PuLID for Flux: Identity Injection Without Training

PuLID Flux (PuLID-FLUX) is a tuning-free identity model that injects one face image into FLUX.1 [dev] generations, so a persona keeps its face with no LoRA training. The current release is v0.9.1, it runs in ComfyUI through community node packs on roughly 12GB of VRAM with an 8-bit Flux build, and its authors still call it a beta.

Versions, memory figures, settings and licenses checked October 2026 against the PuLID repository, its PuLID for FLUX doc and model page, the PuLID paper, the ComfyUI_PuLID_Flux_ll and ComfyUI-PuLID-Flux READMEs, ComfyUI's FLUX.1 tutorial, the FLUX.1 [dev] license, InsightFace and the InstantID repository. Settings are the authors' recommendations; nothing here is a scored test of ours.

This tutorial covers the install, the two settings the authors say need care, how PuLID differs from InstantID, and the licenses stacked under it. It is the Flux identity-adapter branch of our consistent character AI guide. If you work in FLUX.2, read the section on FLUX.2 below before installing anything: you may not need an adapter at all.

What PuLID Flux is, and what you will have at the end

You will have a ComfyUI graph that takes one approved image of your persona and renders that face in Flux scenes, with the pose and framing left to the prompt. The paper's stated goal is "minimizing disruption to the original model", so the background, lighting and composition stay close to what Flux would have drawn without the face.

There is no landmark ControlNet in that chain. That is the first practical difference from InstantID, which poses the head from the reference photo unless you override it.

How much VRAM PuLID Flux needs

Memory depends on Flux precision and how much you offload. The authors publish peak figures for their own demo, and they are the most reliable numbers available.

In ComfyUI the node packs give rounder guidance: about 22GB for the 16-bit FLUX.1 [dev] model and about 12GB for an 8-bit GGUF or fp8 build. The original pack recommends the 16-bit or 8-bit GGUF version and notes that one fp8 variant, e5m2, returned blurry backgrounds. If your card is smaller than that, our cloud GPU hosting comparison lists rental options.

Pre-flight checklist

PuLID Flux ComfyUI setup, step by step

The PuLID authors do not ship a ComfyUI node for Flux; their doc says to stay tuned for the community implementation. Two community packs matter. ComfyUI-PuLID-Flux came first, calls itself an alpha prototype and has not been updated since October 2024. ComfyUI_PuLID_Flux_ll is the fork that fixed its model pollution problem and links the v0.9.1 weights; its last commit is from November 2025. The steps below use the fork.

Folders for the Flux files follow ComfyUI's current FLUX.1 tutorial; the pack READMEs still use the older models/unet and models/clip names for the same files. GGUF builds also need the ComfyUI-GGUF pack. The finished chain:

Load Diffusion Model (FLUX.1 dev)  ->  Apply PuLID Flux  ->  sampler Load PuLID Flux Model (pulid_flux_v0.9.1)  ->  pulid_flux Load Eva Clip (PuLID Flux)                 ->  eva_clip Load InsightFace (PuLID Flux)              ->  face_analysis Load Image (hero)                          ->  image Apply PuLID Flux  [weight: 1.0 | start_at: 0.0 | end_at: 1.0 | optional: attn_mask, options] Text prompt -> FluxGuidance (4) -> sampler (cfg 1) -> VAE Decode -> Save Image

Flux PuLID workflow: the two settings that matter

The authors name two parameters as crucial: when ID insertion starts, and which kind of CFG you run. Everything else can stay at its default while you learn those.

• Insertion timing is the likeness-versus-freedom dial. Insert from the first step and fidelity is highest while editability drops. Insert later and the prompt regains control, at some cost in likeness. If every image wears the reference's expression or ignores an outfit change, start later. If the face reads as a relative, start earlier.
• The demo counts steps; ComfyUI counts fractions. The demo's slider runs from step 0 to 10 on a 28-step default. As plain arithmetic, step 4 of 28 is a start_at of about 0.14. Treat that as a conversion, not a tested value, and adjust by eye.
• Stay on guidance for realism. The doc reports similar ID fidelity for both CFG modes in most cases, with better aesthetics and facial naturalness on guidance alone. It suggests true CFG when likeness is still low, or when a stylised prompt gets a weak response.
• Use bf16 if you can. The authors state that fp8 gives slightly worse face detail with a similar layout.

The fork adds one more control worth knowing: a Pulid Flux Options node that picks which face to read when the reference holds several, sorted left to right, top to bottom or by size. A single-person hero image avoids the question.

PuLID vs InstantID

On Flux there is no contest to run, because InstantID has no Flux release: its ComfyUI pack says it is only for SDXL. The comparison that matters is how each one behaves, so you know what you gain or lose by choosing a base model.

Both depend on InsightFace's antelopev2 models to read the face. Our InstantID guide has that install and its settings, and consistent characters without a LoRA ranks both against the hosted routes.

License: three layers under one face

PuLID itself is permissively licensed. The model it runs on and the face models it calls are not, and those decide what you can do with the result.

The FLUX.1 [dev] license is worth reading in full if money is involved. It says outputs may be used for any purpose, including commercial ones, except as the license prohibits, and separately that using the model for revenue-generating activity is not a non-commercial purpose. Black Forest Labs offers commercial licenses for that case. We are not lawyers; if the persona earns, get the license question answered before you build on this stack.

On likeness, the PuLID disclaimer says users are "expected to comply with local laws and utilize it responsibly", and Black Forest Labs' usage policy (last revised August 4, 2026) bars unlawful impersonation, including unlawful use of a real person's name, image, voice or likeness. Use an original AI persona or your own face. Our AI likeness rights guide explains why a lookalike is as risky as a copy.

PuLID and FLUX.2: do you need an adapter at all?

The PuLID team has released nothing for FLUX.2. A community project, ComfyUI-PuLID-Flux2, adapts the method to FLUX.2 [klein] and [dev] with its own weights and the same InsightFace and EVA-CLIP stack; treat it as an experiment by one developer. FLUX.2 reads reference images natively, so the simpler route there is multi-reference editing, covered in our Flux consistent character guide.

The PuLID authors themselves now point readers to DreamO, their newer customization framework, which has a native ComfyUI implementation. Its README says the ID task reaches higher facial fidelity than earlier methods but "introduces more model contamination" than PuLID. PuLID remains the one built to leave the rest of the image alone.

Troubleshooting

• Low likeness on a male persona. A documented limitation of the Flux model. Lower start_at, try true CFG, and use the clearest reference you have.
• Soft faces on an 8-bit build. Expected: the authors document a quality gap between fp8 and bf16.
• insightface fails to build on Windows portable. The fork's README links the InsightFace issue that covers the missing Python.h error.
• Errors on older GPUs. The original pack notes that the node does not work on hardware with CUDA compute capability below 8.0 when Flux runs in fp8.
• You need a specific pose. PuLID carries the face, not the body. Add a Flux-compatible pose control from our ControlNet guide, or follow the pose-lock workflow.
• Faces drift in wide shots. The face covers too few pixels. Add a detailer pass that reuses the PuLID-patched model; the chain is in our ComfyUI consistent character workflow.

When to stop injecting and train

PuLID reads one image, so it knows one angle and one expression. That is enough to test a face, shoot a first set and build references. When profile views and full-body shots keep missing, move the identity into weights: your approved PuLID renders are the start of the dataset. Our face LoRA image-count guide covers how many you need, and face LoRA vs InstantID vs IP-Adapter weighs training against the adapters. The AI Influencers program goes deeper where this tutorial stops, with lessons on Flux and SDXL setup, LoRA training for consistent faces and body consistency with ControlNet.

PuLID for Flux: FAQ