Z-Image Base Inpainting, Upscaling, and an Encoder Type That Keeps Changing
Three Z-Image Base workflows: inpainting, upscaling, and a Qwen3-VL-driven image cleanup variant. The interesting part is again the loader — the same encoder file carries two different type strings across three graphs.
Three Z-Image Base workflows: inpainting, upscaling, and a Qwen3-VL-driven image cleanup variant. The interesting part is again the loader — the same encoder file carries two different type strings across three graphs.
The graphs
| Workflow | Nodes | CLIP type |
|---|---|---|
| Z-Image Base inpainting | 14 | qwen_image |
| Z-Image Base upscale | 26 | lumina2 |
| Z-Image Base wash + upscale | 49 | lumina2 |
Same encoder (qwen_3_4b.safetensors), same UNET (z_image_bf16.safetensors), same VAE (ae.safetensors). Two type strings.
This has now appeared in three separate Z-Image releases in this series. It is not a one-off typo — it is a convention that was never settled.
The practical rule: copy the loader node from the graph you are extending. If you build a new graph, check which type the source graph used.
The 14-node inpainting graph
Fourteen nodes is close to the minimum for a working inpaint: load model, load image, load mask, encode, sample, decode, paste, save.
Inpainting is the most useful workflow in any image toolkit, and getting it to 14 nodes is a genuine achievement — most implementations carry 30+ nodes of masking and conditioning plumbing.
The Qwen3-VL wash workflow
洗图 (image washing) with Qwen3-VL: use a vision-language model to describe an image, then regenerate it from the description.
Why that produces something useful: a VLM description captures what the image is in semantic terms — subject, composition, lighting, style words — and drops the noise, compression artifacts, and incidental detail. Regenerating from that description produces a cleaner, more coherent image.
This is a different operation from upscaling. Upscaling adds detail; washing replaces detail with a regenerated interpretation. For a low-quality source you want clean output rather than faithful reproduction, washing is the right tool.
The 49-node graph is the largest in the release because it chains two model families: Qwen3-VL to describe, Z-Image to regenerate.
The upscale to near-4K
Z Image base放大 at 26 nodes with block/tiled upscaling.
Tiling matters for a 4K target: one allocation for a 4K image plus its latents is large, and tiling keeps the peak manageable. It also means the upscaler sees local context, which can help or hurt depending on content.
Where this breaks
- The encoder type inconsistency across three graphs.
- Washing loses fidelity deliberately. If the source detail matters, washing is the wrong choice.
- VLM descriptions miss specifics. A description may say "a person in a red jacket" and lose the exact jacket design.
- Tiled upscaling shows tile seams on images with smooth gradients.
- 14-node inpainting does not handle complex masks or multi-region edits.
Practical order
Inpainting first — it is the smallest graph and the most frequently needed operation. Then upscale. Washing last, because it changes the image rather than improving it.
Workflows
Three graphs, published on RunningHub with invite code rh-v1111:
- Z-Image Base inpainting
- Z-Image Base upscale
- Z-Image Base Qwen3-VL wash + upscale
Files
https://huggingface.co/AI-KSK/ksk-video-168-comfyui-workflows
Notes
- Workflow files only. Model weights are not included; each keeps its own licence.
- Tutorial exports, not a product.
- Node counts and CLIP types quoted above are read from the published JSON files.
- Independent AI-KSK release. Not affiliated with Tongyi or RunningHub.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.