"Looks AI-generated" is not one problem. It is at least five: plastic skin, over-smooth gradients, invented detail where the source had none, wrong noise statistics, and a specific kind of symmetrical perfection that real photographs never have.
This episode attacks each separately with a five-in-one workflow set.
The three graphs
| # | Workflow | Mode |
|---|---|---|
| 01 | 4K single image, LLM auto detail enhancement | Interactive |
| 02 | 4K online ZIP batch, fully automatic | RunningHub batch |
| 03 | 4K local folder batch, GGUF vision reverse-prompt | Local batch |
01 and 03 are the same operation at different scales. 02 is the one you use when you have a thousand images and no local GPU.
The five fixes, and what each one actually changes
- De-AI the skin. Local frequency separation — high-frequency grain restored, low-frequency plastic sheen removed. This is the single biggest contributor to the "AI look".
- Faithful redraw. GGUF vision model reverse-prompts the source, then the image is regenerated from that description. Detail that the source did not contain is not invented.
- Local retouch. Mask a region and re-render only that region at full resolution, so a retouched face does not soften the rest of the frame.
- Quality comparison. Side-by-side at 100% crop, because thumbnail comparison hides every artifact that matters.
- 4K batch. The whole chain applied to a folder, with the same parameters, without manual intervention.
The LLM auto-enhancement in graph 01 is doing something specific: it inspects the source and decides how much detail enhancement each region needs, rather than applying one global setting. That is why the batch modes use fixed parameters — the automatic decision does not survive scaling to a folder.
What honest output looks like
A successfully "de-AIed" image still does not look like a photograph taken on set. It looks like a photograph taken on set that has been lightly retouched — which is the actual goal. Expecting a model to reconstruct real sensor noise is the wrong expectation; expecting it to produce plausible noise is the right one.
Files
- HF: https://huggingface.co/AI-KSK/flux3-klein-4k-reconstruction
- ModelScope: https://modelscope.cn/models/Sunjian520/flux3-klein-4k-reconstruction
Note: the tutorial's own YouTube reference for this episode is only partially aligned with the three graphs — the Bilibili version documents the full chain.
Notes
- Three ComfyUI workflow exports; FLUX.3 Klein and the GGUF vision model must exist locally.
- Tutorial material, not a product.
- Independent AI-KSK release; not affiliated with the model authors or RunningHub.
