Generate the base image, detect the face region, run a controlled detail pass, then upscale the finished image. Compare the original on the left and the refined image on the right before deciding whether the change helps.
Watch on YouTube · Video published 2026-08-31 · Companion reviewed 2026-09-23
The demonstrated workflow
The August 2026 lesson uses SDXL with RealVisXL 5.0, a cinematic-lighting LoRA, a face_yolov8m detector and a RealESRGAN ×2 upscale. Face Detailer is a local regeneration step: it can change facial features as well as improve small details. It is not an identity guarantee or a universal fix for an unsuitable base image.
Install the graph’s required nodes, including Impact Pack, its Impact Subpack where required for detector loading, and Doss Node Suite for the supplied controls. Use the dependency list packaged with the release and match model licenses to your intended use.
Work through the graph
- Load the compatible checkpoint and LoRA, then select the detector and upscale model. Missing model names must be resolved in their actual loaders, not by renaming an unrelated file.
- Generate a base image at 896 × 1152, batch size 1. The demonstrated KSampler uses 30 steps, CFG 5, dpmpp_sde, karras and denoise 1.
- Pass the decoded image into Face Detailer and inspect the detector region. Keep the detail pass modest before raising denoise.
- Compare the face with the base image at the same viewing scale. Decide whether eyes, mouth, lighting and identity improved together.
- Apply the ×2 upscale to the selected result. In the shown setup, the final dimensions are 1792 × 2304. Save the base and final images so you can evaluate the whole pipeline.
Settings shown in this lesson
| Control | Demonstrated value |
|---|---|
| Guide size / maximum size | 512 / 1024 |
| Detail steps / CFG | 20 / 5 |
| Detail sampler / scheduler | Euler / simple |
| Detail denoise | 0.40 |
| Detector threshold | 0.50 |
| Bounding-box dilation / crop factor | 10 / 3 |
| Detail cycles | 1 |
These values describe the published SDXL graph. They are not recommendations for every checkpoint, face size or model architecture. SAM fields do not influence the result when the SAM branch is not connected. The package notes should identify which optional connections are active.
Tune one problem at a time
- No face detected: inspect the source image, selected detector and detection threshold before changing generation settings.
- A different-looking person: reduce denoise and simplify the detail prompt. Compare the eyes, jaw and mouth against the base rather than judging sharpness alone.
- Visible seam or lighting mismatch: inspect the crop, mask feathering and surrounding context. A larger refined area may introduce more unwanted change.
- Plastic or overprocessed texture: reduce the strength of the detail or upscale stage. The most detailed output is not always the most believable.
Prompt practice
Suggested detail prompt — use with the compatible SDXL graph
natural facial proportions, clear eyes, subtle skin texture, soft window lighting, consistent expressionUse a short description that supports the source image. Do not ask for a different age, expression or lighting direction unless that change is intentional. Any negative prompt must be supported by the chosen model and conditioning path.
App view and version notes
The app view shown in the lesson is another interface to the same graph, with ten controls and two outputs. It is not a separate hosted service. The published pack documented ComfyUI 0.34.x and frontend 1.49.6 as its historical environment. Later updates require a compatibility check; the article does not claim a fresh GPU run on every version.
Sources and downloads
- Existing Face Detailer Patreon member post
- Face Detailer one-time workflow pack
- Impact Pack project
- Published video and model links