ComfyUI Consistent Character Workflow: Copy-and-Adapt Templates (2026)

A ComfyUI consistent character workflow has two halves: an identity source that conditions every generation (IP-Adapter, PuLID, a LoRA, or FLUX.2 klein's native reference input) and a face detailer pass that re-renders small faces with that same identity. Below are three copy-and-adapt node templates for those jobs plus a no-custom-node FLUX.2 route, each with the settings to start from and the place it fails.

Nodes, settings and licenses checked October 2026 against the ComfyUI_IPAdapter_plus, PuLID_ComfyUI, PuLID and Impact Pack repositories, Impact Pack's detailer reference, the IP-Adapter-FaceID model card, InsightFace, and ComfyUI's FLUX.2 klein guide and Flux2ImageNode docs.

This is the ComfyUI build sheet. If you are still deciding between ComfyUI, Midjourney, Nano Banana or a hosted model, read our consistent character AI guide first; it compares every method. Come back here when you have picked ComfyUI and need the graph.

What you will have at the end

One approved hero image, a saved identity graph you reuse for every post, a detailer pass that holds the face in wide shots, and a set of approved angle views that doubles as a character sheet and a LoRA dataset.

Which template fits your build

Template A: IP-Adapter face reference (SDXL)

IP-Adapter conditions the generation on an image; the IPAdapter plus README calls it a one-image LoRA. For a face, the Plus Face model on a cropped face works without InsightFace. The chain:

Load Checkpoint (SDXL) -> IPAdapter Unified Loader (preset: PLUS FACE (portraits)) -> IPAdapter Advanced  [image: hero face crop | weight: 0.8 | weight_type: adjust for prompt adherence] -> KSampler  [positive: identity block + scene | fixed seed while testing] -> VAE Decode -> (Template C detailer) -> Save Image

• Models. The unified loader only finds files named exactly as the README lists them: the ViT-H CLIP vision encoder in models/clip_vision and ip-adapter-plus-face_sdxl_vit-h.safetensors in models/ipadapter.
• Starting settings. The README suggests lowering weight to at least 0.8 and raising steps, and changing the weight type in IPAdapter Advanced for more prompt adherence.
• Where it fails. The repository has been maintenance-only since April 14, 2025. FaceID variants need InsightFace installed plus a specific paired LoRA, and the FaceID model card says those models are for research, not commercial use, and do not achieve perfect ID consistency.

Template B: PuLID identity injection

PuLID injects an identity embedding taken from the hero face. PuLID_ComfyUI is the native SDXL implementation; PuLID-FLUX runs through community node packs listed by the PuLID authors.

Load Checkpoint (SDXL) -> Load PuLID Model (ip-adapter_pulid_sdxl_fp16 in models/pulid) + Load InsightFace (PuLID): antelopev2  + Load Eva Clip (PuLID): EVA02-CLIP-L-14-336 -> Apply PuLID  [image: hero | method: fidelity | weight: 1.0 default] -> KSampler  [try sgm_uniform scheduler] -> VAE Decode -> (Template C) -> Save Image

• Settings that matter. The method parameter: fidelity sits closer to the reference, style gives the checkpoint more freedom. On the Advanced node, lower fidelity values give higher resemblance.
• On FLUX. The authors' PuLID-FLUX notes suggest starting ID insertion at timestep 4 for realistic images (lower it if similarity is weak) and 0 to 1 for stylised ones.
• Where it fails. The README says reference quality is very important (use a clean, sharp picture) and that quality degraded on checkpoints other than the one it was built around. The FLUX version notes lower ID fidelity for some male inputs and falls under the FLUX.1 dev license. The cubiq repo is maintenance-only since April 14, 2025, and the InsightFace models it needs are research-only.

Template C: the face detailer chain

Identity methods hold close-ups well; in full-body shots the face covers too few pixels and drifts. A detailer detects the face, crops it, re-renders it larger and pastes it back. Impact Pack's FaceDetailer is the standard node; its detector provider for YOLO models lives in the separate Impact Subpack since version 8.0.

Base image (from Template A, B, a LoRA or the native route) -> UltralyticsDetectorProvider (face bbox model, Impact Subpack) -> FaceDetailer  [same model + LoRA/IP-Adapter/PuLID | identity block prompt | guide_size: 512 default | denoise: moderate | noise_mask: on] -> Save Image

• guide_size. Per the detailer reference, faces smaller than guide_size are enlarged to it; faces already larger are skipped as not needing detail, unless force_inpaint is on. Raise guide_size for more face detail.
• noise_mask. Generally recommended on, per the same reference.
• Where it fails. Run it without your identity source and it draws a generic face at higher resolution, which reads as drift. Impact Pack has also changed detailer node structure between versions, so an old saved workflow can error after an update.

ComfyUI consistent character without LoRA: the native FLUX.2 route

In 2026 the simplest no-LoRA graph needs no custom nodes. ComfyUI's FLUX.2 klein guide ships image-edit workflows in the built-in Workflow Templates, with multi-reference composition and iterative edits listed as supported. It reports the 4B base model at about 9.2GB of VRAM, and Black Forest Labs licenses klein 4B under Apache 2.0.

1. Open Workflow Templates and load the FLUX.2 [klein] 4B Image Edit Base workflow.
2. Put the hero image in the first LoadImage node.
3. Prompt by position: "the woman in image 1, [identity block], now sits at a cafe window, overcast light".
4. Approve outputs and add the best angle views as extra references for later scenes.
5. Finish wide shots with Template C.

If you would rather not run the model locally, the built-in Flux2ImageNode sends the job to Black Forest Labs' FLUX.2 [pro] or [max] API with up to eight reference images, resized so the total does not exceed 2048 x 2048 pixels. Our Flux consistent character guide covers prompt word order and the edit chain for poses in detail.

ComfyUI character sheet workflow

A character sheet is the set of approved angles that every later image is checked against. Build it from the hero, one angle at a time, in whichever template you chose:

Prompting for all angles in one grid image is quicker, but faces in small panels drift for the same crop-scale reason the detailer exists. Generating each view separately at full size costs more runs and gives a sheet you can trust. For exact poses per angle, add an OpenPose ControlNet; our ControlNet guide covers the setup.

Which template to start with

• Commercial persona, no custom nodes wanted: the native FLUX.2 klein 4B route plus Template C. Apache 2.0 model, built-in template, nothing to keep patched.
• Existing SDXL setup and checkpoints you like: Template A with the plain Plus Face model, then Template C. Avoid the FaceID variants if the work is commercial.
• Research or personal projects that need the strongest identity from one photo: Template B, within the InsightFace and FLUX.1 dev terms.
• Already have a character LoRA: skip A and B, load the LoRA in the base pass and in Template C.

Save the graph as your template

Once a graph produces faces you approve, freeze it. Save the workflow file with a name that records the base model and identity method, keep the hero image and reference views in a fixed input folder, and change only the scene prompt and seed from post to post. Group the identity nodes so nobody edits them by accident. When you change anything inside the identity group (weight, model, detailer settings), save it as a new version and re-run your character sheet angles through it before posting with it. That habit turns a one-off graph into a template a whole team can run, and it makes drift traceable to a specific change.

Troubleshooting the graph

• Red nodes after an update. Custom node packs lag ComfyUI releases. Update through ComfyUI-Manager, and for production keep a pinned copy of the ComfyUI version your graph was built on. Our ComfyUI workflow library guide covers safe sources and setup for shared workflows.
• Pasted-on or oversaturated face. Weight too high; step it down in 0.1 increments.
• The face drifts in a way none of this explains. Work through the seven causes in our AI face consistency guide, one variable at a time.

When a template stops being enough

Reference templates are the fastest start and the right tool for a new persona. Once you post daily, a LoRA trained on your approved sheet and best outputs holds the face across more angles and lighting than a single reference can, and Template C still runs on top of it. Our character LoRA dataset guide shows how to turn the sheet into training data. The AI Influencers program covers the full ComfyUI persona stack, from hero image to trained LoRA to video.

ComfyUI consistent character: FAQ