This workflow is specifically designed to assist with preparing portrait datasets for stylistic Lora training. Prepare an image dataset in one or more directories and then point this workflow at them. It will automatically upscale, colorise, restore etc. the images and extract a normally rotated square 1024x1024 face image from each source image into an output directory you specify. The workflow utilises some ComfyUI custom nodes available on my github page, or published through ComfyUI. Additional details... # Portrait Dataset Gate — self-driving LoRA dataset prepPoint it at a folder of source photos. For each image it decides what thatimage actually needs, does only that, and names the output so you can see whathappened. Built for face-centric LoRA training sets.**Workflow:** <code>PORTRAIT_DATASET_GATE_v1.json</code>---## What it does, per image```source ─▶ score ──reject (too small, or not a photograph) ──▶ gate_rejected/│├─▶ too soft/small? ──▶ SeedVR2 super-res ──▶ rescore (up to 2 passes)│▼ acceptedhead crop (derotated to eye level, square, +15% padding)│├─▶ black &amp; white or sepia? ──▶ DDColor colourise → tag _col├─▶ single-hue colour cast? ──▶ skin-anchored correction├─▶ skin texture below 1.0? ──▶ SUPIR restoration → tag _sup├─▶ face drifted? ──▶ revert to the crop → tag _noid└─▶ still marginal? → tag _rev▼train/&lt;original stem&gt;[_col][_sup][_noid][_rev].png```**The filename is the report.** No tag means it was clean out of the box.<code>_col</code> means every colour is invented — check those, especially hair.<code>_sup</code> means restoration ran. <code>_sup_rev</code> means it ran and the result is stillmarginal, so look at it. You only eyeball what the workflow couldn't settle.Every scoring also appends a row to <code>quality_gate.csv</code> with the metrics and thereason, so you can tune thresholds against what you actually see.---## Setup### Required custom nodes- **[ComfyUI-PortraitDatasetTools](<a target="_blank" rel="ugc" href="https://github.com/BillyBGit/ComfyUI-PortraitDatasetTools)**">https://github.com/BillyBGit/ComfyUI-PortraitDatasetTools)**</a> — the measurement and routing nodes- [ComfyUI-SeedVR2_VideoUpscaler](<a target="_blank" rel="ugc" href="https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler">https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler</a>) — super-resolution- [ComfyUI-DDColor](<a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-DDColor">https://github.com/kijai/ComfyUI-DDColor</a>) — B&amp;W colourisation- [ComfyUI-KJNodes](<a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">https://github.com/kijai/ComfyUI-KJNodes</a>) — <code>ImageResizeKJv2</code>- [ComfyUI-Custom-Scripts](<a target="_blank" rel="ugc" href="https://github.com/pythongosssss/ComfyUI-Custom-Scripts">https://github.com/pythongosssss/ComfyUI-Custom-Scripts</a>) — <code>ShowText</code>SUPIR is **built into ComfyUI core** <code>SUPIRApply</code> + <code>ModelPatchLoader</code>) — nowrapper needed.### Models| put in | file ||---|---|| <code>models/model_patches/</code> | <code>SUPIR-v0Q_fp16.safetensors</code> — [Kijai/SUPIR_pruned](<a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/SUPIR_pruned">https://huggingface.co/Kijai/SUPIR_pruned</a>) (2.7 GB) || <code>models/checkpoints/</code> | any photographic SDXL checkpoint (SUPIR is an SDXL pipeline) || SeedVR2's folder | <code>seedvr2_ema_7b_sharp_fp16</code> + <code>ema_vae_fp16</code> || <code>models/upscale_models/</code> | a face upscaler such as <code>4xFaceUpSharpDAT</code> || DDColor's <code>checkpoints/</code> | <code>ddcolor_modelscope.pth</code> |### Running1. Set node *<code>[2]</code>** to your source folder — **one directory per line**. Everyoutput path is relative to it, so you can paste in fifty model folders andeach one's outputs land in its own place.2. Queue **once per image** using the **batch-count** field.3. Start ComfyUI **without <code>--cache-none</code>** — it breaks the lazy branches.That's the only path you need to set.---## TuningDefaults came from surveying a ~400-image scraped dataset. Survey your own setbefore trusting them; a corpus of studio portraits sits very differently fromscraped material.| widget | default | meaning ||---|---|---|| <code>min_face_px</code> | 500 | face bbox short edge, **native** pixels || <code>hard_floor_face_px</code> | 200 | below this, reject without attempting super-res || <code>min_clarity</code> | 60 | ~20 soft, ~60 acceptable, 150+ crisp || <code>min_source_grain</code> | 0.15 | below = no sensor noise at all: illustration, airbrushed, or AI || texture gate | 1.0 | below this, run restoration || <code>target_b</code> (skin tone) | 14 | **dataset-specific.** ~14 warm-toned, ~9 East Asian |**Forcing individual images through:** the <code>FORCE</code> primitive <code>accept</code> /<code>refine_1x</code> / <code>refine_2x</code>) ignores every threshold. Put the stragglers in theirown folder, point <code>[2]</code> at it, set it, queue, then set it back to <code>off</code>.---## Findings worth knowing**Restoration is capped by the VAE, not the model.** A pure VAE encode→decodewith no sampling at all loses as much skin texture as any img2img pass at1024. Chroma, Flux.1, Flux.2, SRPO and Qwen-Image-Edit were all tested; allreturned the input or worse. **SUPIR is the exception** — it carries apurpose-trained denoise encoder that replaces the VAE encoder for its hintlatent, so its conditioning skips that low-pass. Measured 0.63 → 0.87 textureat ~20 s/image with identity preserved at 0.94.**SUPIR's denoise is a peak, not a slider.** 0.65 is best; 0.75 collapses backbelow the input. Raising it does not give you more.**SUPIR converges rather than simply improving.** It lifts low texture andpulls down high texture, so running it on everything would damage your bestimages. Hence the 1.0 gate.**A second restoration pass makes things worse** (0.45 → 0.36), and retrying ata different denoise doesn't rescue a failure — the peak doesn't move per image.There is deliberately no loopback.**Over-processed output usually means the source had no real texture.** Alreadyairbrushed exports, illustrations and heavily compressed scans have nothing forsuper-res to recover, so it synthesises wax. That's what <code>min_source_grain</code>catches before any GPU time is spent.**For anything the pipeline can't lift: do nothing.** A soft but genuinephotograph is better training data than invented texture — and the <code>_rev</code> tagalready tells you which ones to look at.---## LicenceWorkflow: MIT. The models it loads carry their own licences.

PortraitDatasetTools

While this workflow uses SDXL for processing (used in conjunction with SUPIR), it is intended to prepare image datasets for lora training - which could be used to train any model.

Files with SHA256: 1a9f20f2e819c6785525df2bb408c295bab5323a8bf17438485776c6e26e8435