h4 - Identity Engine

H4_IdentityEngine is h4_Live's "make the same character every time" node - a full text-to-image generator wrapped around a face-swap, so you get both a fresh scene and a consistent face. It's the character-consistency answer for people who don't want to train a LoRA: you keep a "DNA" prompt (fixed character traits like "blonde, scars") and a scene prompt (what they're doing), and the node generates the scene, then optionally swaps a reference face onto the result using the FaceForge machinery.

The clever design bit is the DNA prompt concept. Instead of repeating a character description in every positive prompt and hoping the model stays consistent, you put the immutable traits in positive_dna, which gets concatenated with the per-scene positive. That separation is genuinely the right mental model for consistent characters: DNA is the person, positive is the moment. The face-swap layer then enforces identity at the pixel level on top of whatever the model guessed - a hybrid that mirrors what the community does by hand ("IP-Adapter gets you 80% there, face swap gets the rest").

The settings that matter

• positive_dna - the character's fixed identity text. positive - the scene. negative - what to avoid. These three are the heart of the node.
• preset - a saved character preset selector (the engine can save/load character configs, so your DNA doesn't get re-typed every session).
• ckpt_name, vae_name, clip_name, seed, steps, cfg, sampler_name, scheduler, denoise, width, height, batch_size - the generation recipe, mostly standard sampler fare. VAE/CLIP default to "Baked"/"Use Checkpoint," which is right for most people.
• The swap block: face_image (drag a photo in to clone that face) or face_model (a saved .safetensors face identity - pair it with H4_BuildFaceModel), swap_enabled, face_similarity (how strictly to copy facial features; higher = more likeness but "maybe weird"), swap_model (inswapper_128).
• The polish block: restore_enabled / restore_model (CodeFormer by default), codeformer_weight, occlusion_enabled (needs a sam_model), preserve_glasses, and upscale_enabled / upscale_model.
• Overrides: model_opt/clip_opt/vae_opt to inject your own models, plus positive_text/negative_text inputs to drive the prompts externally.

The outputs

• character_image - the finished generation with the face applied.
• model, clip, vae - the loaded stack passed through, so the engine can feed a downstream node without reloading.

Installing and dependencies

ComfyUI Manager → search "h4_Live", or:

cd ComfyUI/custom_nodes
git clone https://github.com/m3rr/h4_Live

restart. The swap path needs insightface, onnxruntime, and opencv (no requirements.txt in the pack, lazy-imported, so first run may hit a ModuleNotFoundError):

pip install insightface onnxruntime opencv-python

Restore/swap/SAM model weights auto-download on first use.

Where people get burned

Tune face_similarity down from 1.0 - max likeness produces uncanny results, and the tooltip's "maybe weird" is understating it. Keep positive_dna short and stable across scenes; that's what makes the character read as the same person. And the licensing reminder applies to the whole identity family: InsightFace's weights are non-commercial, so this is personal-work territory, not a commercial pipeline. It's the most ambitious node in the pack - and as with everything in h4_Live, it's a work in progress, so expect the occasional rough edge and keep the "?" help drawer handy.