A ComfyUI consistent character workflow needs a stable definition of the character and a repeatable way to judge the output. Without both, a series can drift into several similar-looking people while each individual image still looks attractive.
Define identity separately from styling
Create a short identity sheet for an original adult character: face shape, eye spacing, hairline, distinguishing features and general proportions. Keep wardrobe, location and pose in a separate scene brief. A new coat should not imply a new face.
Choose one clear neutral portrait as the primary comparison reference. Add a few approved views only when they actually represent the same identity. Feeding a collection of inconsistent generated faces into a workflow does not create a clear target.
Choose a method supported by your model
Some workflows use reference-image conditioning; others use a trained character adapter. A LoRA needs a compatible base model and loader, as described in the official ComfyUI example. Follow our LoRA setup and comparison guide for the surrounding production checks. Plain image-to-image can preserve some input structure, but that alone does not establish reliable identity across large scene changes.
Start with the method’s maintained example. Record its model family, reference requirements and extension dependencies. Do not combine instructions from unrelated model architectures because their node names sound similar.
Use a five-scene acceptance sheet
| Scene | What it tests |
|---|---|
| Neutral chest-up portrait | Baseline facial structure |
| Three-quarter view | Identity under a modest angle change |
| Outdoor waist-up portrait | Lighting and body proportions |
| Seated indoor scene | Posture, clothing and hands |
| Simple product scene | Identity alongside an additional visual requirement |
Keep the face large enough to judge in every test. A tiny person in a wide landscape is not convincing evidence of facial consistency. Compare candidates beside the approved reference at similar sizes.
Review more than the face
Check age appearance, skin texture, body proportions, hairstyle and distinctive accessories. Then inspect scene-specific defects separately. An image can preserve identity and still fail because the hand holding a cup is unusable.
Use three labels: accepted, repairable and rejected. Add a short reason such as “jaw changed,” “correct identity, sleeve repair needed,” or “wrong eye spacing.” Those notes are more actionable than a vague quality score.
Change one source of variation at a time
Suppose the character passes a studio portrait but changes in outdoor scenes. Keep the reference, model and pose fixed while adjusting the scene description. If that helps, test the pose next. Changing the adapter strength, reference set, model and prompt together prevents you from learning which factor mattered.
A fixed seed can help make a comparison more controlled, but it is not a portable identity token across arbitrary models and workflows. Save the full graph and dependency record, not only the seed.
Carry the identity into video carefully
Start from an approved still and ask for one restrained motion. Watch the entire result rather than judging the thumbnail. A face may be correct at the start and change during a turn, smile or hand movement. Keep a video rejection reason distinct from the image-stage review.
The Wan prompt guide provides simple motion briefs. For problems with the source image, use the inpainting process before animation.
Turn accepted results into a production reference
Save the identity sheet, accepted views, workflow version and a small collection of failed examples with notes. When another operator joins, ask them to reproduce the five-scene test before creating a large batch. This checks the handoff as well as the workflow.
For a hosted alternative, build your recurring creator in Clout. Use the same reference sheet and review gates when generating images and videos there. A recognizable character comes from a clear production target and consistent review, whichever environment you choose.
