InstantID Guide: Same Face From a Single Photo

InstantID is a tuning-free method that reproduces one face across SDXL images from a single reference photo, by pairing a face identity embedding with a landmark ControlNet called IdentityNet. In ComfyUI you install one node pack, download three sets of model files, start the weight at 0.8 and lower the CFG to 4 or 5.

Files, node names, settings and license terms checked October 2026 against the InstantID repository, its Hugging Face model card and technical report, the ComfyUI_InstantID README and node source, the InsightFace license section and the IP-Adapter model card. The settings are the projects' own recommendations, not scores from a test of ours.

This page is the persona-first setup: what to download, how to wire it, which numbers to start from, how to fix the artifacts the maintainers document, and where InstantID stops being the right tool. It is one branch of our consistent character AI guide, which compares every method. If you have not chosen a method yet, start there.

What InstantID is, and what you will have at the end

You will have a saved ComfyUI graph that takes one approved image of your persona and renders that face in new scenes, on any SDXL checkpoint, with no training run. The technical report describes the design as "strong semantic and weak spatial conditions": the identity embedding says who, and the landmark image says roughly where.

Two things follow from that design. The face is read by a recognition model rather than copied as a picture, so the method is aimed at identity, not at the hair and lighting of that one photo. And the head is posed from the reference's keypoints by default, which is why untouched InstantID sets all look at the camera the same way. One limit to know before you start: the released model is for SDXL. If your persona lives on Flux, the equivalent tool is PuLID for Flux.

Read the license before you install

InstantID is free to download and restricted in use. The repository's disclaimer makes three separate statements, and the third is the one people miss.

The same disclaimer covers whose face you use: people may create images with the tool but are "obligated to comply with local laws and utilize it responsibly". Treat that as a floor. Use an original AI persona or your own face, never a real person who has not agreed, and remember that a persona resembling a real person carries the same risk as one built from their photo. Our AI likeness rights guide covers where the lines sit.

Pre-flight checklist

The repository publishes no VRAM figure. You are loading an SDXL checkpoint plus a 2.5GB ControlNet and a 1.69GB adapter, so expect it to need more memory than plain SDXL; the README's only memory advice is CPU offloading and VAE tiling for its Python pipeline. If you have no manager yet, our ComfyUI-Manager install guide covers it, and the free AI influencer prompt generator drafts the identity block.

InstantID ComfyUI setup, step by step

The pack to install is ComfyUI_InstantID, which implements InstantID natively instead of wrapping the diffusers pipeline. It is the first ComfyUI option the InstantID authors list.

Sizes are from the Hugging Face file listing, checked October 2026. The finished chain looks like this:

Load Checkpoint (SDXL) Load InstantID Model (ip-adapter.bin)  +  InstantID Face Analysis (provider: CUDA or CPU) Load ControlNet Model (IdentityNet)    +  Load Image (hero) -> Apply InstantID  [weight: 0.8 | start_at: 0 | end_at: 1 | optional: image_kps, mask] -> KSampler  [positive: identity block + scene | CFG 4 to 5 | fixed seed while testing] -> VAE Decode -> Save Image

Apply InstantID takes the model and both conditionings in and passes them out again, so it sits between your prompt nodes and the sampler. The other example files in the pack cover a posed keypoint image, a depth ControlNet, IP-Adapter styling and a multi-ID graph the maintainer calls "hackish" and slower.

InstantID settings: ID weight and strength by checkpoint type

Nobody publishes per-checkpoint numbers for InstantID, and we have not scored any. What the projects document is one default and a set of directions, and that is enough to tune a persona.

• Two dials, one default. The project's example code sets ip_adapter_scale and controlnet_conditioning_scale to 0.8. In ComfyUI the standard node has a single weight at 0.8 that drives both.
• Split them on the Advanced node. Apply InstantID Advanced exposes ip_weight for the adapter and cn_strength for the ControlNet, both defaulting to 0.8. The README says the adapter accounts for about 25% of the composition and the ControlNet for the rest.
• Directions from the authors. Raise both for more similarity. For over-saturation lower the adapter first, then the ControlNet. For more prompt control lower the adapter.
• Noise. The standard node injects 35% noise into the negative embeds to reduce the burnt look. The Advanced node leaves noise at 0 until you set it.

Tune in this order: CFG, then resolution, then weights. Most first-run complaints are CFG and resolution problems that no weight fixes. One more input is worth knowing about: the node source averages the face embeddings when you feed it a batch of reference images, and takes keypoints from the first image only. Three approved views of your persona, batched, give the adapter more than one photo's worth of identity while the pose still comes from one place.

The InstantID README also points to a follow-up, InstantID-Rome, "if you are pursuing better results". Its repository shows comparison images and ends with "coming", with no install instructions, so there is nothing to add to a graph today.

InstantID face consistency: fixing the common artifacts

Each of these has a documented cause, so fix them by cause, not by nudging weights at random.

• Small faces drift. In a full-body shot the face covers few pixels and the identity thins out. Re-render it with a face detailer pass that reuses the same InstantID nodes; the chain is in our ComfyUI consistent character workflow.
• Poses beyond the head. InstantID's keypoints cover the face only. For body pose add an OpenPose ControlNet beside it, which the pack supports as an additional ControlNet; see our ControlNet guide and the pose-lock workflow for running a whole pose sheet.
• Red nodes after an update. The pack has been in maintenance-only mode since April 14, 2025. Keep a copy of the ComfyUI version your graph was built on.
• Drift none of this explains. Work through the causes in our AI face consistency guide, one variable at a time.

InstantID vs IP-Adapter

InstantID's authors credit IP-Adapter and ControlNet as the work it builds on. The practical difference is what each one reads from your photo, and what license comes with the files.

Between the two sits IP-Adapter FaceID, which swaps the image embedding for an InsightFace one and adds a paired LoRA; its model card says it is released for research and not intended for commercial use. Our IP-Adapter face guide decodes which file fits which base model. For a full ranking of the no-training routes, hosted ones included, see consistent characters without a LoRA.

When InstantID beats a trained LoRA, and when it does not

InstantID wins on time to first image. Its README shows results the authors call competitive with character LoRAs "without any training", and that is the right way to use it: to find out in an afternoon whether a face is worth building on. For the three-way decision with published costs and a test you can run, see face LoRA vs InstantID vs IP-Adapter.

A LoRA has seen the profile, the laugh and the full-body shot because you put them in the dataset. InstantID has seen one photo and guesses the rest. When you start rejecting more outputs than you keep, your approved InstantID renders become the training set: see how many images a face LoRA needs and the LoRA training guide for consistent faces. The AI Influencers program goes deeper where this page stops, with lessons on creating the persona, LoRA training for consistent faces and body consistency with ControlNet.

InstantID: FAQ