This is a simplified guide to an AI model called Flux-Pulid maintained by Jichengdu. If you like these kinds of analysis, you should join AImodels.fyi or follow us on Twitter.
flux-pulid
is a face identity customization model built on the FLUX diffusion architecture that generates images matching specific identity characteristics extracted from reference photos. Developed by jichengdu at ByteDance, this model implements the PuLID (Pure and Lightning ID Customization via Contrastive Alignment) method presented at NeurIPS 2024. The model uses contrastive alignment to inject identity information into the FLUX generation process while maintaining high image quality and editability through text prompts. The critical distinction before using it: this model trades some identity fidelity compared to newer v0.9.1 releases in exchange for broader male face compatibility, and the identity injection mechanism via the start_step
parameter fundamentally alters how much you can edit the generated image through text prompts—starting at step 0 preserves identity most faithfully but reduces prompt control, while starting at step 4 enables more creative editing but weakens identity retention.
Portrait generation with specific identity preservation for creative projects. When you have a reference photo of a person and need to generate new images of them in different contexts, styles, or compositions, flux-pulid
excels at maintaining recognizable facial characteristics while applying text-based modifications. For example, generating professional headshots in various business settings, reimagining a portrait subject in historical clothing, or creating multiple stylistic interpretations of the same person's face. The identity weight parameter (0.0–3.0) lets you tune how strongly the model enforces facial similarity, useful when the text prompt demands significant appearance changes that conflict with strict identity preservation.
Diverse pose and expression variations from a single reference photo. Rather than hiring models for multiple photoshoots, you can use one reference portrait and generate variations with different poses, expressions, lighting conditions, and camera angles. This is valuable for headshot galleries, dating profile variations, or character exploration in game and film pre-production. The guidance scale and prompt sophistication of FLUX means you have fine control over these variations through natural language rather than manual parameter tweaking.
Identity-consistent creative fashion and styling exploration. Fashion designers and stylists can reference a person's face and generate mockups showing how they look in different clothing, hairstyles, and makeup styles without requiring fitting sessions. The model preserves facial identity while the text prompt controls visual elements like "wearing a Victorian gown with elaborate updo" or "punk rock aesthetic with heavy eyeliner." This workflow compresses design iteration cycles from hours to minutes.
Conditional face generation for content moderation and synthetic data. The ability to generate variations of specific identities from reference images can support testing facial recognition systems, generating balanced datasets for bias evaluation, or creating templated variations of approved identities for controlled content. The seed parameter ensures reproducibility when needed for validation workflows.
Identity fidelity gaps on certain male face inputs. The v0.9.0 model used here (based on the Replicate metadata) exhibits lower ID fidelity on some male face reference images compared to later v0.9.1 releases. The README explicitly notes "ID fidelity is not high enough for some male inputs," meaning certain male faces may not generate with sufficient similarity to the reference. This is a known limitation of this specific version and suggests male identity preservation requires careful prompt engineering or switching to v0.9.1 if available.
Fundamental tradeoff between identity fidelity and prompt editability. The start_step
parameter (0–10, recommended 0–4) controls when identity information injection begins in the denoising process. Starting at step 0 maximizes identity preservation but severely constrains what text prompts can achieve—the generated image will look most like the reference but may ignore creative prompt requests. Starting at step 4 allows much more prompt influence and creative control but weakens identity similarity by roughly 5 percentage points according to the README. You cannot simultaneously achieve maximum fidelity on both dimensions.
Maximum resolution constraints limit billboard and large-format use. The model generates images up to 1536×1536 pixels (width and height individually capped at 1536 minimum 256), but inference time and quality degrade at maximum resolutions. For large-format printing, cinema-quality assets, or applications requiring 4K resolution, this model falls short. The default is 896×1152, a reasonable compromise.
Inference speed and computational requirements not publicly documented. The README mentions the model requires 16GB VRAM for FLUX-dev standard precision (or can use fp8 quantization to reduce to 12GB), but actual inference latency per API call is not specified. This makes it difficult to predict response times for time-sensitive applications or batch processing workflows without running benchmarks.
Limited batch generation and output control. You can generate 1–4 images per call (via num_outputs
), but cannot generate more without multiple API calls. The model offers no built-in upscaling, inpainting, or face detection/alignment features—you handle image preprocessing and post-processing separately.
Text prompt encoding constraints. The max_sequence_length
parameter caps T5 tokenization at 128–512 tokens (default 128), meaning extremely long or detailed prompts may be truncated. This is much shorter than modern LLM context windows and may force you to compress complex creative direction.
License and commercial use restrictions. The license is available at the GitHub repository, but the README includes a disclaimer: "This project strives to impact the domain of AI-driven image generation positively. Users are granted the freedom to create images using this tool, but they are expected to comply with local laws and utilize it responsibly. The developers do not assume any responsibility for potential misuse by users." This suggests caution around identity-based synthetic images in regulated contexts (deepfakes, impersonation, etc.).
** flux-pulid by zsxkib.** This appears to be a different implementation or version of PuLID-FLUX on Replicate. Without detailed schema comparison, both likely share the same underlying model but may differ in preprocessing, quantization, or wrapper code. Pick this model if the zsxkib version is unavailable or has different latency characteristics; both should produce similar quality outputs for the same inputs.
** flux-pulid by bytedance.** The official ByteDance implementation of PuLID-FLUX, likely identical or nearly identical to this model since both are from the original research team. If you need guaranteed official support or the latest optimizations, prefer the bytedance version; otherwise, functionality should be equivalent.
** pulid by zsxkib.** This is the SDXL-based PuLID model (v1 or v1.1), not the FLUX version. Choose SDXL-PuLID if you need faster inference on older hardware, more mature model versions with v1.1 improvements, or existing integrations with SDXL-based workflows. Choose this FLUX version for superior image quality, better prompt adherence, and more advanced generation features at the cost of higher computational requirements.
** oot_diffusion_dc by qiweiii.** This model specializes in full-body outfit-try-on synthesis, while
flux-pulid
focuses on face identity. Use OOT Diffusion when you need to see how clothing fits on a full body; use flux-pulid
when the task is face-centric (portraits, expressions, styles applied to the head). These are complementary rather than competing tools.** flux-multi-pulid-controlnet by mattheum.** This is a multi-identity fork of FLUX-PuLID that accepts multiple faces plus a depth map and bounding boxes to place different identities in specific regions of an image. Use the standard
flux-pulid
for single-identity generation with maximum fidelity and ease of use; use the ControlNet variant when you need simultaneous generation of multiple distinct people in one image with precise spatial control.Architecture and training. The model implements PuLID as described in the NeurIPS 2024 paper using FLUX-dev as the base diffusion architecture. The approach uses contrastive alignment to embed identity information extracted from reference face images, enabling precise identity conditioning during the generation process. The underlying FLUX model provides significantly higher image quality and prompt adherence compared to SDXL-based alternatives.
Model versioning. This Replicate deployment runs PuLID-FLUX-v0.9.0 (released September 12, 2024, per the README). A newer v0.9.1 version exists with approximately 5 percentage points higher ID similarity on quantitative metrics, particularly for male face inputs, though this deployment has not yet been updated.
Hardware requirements. The model requires a GPU with at least 16GB VRAM for standard FLUX-dev inference in bf16 precision. It can be optimized to run on 12GB VRAM using fp8 quantization (requires PyTorch 2.4.1+), according to the README.
Input specifications:
Output specifications: Array of image URIs in the specified format (webp or jpeg), with quality determined by the output_quality parameter.
File format and model artifacts. The underlying model weights ship as safetensors format for security. The Replicate deployment wraps this with Cog (version 0.13.7 per metadata) for containerized inference.
import replicate
client = replicate.Replicate()
output = client.run(
"jichengdu/flux-pulid",
input={
"main_face_image": "https://example.com/reference_face.jpg",
"prompt": "a professional headshot in a modern office, natural lighting, wearing business attire",
"negative_prompt": "bad quality, worst quality, text, signature, watermark, extra limbs",
"width": 896,
"height": 1152,
"num_steps": 20,
"start_step": 0,
"guidance_scale": 4.0,
"id_weight": 1.0,
"seed": 42,
"output_format": "webp",
"output_quality": 80,
"num_outputs": 1
}
)
print(output)
Q: How does the start_step parameter affect output quality and identity preservation?
A: The start_step
parameter controls when identity information is injected into the diffusion process (0–10, with 0–4 recommended). Starting at step 0 maximizes facial similarity to the reference image but severely limits what the text prompt can achieve, resulting in less creative variation. Starting at step 4 allows the text prompt much more influence and creative control but reduces identity fidelity by approximately 5 percentage points. Choose based on whether you prioritize identity accuracy or prompt flexibility for your use case.
Q: What is the difference between guidance_scale and true_cfg?
A: The guidance_scale
parameter (1.0–10.0) controls standard classifier-free guidance, strengthening how much the model follows your text prompt. The true_cfg
parameter (1.0–10.0) is an advanced option where values >1.0 enable "True CFG," providing finer-grained control over generation characteristics. For most users, leave true_cfg
at the default 1.0 and adjust guidance_scale
instead.
Q: Can I use this model to generate images of people who are not in my reference photo?
A: No, this model is designed for identity customization—it extracts facial characteristics from the reference image you provide and generates new images preserving those characteristics. It cannot synthesize arbitrary identities without a reference photo. If you try to use a photo of person A to generate person B, the output will reflect person A's identity features.
Q: What happens if my reference face image is low quality or partially obscured?
A: The model relies on clear facial information to extract identity characteristics. Low-resolution, heavily filtered, or partially obscured reference images may result in poor identity transfer or artifacts in the output. The README explicitly mentions that identity fidelity varies, particularly on some male face inputs in the v0.9.0 version, so quality issues with specific face types are known limitations.
Q: Is this model suitable for production use with user-uploaded reference images?
A: Use caution. The README includes a disclaimer that users must comply with local laws and use the tool responsibly. Generating synthetic images of identities from user photos raises legal and ethical concerns around deepfakes, consent, and impersonation. In production contexts, you should implement identity verification, usage policies, and legal review before deploying identity customization features to end users.
Q: How does id_weight differ from start_step in controlling identity influence?
A: The id_weight
parameter (0.0–3.0) controls the magnitude of identity information injected, acting like a volume knob for how strongly the reference face influences the output. The start_step
parameter controls when that injection begins, fundamentally changing the interplay between identity and prompt control. Increasing id_weight
strengthens identity at any given start_step
; decreasing start_step
injects identity earlier and locks in more facial similarity regardless of id_weight
.
Q: What image dimensions should I use for best results?
A: The default 896×1152 is a reasonable balance, representing landscape-to-portrait aspect ratio typical of portrait photography. You can use 1536×1536 for maximum resolution, but inference time increases and quality may vary. For specific use cases (LinkedIn headshots are typically square or vertical; fashion photography may prefer wider shots), adjust width and height accordingly, keeping both between 256 and 1536.
Q: Is the model still actively maintained?
A: Yes, the original PuLID project is actively maintained by ByteDance researchers. The v0.9.1 release (October 2024) improved ID fidelity, and the main repository continues accepting issues and contributions. However, this specific Replicate deployment runs v0.9.0; the maintainer may update it to v0.9.1 in the future if they choose to refresh the wrapper.