2's GrowMaskWithBlur + Differential Diffusion

6. Advanced Techniques

1. Inpaint Crop & Stitch (VRAM-friendly large images)

ComfyUI-Inpaint-CropAndStitchCustom crops the redraw region, samples it separately, and stitches it back seamlessly. For 2K/4K images, VRAM usage drops from "full-image sampling" to "local sampling" — runs fine on an RTX 4060.

2. ControlNet Union 2.1 (near-official redraw)

In December 2025, the community released the Z-Image-Turbo Fun Controlnet Union 2.1 model patch (HuggingFace). Combined with ComfyUI's new node, it enables more advanced inpainting — the masked region follows extra control conditions (lineart, depth, pose), perfect for "redraw while preserving structure." NerdyRodent published a matching free workflow.

3. Outpainting (expansion)

The ConditioningZeroOut node clears conditioning outside the mask; combined with a suitable mask you can expand outward — turning 1024×1024 into a 1536×1024 panorama composition.

4. Complete Workflow Template (25-node community solution)

The high-rated community workflow (Prompting Pixels, 78 downloads) uses three models:

• z-image-turbo-fp8-e4m3fn.safetensors (UNETLoader)
• qwen_3_4b.safetensors (CLIPLoader)
• ae.safetensors (VAELoader, flux_ae)

Core node chain: LoadImage → SAM3 (Grounding/PointCollector/Segmentation) → ImageToMask → InpaintCropImproved → GrowMaskWithBlur → VAE Encode (for Inpainting) → InpaintModelConditioning → KSampler → VAEDecode → InpaintStitchImproved → Image Comparer.

Conclusion

Z-Image-Turbo has no official inpainting model yet, but three community routes are mature: native nodes for quick fixes, KJNodes for edge blending, and SAM3 for "describe-it-don't-draw-it" masking. When Z-Image-Omni-Base and Z-Image-Edit finally ship, these workflows will migrate seamlessly to official editing capabilities — but until then, mastering these three approaches means your Z-Image Turbo already handles 90% of local redraw needs.